Statistics for Social and Behavioral Sciences

Estela Bee Dagum
Silvia Bianconcini

Seasonal
Adjustment
Methods and
Real Time Trend-
Cycle Estimation
Statistics for Social and Behavioral Sciences

Series editor
Stephen E. Fienberg
Carnegie Mellon University
Pittsburgh
Pennsylvania
USA
Statistics for Social and Behavioral Sciences (SSBS) includes monographs and
advanced textbooks relating to education, psychology, sociology, political science,
public policy, and law.

More information about this series at http://www.springer.com/series/3463

Estela Bee Dagum • Silvia Bianconcini

Seasonal Adjustment
Methods and Real Time
Trend-Cycle Estimation

123
Estela Bee Dagum Silvia Bianconcini
Department of Statistical Sciences Department of Statistical Sciences
University of Bologna University of Bologna
Bologna, Italy Bologna, Italy

ISSN 2199-7357 ISSN 2199-7365 (electronic)

Statistics for Social and Behavioral Sciences
ISBN 978-3-319-31820-2 ISBN 978-3-319-31822-6 (eBook)
DOI 10.1007/978-3-319-31822-6

Library of Congress Control Number: 2016938799

Mathematics Subject Classification (2010): 62G08, 62M10, 62P20, 62P25

© Springer International Publishing Switzerland 2016

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG Switzerland
To my sons, Alex, Paul, and Leo, with their
lovely families for their strong valuable
support and lively discussions during the
writing of this book
To Paolo who always inspires me to try my
best
Preface

In order to assess the current stage of the business cycle at which the economy
stands, real time trend-cycle estimates are needed. The basic approach is that of
assessing the real time trend-cycle of major socioeconomic indicators (leading,
coincident, and lagging) using percentage changes, based on seasonally adjusted
data, calculated for months and quarters in chronological sequence. The main goal
is to evaluate the behavior of the economic indicators during incomplete phases
by comparing current contractions or expansions with corresponding phases in the
past. This is done by measuring changes of single time series (mostly seasonally
adjusted) from their standing at cyclical turning points with past changes over a
series of increasing spans. This differs from business cycle studies where cyclical
fluctuations are measured around a long-term trend to estimate complete business
cycles. The real time trend corresponds to an incomplete business cycle and is
strongly related to what is currently happening on the business cycle stage. Major
changes of global character in the financial and economic sector have introduced
high levels of variability in time series making difficult to detect the direction of
the short-term trend by simply looking at seasonally adjusted data, and the use of
trend-cycle data or smoothed seasonally adjusted series has been supported. Failure
in providing reliable real time trend-cycle estimates could give rise to dangerous
drift of the adopted policies. Therefore, a reliable estimation is of fundamental
importance.
This book includes two opening chapters, Chap. 1 which is a general introduction
and Chap. 2 on time series components. The remaining nine chapters are divided
in two parts, one on seasonal adjustment and the other on real time trend-cycle
estimation.
Since the input for trend-cycle estimation is seasonally adjusted data, Part I of this
book thoroughly discusses the definitions and concepts involved with three major
seasonal adjustment methods as follows:
Chapter 3. Seasonal adjustment, meaning, purpose, and methods.
Chapter 4. Linear smoothing or moving average seasonal adjustment methods.
Chapter 5. Seasonal adjustment based on ARIMA model decomposition:
TRAMO-SEATS.

vii
viii Preface

Chapter 6. Seasonal adjustment based on structural time series models.

Two of the seasonal adjustment methods are officially adopted by statistical
agencies, namely, X12ARIMA and TRAMO-SEATS, and their respective software
default options are illustrated with an application to the US Orders for Durable
Goods series. The third method, structural time series models and its software
STAMP, is also illustrated with an application to the US Unemployment Males (16
years and over) series.
Part II of the book comprises:
Chapter 7. Trend-cycle estimation.
Chapter 8. Recent developments on nonparametric trend-cycle estimation.
Chapter 9. A unified view of trend-cycle predictors in reproducing kernel Hilbert
spaces.
Chapter 10. Real time trend-cycle estimation.
Chapter 11. The effect of seasonal adjustment methods on real time trend-cycle
estimation.
Chapter 7 systematically discusses the definitions and concepts of the trend-cycle
component of the various seasonal adjustment methods previously introduced.
Chapter 8 concentrates on the last 20 years’ developments made to improve the
Henderson filter used to estimate the trend-cycle in the software of the US Bureau
of Census X11 and its variants, the X11/ X12ARIMA. The emphasis has been on
determining the direction of the short-term trend for an early detection of a true
turning point. It introduces in detail three major contributions: (1) a nonlinear trend-
cycle estimator also known as Nonlinear Dagum Filter (NLDF), (2) a cascade linear
filter (CLF) that closely approximates the NLDF, and (3) an approximation to the
Henderson filter via the reproducing kernel Hilbert space (RKHS) methodology.
Chapter 9 presents a unified approach for different nonparametric trend-cycle
estimators by means of the reproducing kernel Hilbert space (RKHS) methodology.
These nonparametric trend-cycle estimators are based on different criteria of fitting
and smoothing, and they are (1) density functions, (2) local polynomial fitting,
(3) graduation theory, and (4) smoothing spline regression. It is shown how
nonparametric estimators can be transformed into kernel functions of order two,
which are probability densities and from which corresponding higher-order kernels
are derived. This kernel representation enables the comparison of estimators based
on different smoothing criteria and has important consequences in the derivation of
the asymmetric filters which are applied to the most recent seasonally adjusted data
for real time trend-cycle estimation.
Chapter 10 is dedicated to real time trend-cycle estimation. Official statistical
agencies generally produce estimates derived from asymmetric moving average
techniques which introduce revisions as new observations are incorporated to the
series as well as delays in detecting true turning points. This chapter presents a
reproducing kernel approach to obtain asymmetric trend-cycle filters that converge
fast and monotonically to the corresponding symmetric ones. This is done with time-
varying bandwidth parameters because the asymmetric filters are time-varying. It
shows that the preferred one is the bandwidth parameter that minimizes the distance
between the gain functions of the asymmetric and symmetric filters. The theoretical
Preface ix

results are empirically corroborated with a set of leading, coincident, and lagging
indicators of the US economy.
Chapter 11 deals with the effects of the seasonal adjustment methods when the
real time trend is predicted with nonparametric kernel filters. The seasonal adjust-
ments compared are the two officially adopted by statistical agencies, X12ARIMA
and TRAMO-SEATS, applied to a sample of US leading, coincident, and lagging
indicators.
The eleven chapters have been written as complete as possible, and each one can
be read rather independently. We have also introduced a uniform notation all along
the chapters to facilitate the reading of the book as a whole.
This book will prove useful for graduate and final-year undergraduate courses
in econometrics and time series analysis and as a reference book for researchers
and practitioners in statistical agencies, other government offices, and business.
The prerequisites are a good knowledge of linear regression, matrix algebra, and
knowledge of ARIMA modeling.
We are indebted to participants and students who during many seminars and
presentations raised valuable questions answered in this book. Our most sincere
gratitude goes to our colleagues who encouraged us to write this book and all those
who provided many valuable and useful suggestions through lively discussions or
written comments. Our thanks also go to Veronika Rosteck, associate statistics
editor of Springer, for her sustained and valuable support while writing the book.
We are solely responsible for any errors and omissions.

Bologna, Italy Estela Bee Dagum

February 2016 Silvia Bianconcini
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.1 Book Structure.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.2 Part I: Seasonal Adjustment Methods . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2
1.2.1 Smoothing Linear Seasonal Adjustment Methods . . . . . . . . 3
1.2.2 ARIMA Model-Based Seasonal Adjustment Method . . . . 9
1.2.3 Structural Time Series Models .. . . . . . . .. . . . . . . . . . . . . . . . . . . . 14
1.3 Part II: Real Time Trend-Cycle Estimation . . . . .. . . . . . . . . . . . . . . . . . . . 21
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 26
2 Time Series Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 29
2.1 Time Series Decomposition Models.. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 30
2.2 The Secular or Long-Term Trend.. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 33
2.2.1 Deterministic Trend Models . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 34
2.2.2 Stochastic Trends . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 34
2.3 The Business Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 36
2.3.1 Deterministic and Stochastic Models
for the Business Cycle . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 39
2.4 The Seasonal Variations.. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 39
2.4.1 Seasonal Adjustment Methods .. . . . . . . .. . . . . . . . . . . . . . . . . . . . 41
2.5 Calendar Variations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 47
2.5.1 The Moving Holiday Component .. . . . .. . . . . . . . . . . . . . . . . . . . 47
2.5.2 The Trading Day Component . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 49
2.6 The Irregular Component . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 52
2.6.1 Redistribution of Outliers and Strikes .. . . . . . . . . . . . . . . . . . . . 52
2.6.2 Models for the Irregulars and Outliers .. . . . . . . . . . . . . . . . . . . . 53
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 56

xi
xii Contents

Part I Seasonal Adjustment Methods

3 Seasonal Adjustment: Meaning, Purpose, and Methods . . . . . . . . . . . . . . 61
3.1 Seasonality, Its Causes and Characteristics . . . . .. . . . . . . . . . . . . . . . . . . . 61
3.2 The Economic Significance of Seasonality
and the Need for Seasonally Adjusted Series . . .. . . . . . . . . . . . . . . . . . . . 63
3.3 Basic Assumptions of Main Seasonal Adjustment Methods .. . . . . . 65
3.3.1 Regression Methods . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 66
3.3.2 Stochastic Model-Based Methods . . . . .. . . . . . . . . . . . . . . . . . . . 71
3.3.3 Linear Smoothing Methods . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 72
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 76
4 Linear Filters Seasonal Adjustment Methods: Census
Method II and Its Variants . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 79
4.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 79
4.1.1 Main Steps to Produce a Seasonally Adjusted Series . . . . . 81
4.2 Basic Properties of the Two-Sided Linear Smoothing
Filters of Census Method II-X11 Variant . . . . . . .. . . . . . . . . . . . . . . . . . . . 83
4.2.1 The Centered 12 Months Moving Average .. . . . . . . . . . . . . . . 83
4.2.2 The 9-, 13-, and 23-Term Henderson Moving
Averages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 84
4.2.3 The Weighted 5-Term (3  3) and 7-Term
(3  5) Moving Averages.. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 84
4.3 Basic Properties of the One-Sided Linear Smoothing
Filters of Census Method II-X11 Variant . . . . . . .. . . . . . . . . . . . . . . . . . . . 85
4.4 The X11ARIMA Method . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 86
4.4.1 General Outline and Basic Assumptions . . . . . . . . . . . . . . . . . . 86
4.4.2 The Forecasting Filters of ARIMA Models
and Their Properties . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 88
4.4.3 Other Main Improvements Incorporated
into the Automated Version of X11ARIMA . . . . . . . . . . . . . . 91
4.5 The X12ARIMA Method . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 95
4.5.1 General Outline .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 95
4.5.2 The General RegARIMA Model .. . . . . .. . . . . . . . . . . . . . . . . . . . 97
4.6 Illustrative Example: X12ARIMA Seasonal
Adjustment of the US NODG Series . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 100
4.6.1 Input: Specification File . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 101
4.6.2 Testing for the Presence of Identifiable Seasonality . . . . . . 103
4.6.3 Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 105
4.6.4 Decomposition .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 110
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 113
5 Seasonal Adjustment Based on ARIMA Model
Decomposition: TRAMO-SEATS . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 115
5.1 TRAMO: Time Series Regression with ARIMA Noise,
Missing Observations, and Outliers . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 116
Contents xiii

5.2 SEATS: Signal Extraction in ARIMA Time Series .. . . . . . . . . . . . . . . . 121

5.3 Illustrative Example: TRAMO-SEATS Seasonal
Adjustment of the US NODG Series . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 126
5.3.1 Input: Specifications .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 127
5.3.2 Testing for the Presence of Identifiable Seasonality . . . . . . 128
5.3.3 Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 128
5.3.4 Decomposition .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 136
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 144
6 Seasonal Adjustment Based on Structural Time Series Models . . . . . . . 147
6.1 Structural Time Series Models.. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 148
6.1.1 Trend Models . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 148
6.1.2 The Cyclical Component .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 150
6.1.3 Seasonality .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 151
6.1.4 Regression Component .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 153
6.2 Linear State Space Models. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 153
6.2.1 The Kalman Filter . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 155
6.2.2 Likelihood Estimation .. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 156
6.2.3 Diagnostic Checking . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 158
6.3 Illustrative Example: Analysis of the US
Unemployment Rate for Males Using STAMP .. . . . . . . . . . . . . . . . . . . . 159
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 163

Part II Trend-Cycle Estimation

7 Trend-Cycle Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 167
7.1 Deterministic Global Trend Models .. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 168
7.2 Stochastic Global Trend Models.. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 170
7.2.1 TRAMO-SEATS Trend Models . . . . . . .. . . . . . . . . . . . . . . . . . . . 172
7.2.2 STAMP Trend Models . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 175
7.3 Stochastic Local Trend-Cycle Models . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 178
7.3.1 Locally Weighted Regression Smoother (LOESS).. . . . . . . 180
7.3.2 Henderson Smoothing Filter . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 181
7.3.3 Gaussian Kernel Smoother .. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 184
7.3.4 Cubic Smoothing Spline . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 185
7.3.5 Theoretical Properties of Symmetric
and Asymmetric Linear Trend-Cycle Filters . . . . . . . . . . . . . . 188
7.4 Illustrative Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 191
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 193
8 Further Developments on the Henderson Trend-Cycle Filter . . . . . . . . . 197
8.1 The Nonlinear Dagum Filter (NLDF) . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 198
8.2 The Cascade Linear Filter. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 201
8.2.1 The Symmetric Linear Filter . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 202
8.2.2 The Asymmetric Linear Filter . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 205
xiv Contents

8.3 The Henderson Filter in the Reproducing Hilbert Space

(RKHS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 210
8.3.1 Linear Filters in Reproducing Kernel Hilbert Spaces .. . . . 211
8.3.2 The Symmetric Henderson Smoother and Its
Kernel Representation .. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 215
8.3.3 Asymmetric Henderson Smoothers and Their
Kernel Representations .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 220
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 222
9 A Unified View of Trend-Cycle Predictors in Reproducing
Kernel Hilbert Spaces (RKHS).. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 225
9.1 Nonparametric Estimators in RKHS . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 226
9.1.1 Polynomial Kernel Regression .. . . . . . . .. . . . . . . . . . . . . . . . . . . . 230
9.1.2 Smoothing Spline Regression .. . . . . . . . .. . . . . . . . . . . . . . . . . . . . 231
9.2 Boundary Behavior .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 235
9.2.1 Empirical Evaluation . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 237
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 240
10 Real Time Trend-Cycle Prediction . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 243
10.1 Asymmetric Filters and RKHS . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 245
10.1.1 Properties of the Asymmetric Filters . .. . . . . . . . . . . . . . . . . . . . 249
10.2 Optimal Bandwidth Selection . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 250
10.3 Empirical Application .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 255
10.3.1 Reduction of Revision Size in Real Time
Trend-Cycle Estimates . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 256
10.3.2 Turning Point Detection .. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 258
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 262
11 The Effect of Seasonal Adjustment on Real-Time
Trend-Cycle Prediction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 263
11.1 Seasonal Adjustment Methods.. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 264
11.1.1 X12ARIMA .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 264
11.1.2 TRAMO-SEATS.. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 268
11.2 Trend-Cycle Prediction in Reproducing Kernel Hilbert
Space (RKHS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 270
11.3 Empirical Application .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 272
11.3.1 Reduction of Revisions in Real Time
Trend-Cycle Estimates . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 275
11.3.2 Turning Point Detection .. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 275
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 277

Glossary . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 279
Acronyms

ACGF Autocovariance generating function

ADR Average duration of run
AICC Akaike information criterion (AIC) with a correction for finite sample
sizes
AMB ARIMA model-based method
AO Additive outlier
AR Autoregressive process
ARMA Autoregressive moving average process
ARIMA Autoregressive integrated moving average process
BIC Bayesian information criterion
CLF Cascade linear filter
ESS European statistical system
GCV Generalized cross validation
H13 13-term symmetric Henderson filter
I(k) Integrated of order k process
IGLS Iterative generalized least squares
i.i.d. Independent and identically distributed
IO Innovation outlier
LEI Index of leading economic indicators
LOESS Locally weighted regression smoother
LOM Length of month
LS Level shift outlier
MA Moving average process
MCD The number of months it takes the average absolute change in the
trend-cycle to dominate that in the irregular
MMSLE Minimum mean square linear estimator
MSE Mean square error
MSAR Markov switching autoregressive process
NBER National Bureau of Economic Research
NLDF Nonlinear Dagum filter
OCV Ordinary cross validation

xv
xvi Acronyms

OLS Ordinary least squares

QCD The number of quarters it takes the average absolute change in the
trend cycle to dominate that in the irregular
regARIMA Autoregressive integrated moving average process with regression
components
RKHS Reproducing kernel Hilbert space
RMSE Root mean square error
SAR Seasonal autoregressive process
SMA Seasonal moving average process
SEATS Signal extraction in ARIMA time series
SIGEX Signal extraction
STAMP Structural time series analyzer, modeler, and predictor
STAR Smooth transition autoregressive process
STL Seasonal trend decomposition procedure based on LOESS
TAR Threshold autoregressive process
TC Temporary change outlier
TRAMO Time series regression with ARIMA noise, missing observations, and
outliers
UC Unobserved component
WK Wiener–Kolmogorov filter
WN.0;  2 / White noise process with zero mean and constant variance  2 , with
uncorrelated components
2  12 m.a. Centered 12-month moving average
3  3 m.a. Weighted five-term moving average
3  5 m.a. Weighted seven-term moving average
Chapter 1
Introduction

Abstract This chapter introduces the main topics of the book that are divided into
two parts: (I) seasonal adjustment methods and (II) real time trend-cycle estimation.
Most socioeconomic series and mainly indicators (leading, lagging, and coincident)
are affected by the impact of seasonal variations. In order to assess what is the
stage of the cycle at which the economy stands it is necessary to identify, estimate,
and remove the effect of seasonality from the data. The seasonal adjustment
methods discussed in Part I are grouped into three main categories: smoothing linear
methods, ARIMA model-based and structural time series models. There are three
seasonal adjustment softwares, one for each method, that are currently available.
Two are officially adopted by statistical agencies, named X12ARIMA and TRAMO-
SEATS, and the third major software, called STAMP, is mainly used in econometric
studies and for academic research. In Part II, the trend-cycle estimates from these
seasonal adjustment methods are analyzed and compared to a new set of trend-
cycle symmetric and asymmetric filters based on Reproducing Kernel Hilbert Space
(RKHS) methodology.

1.1 Book Structure

This book is divided into two parts: one on seasonal adjustment and the other on
real time trend-cycle estimation. It has eleven chapters as follows:
1. Introduction.
2. Time series components.
PART I: SEASONAL ADJUSTMENT.
3. Seasonal adjustment, meaning, purpose, and methods.
4. Linear smoothing or moving averages seasonal adjustment methods.
5. Seasonal adjustment based on ARIMA model decomposition: TRAMO-
SEATS.
6. Seasonal adjustment based on structural time series models.
PART II: REAL TIME TREND-CYCLE ESTIMATION.
7. Trend-cycle estimation.
8. Recent developments on nonparametric trend-cycle estimation.

© Springer International Publishing Switzerland 2016 1

E. Bee Dagum, S. Bianconcini, Seasonal Adjustment Methods and Real Time
Trend-Cycle Estimation, Statistics for Social and Behavioral Sciences,
DOI 10.1007/978-3-319-31822-6_1
2 1 Introduction

9. A unified view of trend-cycle predictors in reproducing kernel Hilbert spaces.

10. Real time trend-cycle estimation.
11. The effect of seasonal adjustment methods on real time trend-cycle estimation.
A summary of the contents of these two parts follows.

1.2 Part I: Seasonal Adjustment Methods

A great deal of data in business, economics, and natural sciences occurs in the
form of time series where observations are dependent, and where the nature of
this dependence is of interest in itself. The time series is generally compiled for
consecutive and equal periods, such as weeks, months, quarters, and years.
From a statistical point of view, a time series is a sample realization of a
stochastic process, i.e., a process controlled by probability laws. The observations
made as the process continues indicate the way it evolves. In the analysis of
economic time series, Persons [48] was the first to distinguish four types of
evolution, namely: (a) trend, (b) cycle, (c) seasonal variations, and (d) irregular
fluctuations.
The trend corresponds to a variation in some defined sense persisting over a long
period of time, that is, a period which is long in relation to the cycle. In some cases,
it is a steady growth; while in others, the trend may move downward as well as
upward. The cycle, usually referred to as the business cycle, is a quasi-periodic
oscillation characterized by alternating periods of expansion and contraction. The
seasonal variations represent the composite effect of climatic and institutional events
which repeat more or less regularly each year. These three types of fluctuation are
assumed to follow systematic patterns (they are the signal of the process).
On the other hand, the irregulars are unforeseeable movements related to events
of all kinds. In general, they have a stable random appearance but, in some
series, extreme values may be present. These extreme values or outliers have
identifiable causes, e.g., floods, unseasonal weather, and strikes; and, therefore, can
be distinguished from the much smaller irregular variations.
Among all these components, the influence of the seasonal fluctuations in the
human activity has been felt from earlier times. The organization of society, the
means of production and communication, the habits of consumption, and other
social and religious events have been strongly conditioned by both climatic and
conventional seasons. The seasonal variations in agriculture, the high pre-Easter and
pre-Christmas retail sales, and the low level of winter construction are all general
knowledge.
The main causes of seasonality, such as climatic and institutional factors, are
mainly exogenous to the economic system and cannot be controlled or modified
by the decision makers in the short run. The impact of seasonality in the economic
activity, however, is usually not independent on the stages of the business cycle. It
is well-known for example, that the seasonal unemployment among adult males is
1.2 Part I: Seasonal Adjustment Methods 3

much larger during periods of recession than during periods of prosperity. Another
main feature of seasonality is that the phenomenon repeats with certain regularity
every year but it may evolve. The latter is mainly due to systematic intervention
of the government and to technological and institutional changes as well; and it is
more the general case than the exception for economic time series. Therefore, the
assumption of stable seasonality, i.e., of seasonality being represented by a strictly
periodic function, can be used as a good approximation for few series only. In effect,
even in the extreme cases of those activities where seasonality is mainly caused
by climatic conditions, e.g., agriculture, fishing, forestry, the seasonal variations
change, for weather itself measured by the temperature, and quantity of precipitation
changes.
Once the assumption of stable seasonality is abandoned, new assumptions must
be made regarding the nature of its evolution; if seasonality changes, is it slowly or
rapidly? Is it gradually or abruptly? Is it in a deterministic or a stochastic manner?
Today, the most widely accepted hypothesis is that seasonality moves gradually,
slowly, and in a stochastic manner. A third characteristic of seasonality is that the
phenomenon can be separated from other forces (trend, cycle, and irregulars) that
influence the movement of an economic time series and can be measured.
The seasonal variations are distinguished from trend by their oscillated character,
from the cycle by being confined within the limits of an annual period, and from the
irregulars, by the fact of being systematic. Most socioeconomic indicators, such as
leading, coincident, and lagging, are affected by the impact of seasonal variations.
In order to assess what is the stage of the current economic conditions it is necessary
to identify, estimate, and remove the effect of seasonality. Seasonal adjustment
methods can be grouped into three main categories: smoothing linear methods,
ARIMA model-based and structural time series models. There are three seasonal
adjustment softwares, one for each method, that are currently available but only
two are officially adopted by statistical agencies, named X12ARIMA and TRAMO-
SEATS. The third major software, called STAMP, is mainly used in econometric
studies and for academic research. These topics are largely discussed in Chaps. 4, 5,
and 6, respectively, and only brief summaries are given below.

1.2.1 Smoothing Linear Seasonal Adjustment Methods

The best known and most often applied seasonal adjustment methods are based on
smoothing linear filters or moving averages applied sequentially by adding (and
subtracting) one observation at a time. These methods assume that the time series
components change through time in a stochastic manner. Given a time series yt ; t D
1; : : : ; n, for any t far removed from both ends, say m < t < n  m, the seasonally
adjusted value yat is obtained by applying, in a moving manner, a symmetric average
4 1 Introduction

W.B/, that is,

X
m X
m
yat D W.B/yt D wj Bj yt D wj ytj ; (1.1)
jDm jDm

where the weights wj are symmetric, that is, wj D wj , and the length of the average
is 2m C 1. For current and recent data (n  m < t  n), a symmetric linear filter
cannot be applied, and therefore truncated asymmetric filters are used. For example,
for the last available observation yn , the seasonally adjusted value is given by

X
m
yan D W0 .B/yn D w0;j ynj : (1.2)
jD0

The asymmetric filters are time-varying in the sense that different filters are applied
to the m first and last observations. The end estimates are revised as new observa-
tions are added because of: (a) the new innovations and (b) the differences between
the symmetric and asymmetric filters. The estimates obtained with symmetric filters
are often called “final.”
The development of electronic computers contributed to major improvements
in seasonal adjustment based on moving averages and facilitated their massive
application. In 1954, Julius Shiskin of the US Bureau of Census developed a
software called Method I, based mainly on the works of Macauley [44] already
being used by the US Federal Reserve Board. Census Method I was followed by
Census Method II and 11 more experimental versions (X1, X2, : : :, X11). The best
known and widely applied was the Census Method II-X11 variant developed by
Shiskin et al. [51], but produced poor seasonally adjusted data at the end of the
series which is of crucial importance to assess the direction of the short-term trend
and the identification of real time turning points in the economy.
Estela Bee Dagum [13] working at Statistics Canada developed the X11ARIMA
mainly to correct for this serious limitation and, later, Findley et al. [26] developed
X12ARIMA that offers a regARIMA option to estimate deterministic components,
such as trading day variations, moving holiday effects and outliers, simultaneously
with the ARIMA model for extrapolation. It also includes new diagnostic tests and
spectral techniques to assess the goodness of the results. The X12ARIMA method
is today the one most often applied by statistical agencies in the world. The US
Bureau of Census continued research on the development of seasonal adjustment
methods, and recently produced a version called X13ARIMA-SEATS which enables
the estimation of the seasonal component either via linear filters as those available
in X12ARIMA or based on an ARIMA decomposition model.
All the seasonal adjustment methods based on moving averages discussed in this
book have nonlinear elements. Hence, the seasonally adjusted total of an aggregated
series is not equal to the algebraic sum of the seasonally adjusted series that enter
1.2 Part I: Seasonal Adjustment Methods 5

into the aggregation. The main causes of nonlinearity are generated by:
1. a multiplicative decomposition model for the unobserved components,
2. the identification and replacement of outliers,
3. the ARIMA extrapolations, and
4. automatic selection of the moving average length for the estimation of the trend-
cycle and seasonality.
The properties of the combined linear filters applied to estimate the various
components were originally calculated by Young [54] for the standard option
of Census II-X11 variant. Later, Dagum et al. [18] calculated and analyzed all
possible filter combinations of Census II-X11 and X11ARIMA. Given the fact that
X11ARIMA and X12ARIMA are methods based on the Census Method II-X11
[51], widely discussed in Chap. 4, a brief summary is given here. The reader is
referred to Ladiray and Quenneville [41] for a very detailed description of the X11
method.
The X11 method assumes that the main components of a time series follow a
multiplicative or an additive model, that is,
(multiplicative model) yt D TCt  St  It ,
(additive model) yt D TCt C St C It ,
where yt stands for the original series, TCt for the trend-cycle, St for the seasonality,
and It for the irregulars.
There are no mixed models in this program, such as yt D TCt  St C It or other
possible combinations. The estimation of the components is made with different
kinds of smoothing linear filters. The filters are applied sequentially and the whole
process is repeated twice.
The X11 variant produced seasonal factor forecasts for each month which were
given by

1
Sj;tC1 D Sj;t C .Sj;t  Sj;t1 /; j D 1; : : : ; 12;
2
where j is the month and t the current year. The use of seasonal factor forecasts was
very popular in statistical bureaus till the end of the 1970s. In 1975, Dagum [12]
proposed the use of concurrent seasonal factors obtained from data that included
the most recent values and this was adopted by Statistics Canada. Gradually, other
statistical agencies applied concurrent seasonal factors and this is now the standard
practice.
The final trend-cycle is obtained by a 9-, 13-, or 23-term Henderson moving
average applied to the final monthly seasonally adjusted series (a 5- or 7-term
Henderson filter for quarterly data). The selection of the appropriate filter is made
on the basis of a preliminary estimate of the I=C ratio (the ratio of the average
absolute month-to-month change in the irregular to that in the trend-cycle). A
9-term is applied to less irregular series and a 23-term to highly irregular series.
The linear smoothing filters applied by Census Method II-X11 variant to produce
6 1 Introduction

seasonally adjusted data can be classified according to the distribution of their set
of weights in symmetric (two-sided) and asymmetric (one-sided). The symmetric
moving averages are used to estimate the component values that fall in the middle
of the span of the average, say 2m C 1, and the asymmetric moving averages, to the
m first and last observations.
The sum of the weights of both kinds of filters is one, and thus, the mean of
the original series is unchanged in the filtering process. The sum of the weights of
a filter determines the ratio of the mean of the smoothed series to the mean of the
unadjusted series assuming that these means are computed over periods long enough
to ensure stable results. It is very important in filter design that the filter does not
displace in time the components of the output relative to those of the input. In other
words, the filter must not introduce phase shifts.
Symmetric moving averages introduce no time displacement for some of the
components of the original series and a displacement of 180ı for others. A phase
shift of 180ı is interpreted as a reverse in polarity which means that maxima are
turned into minima and vice versa. In other words, peaks (troughs) in the input
are changed into troughs (peaks) in the output. For practical purposes, however,
symmetric moving averages act as though the time displacement is null. This is so
because the sinusoids that will have a phase shift of 180ı in the filtering process are
cycles of short periodicity (annual or less) and moving averages tend to suppress or
significantly reduce their presence in the output. In spectral analysis, the phase is a
dimensionless parameter that measures the displacement of the sinusoid relative to
the time origin. Because of the periodic repetition of the sinusoid, the phase can be
restricted to ˙180ı . The phase is a function of the frequency of the sinusoid, being
the frequency equal to the reciprocal of the length of time or period required for one
complete oscillation.
It is inherent to any moving average procedure that the m first and last points
of an original series cannot be smoothed with the same set of symmetric 2m C 1
weights applied to the middle values.
The X11 method uses one-sided filters to smooth these end points. Moving
averages with asymmetric weights are bounded to introduce phase shifts for all the
components of the original series. This is a very undesirable characteristic since it
may cause, for example, a business cycle in the smoothed series to turn down before
or after it actually does in the unadjusted data. The asymmetric weights associated
with the Henderson filters no longer estimate a cubic within their span, but only a
linear trend-cycle within the length of the filters.
The X11ARIMA is a modified version of the Census Method II-X11 variant that
was developed by Estela Bee Dagum, at Statistics Canada, mainly to produce a
better current seasonal adjustment of series with seasonality that changes rapidly in
a stochastic manner. The latter characteristic is often found in main socioeconomic
indicators, e.g., retail trade, imports and exports, unemployment, and so on.
A detailed description of the X11ARIMA method of seasonal adjustment is given
1.2 Part I: Seasonal Adjustment Methods 7

in [13] and [14]. The X11ARIMA method consists of:

(i) modeling the original series with an ARIMA model of the Box and Jenkins
type,
(ii) extending the original series 1–3 years with forecasts from the ARIMA model
that fits and extrapolates well according to well-defined acceptance criteria,
and
(iii) estimating each component with moving averages that are symmetric for
middle observations and asymmetric for both end years. The latter are obtained
via the convolution of the Census II-X11 variant filters and the ARIMA model
extrapolations.
For series that result from the accumulation of daily activities, called flow series,
deterministic components such as trading day variations and Easter holiday effects
are estimated with dummy variable regression models and removed from the series,
so that only the remainder is subject to steps (i)–(iii) above. The X11ARIMA was
again modified by Dagum [14]. ARIMA models of the Box and Jenkins type were
chosen which were previously shown to be very effective for forecasting a large
number of series [46, 50].
The ARIMA models applied to seasonal series belong to the general multiplica-
tive Box and Jenkins type, that is,

p .B/˚P .Bs /.1  B/d .1  Bs /D yt D q .B/Q .Bs /at (1.3)

where s denotes the periodicity of the seasonal component (equal to 12 for monthly
series and 4 for quarterly data), B is the backshift operator, such that Byt D yt1 ,
p .B/ D .1  1 B      p Bp / is the nonseasonal autoregressive (AR) operator of
order p, ˚P .Bs / D .1  ˚1 Bs      ˚P BPs / is the seasonal autoregressive (SAR)
operator of order P, q .B/ D .1  1 B      q Bq / is the nonseasonal moving
average (MA) operator of order q, Q .Bs / D .11 Bs   Q BQs / is the seasonal
moving average (SMA) operator of order Q, and the at ’s are i.i.d. with mean zero
and variance a2 (white noise). The .1  B/d .1  Bs /D term implies nonseasonal
differencing of order d and seasonal differencing of order D. If d D D D 0 (no
differencing), it is common to replace yt in (1.3) by deviations from its mean. The
general multiplicative model (1.3) is said to be of order .p; d; q/.P; D; Q/s.
The X12ARIMA is today the most often applied seasonal adjustment method by
statistical agencies. It was developed by Findley et al. [26], and it is an enhanced
version of the X11ARIMA method.
The major modifications concern: (1) extending the automatic identification and
estimation of ARIMA models for the extrapolation option to many more than the
three models available in X11ARIMA and (2) estimating trading day variations,
moving holidays, and outliers in what is called regARIMA.
RegARIMA consists of regression models with autoregressive integrated moving
average (ARIMA) errors. More precisely, they are models in which the mean
function of the time series (or its logs) is described by a linear combination of
regressors, and the covariance structure of the series is that of an ARIMA process.
8 1 Introduction

If no regressors are used, indicating that the mean is assumed to be zero, the
regARIMA model reduces to an ARIMA model.
Whether or not special problems requiring the use of regressors are present in
the series to be adjusted, a fundamentally important use of regARIMA models is
to extend the series with forecasts (and backcasts) in order to improve the seasonal
adjustments of the most recent (and the earliest) data. Doing this reduces problems
inherent in the trend-cycle estimation and asymmetric seasonal averaging processes
of the type used by the X11 method near the ends of the series. The provision of
this extension was the most important improvement offered by the X11ARIMA
program. Its theoretical and empirical benefits have been documented in many
publications, such as Dagum [14], Bobbit and Otto [5], and the references therein.
The X12ARIMA method has all the seasonal adjustment capabilities of the
X11ARIMA variant. The same seasonal and trend moving averages are available,
and the program still offers the X11 calendar and holiday adjustment routines
incorporated in X11ARIMA. But several important new options have been included.
The modeling module is designed for regARIMA model building with seasonal
socioeconomic time series. To this end, several categories of predetermined regres-
sion variables are available, including trend constants or overall means, fixed
seasonal effects, trading day effects, holiday effects, pulse effects (additive outliers),
level shifts, temporary change outliers, and ramp effects. User-specified regression
variables can also be included in the models.
The specification of a regARIMA model requires specification of both the regres-
sion variables to be included in the model and the type of ARIMA model for the
regression errors (i.e., the order .p; d; q/.P; D; Q/s ). Specification of the regression
variables depends on the user knowledge about the series being modeled. Identi-
fication of the ARIMA model for the regression errors follows well-established
procedures based on examination of various sample autocorrelation and partial
autocorrelation functions produced by the X12ARIMA program. Once a regARIMA
model has been specified, X12ARIMA estimates its parameters by maximum like-
lihood using an Iterative Generalized Least Squares (IGLS) algorithm. Diagnostic
checking involves examination of residuals from the fitted model for signs of model
inadequacy. X12ARIMA produces several standard residual diagnostics for model
checking, as well as provides sophisticated methods for detecting additive outliers
and level shifts.
Trading day effects occur when a series is affected by the different day-of-the-
week compositions of the same calendar month in different years. Trading day
effects can be modeled with seven variables that represent (no. of Mondays), . . . ,
(no. of Sundays) in month t. Bell and Hillmer [2] proposed a better parameterization
of the same effects using six variables defined as (no. of Mondays)  (no. of
Sundays), . . . , (no. of Saturdays)  (no. of Sundays), along with a seventh variable
for length of month (LOM) or its deseasonalized version, the leap-year regressor
(lpyear). In X12ARIMA the six variables are called the tdnolpyear variables.
Instead of using a seventh regressor, a simpler and often better way to handle
multiplicative leap-year effects is to re-scale the February values of the original
time series before transformation to m N Feb yt =mt , where yt is the original time series
1.2 Part I: Seasonal Adjustment Methods 9

before transformation, mt is the length of month t (28 or 29), and m N Feb D 28:25 is
the average length of February. If the regARIMA model includes seasonal effects,
these can account for the length-of-month effect except in Februaries, so the trading
day model only has to deal with the leap-year effect. When this is done, only the
tdnolpyear variables need be included in the model. X12ARIMA allows explicit
choice of either approach, as well as an option (td) that makes a default choice
of how to handle length-of-month effects. When the time series being modeled
represents the aggregation of some daily series (typically unobserved) over calendar
months they are called monthly flow series. If the series instead represents the value
of some daily series at the end of the month, called a monthly stock series, then
different regression variables are appropriate.
Holiday effects in a monthly flow series arise from holidays whose dates vary
over time if: (1) the activity measured by the series regularly increases or decreases
around the date of the holiday and (2) this affects 2 (or more) months depending on
the date the holiday occurs each year. Effects of holidays with a fixed date, such as
Christmas, are indistinguishable from fixed seasonal effects. Easter effects are the
most frequently found holiday effects in American and European economic time
series, since the date of Easter Sunday varies between March 22 and April 25. Labor
Day and Thanksgiving also are found in American and Canadian time series.
X12ARIMA provides four other types of regression variables to deal with abrupt
changes in the level of a series of temporary or permanent nature: additive outliers
(AO), level shifts (LS), temporary changes (TC), and ramps. Identifying the location
and nature of potential outliers is the object of the outlier detection methodology
implemented. This methodology can be used to detect AOs, TCs, and LSs (not
ramps). Any outlier detected is automatically added to the model as regression
variable. Prespecified AOs, LSs, TCs, and ramps are actually simple forms of
interventions as discussed by Box and Tiao [6].
Figure 1.1 shows the original values of the time series US Unemployment Rate
for Males (16 years and over) observed from January 1992 to December 2013
together with the final seasonally adjusted series obtained with the X12ARIMA
software using the default option. Figure 1.2 exhibits the corresponding seasonal
factors.

1.2.2 ARIMA Model-Based Seasonal Adjustment Method

Peter Burman [7] was the first to develop a seasonal adjustment method based on
ARIMA model decomposition, named SIGEX (Signal Extraction). Later, working
on the same topic Hillmer and Tiao [35] developed what is known as ARIMA
model-based seasonal adjustment, largely discussed in Bell and Hillmer [2].
An ARIMA model is identified from the observed data and by imposing certain
restrictions, models for each component are derived. Since the components are
unknown, to obtain a unique solution Hillmer and Tiao proposed a canonical
decomposition which has the property of maximizing the variance of the irregulars
10 1 Introduction

Original series
12
10
8
6
4 X12ARIMA seasonally adjusted series

1995 2000 2005 2010

Fig. 1.1 Original and X12ARIMA (default option) seasonally adjusted US Unemployment Rate
for Males (16 years and over) observed from January 1992 to December 2013
115
105
95
90

1995 2000 2005 2010

Fig. 1.2 Seasonal factors of the US Unemployment Rate for Males (16 years and over) estimated
using the X12ARIMA software (default option)
1.2 Part I: Seasonal Adjustment Methods 11

and minimizing the variance of the estimated components. Because ARIMA

model identification and estimation are not robust to outliers, and cannot deal
with deterministic components such as trading days and moving holidays, further
changes were made by combining dummy variables regression models with ARIMA
models. In this regard, Gomez and Maravall [28] developed at the Bank of Spain,
a seasonal adjustment software called TRAMO-SEATS which is currently applied
mainly by European statistical agencies.
TRAMO stands for “Time series Regression with ARIMA noise, Missing
observations and Outliers” and SEATS for “Signal Extraction in ARIMA Time
Series.” First, TRAMO estimates via regression the deterministic components,
which are later removed from the raw data. In the second round, SEATS estimates
the seasonal and trend-cycle components from the ARIMA model fitted to the data
where the deterministic components are removed. SEATS uses the filters derived
from the linearized ARIMA model that describes the stochastic behavior of the time
series. It should be mentioned that Eurostat, in collaboration with the National Bank
of Belgium, the US Bureau of the Census, the Bank of Spain, and the European
Central Bank, has developed an interface of TRAMO-SEATS and X12ARIMA
called Demetra+. In the Bank of Spain and Eurostat websites, it is also possible to
find a considerable number of papers relevant to TRAMO-SEATS as well as in the
European Statistical System (ESS) Guideline. TRAMO is a regression method that
performs the estimation, forecasting, and interpolation of missing observations with
ARIMA errors. The ARIMA model can be identified automatically or by the user.
Given the vector of observations y D .y1 ; : : : ; yn /0 , the program fits the regression
model

yt D x0t ˇ C t (1.4)

where ˇ D .ˇ1    ˇp /0 is a vector of regression coefficients, xt D .x1t    xpt /0 are

p regression variables that define the deterministic part of the model. On the other
hand, t represents the stochastic part of the model, that is assumed to follow the
general ARIMA model

 .B/t D .B/at ; (1.5)

where B is the backshift operator,  .B/ and .B/ are finite polynomials in B, and at
is assumed to be normally identically distributed, that is, NID(0, a2 ). The polyno-
mial  .B/ contains the unit roots associated with regular and seasonal differencing,
as well as the polynomial with stationary autoregressive roots (and complex unit
roots, if present). .B/ denotes the invertible moving average polynomial. The
regression variables xt can be given by the user or generated by the program. In the
latter case, the variables are for trading day variations, Easter effects, and outliers.
Outliers reflect the effect of some special, non regularly repetitive events, such
as implementation of a new regulation, major political or economic changes,
modifications in the way the variable is measured, occurrence of a strike or natural
disaster, etc. Consequently, discordant observations and various types of abrupt
12 1 Introduction

changes are often present in times series data. The location and type of outliers are
“a priori” unknown. TRAMO uses an improved Chen and Liu [8] type procedure
for outlier detection and correction. The effect of an outlier is modeled by yt D
! .B/It .t0 / C t , where .B/ is a quotient of polynomials in B that models the type
of outlier (its effect over time), and It .t0 / is an indicator variable of the occurrence
of the outlier, that is,

1 if t D t0
It .t0 / D
0 otherwise

! represents the impact of the outlier at time t0 and t is the outlier free series which
follows the model specified in Eq. (1.5). In the automatic detection and correction,
by default, three types of outliers can be considered
(1) Additive Outlier (AO): .B/ D 1;
(2) Transitory Change (TC): .B/ D 1=.1  ıB/, where, by default, ı D 0:7;
(3) Level Shift (LS): .B/ D 1=.1  B/.
One can also include a fourth type of outlier, that is, Innovational Outlier (IO) for
which .B/ D .B/=.B/ı.B/ that resembles a shock in the innovations at .
The procedure followed by TRAMO consists of:
1. Pretest for the log/level specification, that is, a trimmed range mean regression
test is performed to select whether the original series will be transformed into log
or maintain the level.
2. Pretest for trading days and Easter effects made with regressions using the
default model for the noise. If the model is subsequently changed, the test is
redone. Thus, the output file of TRAMO may say at the beginning “Trading
day is not significant,” but the final model estimated may contain trading day
variables (or vice versa).
3. Automatic detection and correction of outliers. The program has a facility for
detecting outliers and for removing their effects. The outliers can be entered by
the user or they can be automatically identified by the program.
4. Automatic model selection. The program further performs an automatic identi-
fication of the ARIMA model. This is done in two steps. The first one yields
the nonstationary polynomial .1  B/d .1  Bs /D of model (1.5). This is done by
iterating on a sequence of AR and ARIMA models (with constant) which have
a multiplicative structure when the data is seasonal. The procedure is based on
results of Tiao and Tsay [52] and Tsay [53]. Regular and seasonal differences
are obtained up to a maximum order of .1  B/2 .1  Bs /. The second step
identifies an ARMA model for the stationary series (modified for outliers and
regression-type effects) following the Hannan–Rissanen procedure [31], with
an improvement which consists of using the Kalman filter instead of zeros to
calculate the first residuals in the computation of the estimator of the variance
of the innovations of model (1.5). Finally, the program combines the facilities
for automatic identification and correction of outliers and automatic ARIMA
1.2 Part I: Seasonal Adjustment Methods 13

model identification just described so that it has an option for automatic model
identification of a nonstationary series in presence of outliers.
5. Diagnostic checks. The main diagnostic tests are on the residuals. The estimated
residuals aO t are analyzed to test the hypothesis that at are independent and
identically normally distributed with zero mean and constant variance a2 .
Besides inspection of the residual graph, the corresponding sample autocorre-
lation function is examined. The lack of residual autocorrelation is tested using
the Ljung–Box test statistics, and skewness and kurtosis tests are applied to
test for normality of the residuals. Specific out-of-sample forecast tests are also
performed to evaluate if forecasts behave in agreement with the model.
6. Optimal forecasts. If the diagnostics are satisfied, the model is used to compute
optimal forecasts for the series, together with their mean square error (MSE).
These are obtained using the Kalman filter applied to the original series.
TRAMO is to be used as a preadjustment process to eliminate all the determinis-
tic components such that the residual is a linearized series that can be modeled with
an ARIMA process. The latter is decomposed by SEATS in stochastic trend-cycle,
seasonality, and irregulars. Both programs can handle routine applications to a large
number of series and provide a complete model-based solution to the problems of
forecasting, interpolation, and signal extraction for nonstationary time series.
SEATS belongs to the class of procedures based on ARIMA models for the
decomposition of time series into unobserved components and consists of the
following steps:
1. ARIMA model estimation. SEATS starts by fitting an ARIMA model to a series
not affected by deterministic components, such as trading day variations, moving
holidays, and outliers. Let yt denote this linearized series, and consider an
additive decomposition model (multiplicative if applied to the log transformation
of yt ). The complete model can be written in detailed form as

p .B/˚P .Bs /.1  B/d .1  Bs /D yt D q .B/Q .Bs /at C c; (1.6)

and, in concise form, as

 .B/yt D .B/at C c; (1.7)

where c is equal to  .B/Ny, being yN the mean of the linearized series yt . In

words, the model that SEATS assumes is that of a linear time series with
Gaussian innovations. When used with TRAMO, estimation of the ARIMA
model is made by the exact maximum likelihood method described by Gomez
and Maravall [27].
2. Derivation of the ARIMA models for each component. The program proceeds
by decomposing the series that follows the ARIMA model (1.7) into several
components. The decomposition can be multiplicative or additive. Next, we shall
discuss the additive model, since the multiplicative form is equivalent in the log
14 1 Introduction

transformation of the data. That is,

yt D Tt C Ct C St C It ; (1.8)

where Tt denotes the trend component, Ct the cycle, St represents the seasonal
component, and It the irregulars.
The decomposition is done in the frequency domain. The spectrum (or
pseudospectrum) is partitioned into additive spectra, associated with the different
components which are determined mainly from the AR roots of the model. The
trend component represents the long-term evolution of the series and displays
a spectral peak at frequency 0, whereas the seasonal component captures the
spectral peaks at seasonal frequencies (e.g., for monthly data these are 0.524,
1.047, 1.571, 2.094, 2.618, and 3.142). The cyclical component captures periodic
fluctuations with period longer than a year, associated with a spectral peak
for a frequency between 0 and (2=s), and short-term variation associated
with low order MA components and AR roots with small moduli. Finally, the
irregular component captures white noise behavior, and hence has a flat spectrum.
The components are determined and fully derived from the structure of the
(aggregate) ARIMA model (1.7) for the linearized series directly identified from
the data. The program is aimed at monthly or quarterly frequency data and the
maximum number of observations that can be processed is 600.
One important assumption is that of orthogonality among the components, and
each one will have in turn an ARIMA model. In order to identify the components,
the canonical decomposition is used which implies that the variance of the
irregulars is maximized, whereas the trend, seasonal, and cycle are as stable as
possible.
The canonical condition on the trend, seasonal, and cyclical components
identifies a unique decomposition, from which the ARIMA models for the
components are obtained (including the component innovation variances).
Figure 1.3 shows the original values of the time series US Unemployment Rate for
Males (16 years and over) observed from January 1992 to December 2013 together
with the final seasonally adjusted series obtained with the TRAMO-SEATS software
using the default option. Figure 1.4 exhibits the corresponding seasonal factors.

1.2.3 Structural Time Series Models

The structural time series approach involves decomposing a series into components
which have a direct interpretation. A structural model, also called unobserved
component (UC) model, consists of a number of stochastic linear processes that
stand for the trend, cycle, seasonality, and remaining stationary dynamic features
in an observed time series. The trend component typically represents the longer
1.2 Part I: Seasonal Adjustment Methods 15

12
Original series
TRAMO-SEATS seasonally adjusted series
10
8
6
4

1995 2000 2005 2010

Fig. 1.3 Original and TRAMO-SEATS (default option) seasonally adjusted US Unemployment
Rate for Males (16 years and over) observed from January 1992 to December 2013
115
105
95
90

1995 2000 2005 2010

Fig. 1.4 Seasonal factors of the US Unemployment Rate for Males (16 years and over) estimated
using the TRAMO-SEATS software (default option)

term movement of the series and is often specified as a smooth function of time.
The recurring but persistently changing patterns within the years are captured by the
seasonal component. In economic time series, the cycle component can represent the
16 1 Introduction

dynamic features associated with the business cycle. Each component is represented
as an ARIMA process. Hence, the typical parameters that need to be estimated are
the variances of the innovations driving the components, and a selection of other
coefficients associated with the ARIMA processes.
The key to handling structural time series models is the state space form with
the state of the system representing the various unobserved components such
as trend, cycle, and seasonality. The estimate of the unobservable state can be
updated by means of a filtering procedure as new observations become available.
Predictions are made by extrapolating these estimated components into the future,
while smoothing algorithms give the best estimate of the state at any point within the
sample. The statistical treatment can therefore be based on the Kalman filter and its
related methods. A detailed discussion of the methodological and technical concepts
underlying structural time series models is given in Chap. 6. The reader is also
referred to the monographs by Harvey [32], Kitagawa and Gersh [39], and Kim and
Nelson [37]. Durbin and Koopman [24] also provide an exhaustive overview on state
space methods for time series, whereas an introduction is given by Commandeur and
Koopman [11].
We now introduce the basic structural time series model with explicit specifica-
tions for each component. For quarterly or monthly observations yt ; t D 1; : : : ; n, it
is given by

yt D Tt C St C It ; (1.9)

where Tt stands for the trend, St for seasonality, and It is the irregular component
generally assumed to be NID.0; I2 /. All these components are stochastic and the
disturbances driving them are assumed to be mutually uncorrelated. The definitions
of the components are given below, but a full explanation of the underlying rationale
can be found in Harvey [32]. The effectiveness of structural time series models
compared to ARIMA-type models is discussed in Harvey et al. [33].
The trend component can be specified in many different ways, but the most
common specification is an extension of the random walk trend by including a
stochastic drift component

TtC1 D Tt C ˇt C aTt (1.10)

ˇtC1 D ˇt C
t
t  NID.0; 
2 /; (1.11)

where the disturbance series aTt is normally independently distributed. The initial
values T1 and ˇ1 are treated as unknown coefficients. When Tt is given by (1.10),
Harvey [32] defines yt D Tt C It as the local linear trend model. In case 
2 D 0, the
trend (1.10) reduces to an I.1/ process given by TtC1 D Tt Cˇ1 CaTt , where the drift
ˇ1 is fixed. This specification is referred as a random walk plus drift process. If in
addition T2 D 0, the trend reduces to the deterministic linear trend TtC1 D T1 C ˇ1 t.
When T2 D 0 and 
2 > 0, the trend Tt in (1.10) remains an I.2/ process.
Common specifications for the seasonal component St are provided in the
following.
1.2 Part I: Seasonal Adjustment Methods 17

Dummy seasonal. Since the seasonal effects should sum to zero over a year, a
basic model for this component is given by

X
s1
St D  Stj C !t t D s; : : : ; n; (1.12)
jD1

where s denotes the number of “seasons” in a year. In words, the seasonal effects are
allowed to change over time by letting their sum over the previous year be equal to a
random disturbance term !t with mean zero and variance !2 . Writing out Eq. (1.12)
in terms of the lag operator B gives

.1 C B C    C Bs1 /St D .B/St D !t t D s; : : : ; n: (1.13)

However, since .1  Bs / D .1 C B C    C Bs1 /.1  B/ D .B/.1  B/, the

model can also be expressed in terms of the seasonal difference operator as .1 
Bs /St D .1  B/!t . The normally distributed disturbance !t drives the changes in
the seasonal effect over time and is serially and mutually uncorrelated with all other
disturbances and for all time periods. In the limiting case where !2 D 0 for all t,
the seasonal effects are fixed over time and are specified as a set of unknown fixed
dummy coefficients that sum up to zero.
Trigonometric seasonal. Alternatively, a seasonal pattern can also be modeled
by a set of trigonometric terms at the seasonal frequencies, j D 2j=s, j D
1; : : : ; Œs=2 , where Œs=2 is equal to s=2 if s is even, and .s  1/=2 if s is odd.
The seasonal effect at time t is then described as
Œs=2
X
St D .Sj cos j t C Sj sin j t/: (1.14)
jD1

When s is even, the sine term disappears for j D s=2, and so the number of
trigonometric parameters, the Sj and Sj , is always .s1/=2, which is the same as the
number of coefficients in the seasonal dummy formulation. A seasonal pattern (1.14)
is the sum of [s=2] cyclical components that can be reformulated as follows:
!   ! !
Sj;t cos j sin j Sj;t1 !j;t
 D  C  (1.15)
Sj;t  sin j cos j Sj;t1 !j;t

with !j;t and !j;t ; j D 1; : : : ; Œs=2 , being zero mean white noise processes which are
uncorrelated with each other and have a common variance !2 . Note that, when s is
even, the component at j D s=2 collapses to

Sj;t D Sj;t1 cos j C !jt :

The statistical treatment of structural time series models is based on the

corresponding state space representation, according to which the observations are
assumed to depend linearly on a state vector that is unobserved and generated by a
18 1 Introduction

stochastic time-varying process. The observations are further assumed to be subject

to a measurement error that is independent on the state vector. The state vector can
be estimated or identified once a sufficient set of observations becomes available.
The state space form provides a unified representation of a wide range of linear
time series models, as discussed by Harvey [32], Kitagawa and Gersch [39], and
Durbin and Koopman [24]. The general linear state space model for a sequence of
n observations, y1 ; : : : ; yn , is specified as follows:

yt D zTt ˛t C It ; t D 1; : : : ; n (1.16)
˛tC1 D  t ˛t C Rt "t : (1.17)

Equation (1.16) is called the observation or measurement equation which relates

the observations yt ; t D 1; : : : ; n, to the state vector ˛t through zt , that is, an m  1
vector of fixed coefficients. In particular, ˛t is the m  1 state vector, that contains
the unobserved trend Tt , cycle Ct , and the seasonal component St . The irregular
component It is generally assumed to follow a white noise process with zero mean
and variance I2t .
On the other hand, Eq. (1.17) is called the state or transition equation, where the
dynamic evolution of the state vector ˛t is described through the fixed matrix  t of
order m  m. "t is an r  1 vector of disturbances which are assumed to follow a
multivariate white noise process with zero mean vector and covariance matrix ˙ "t .
It is assumed to be distributed independently on It at all time points. The matrix Rt
is an m  r selection matrix with r < m, that in many standard cases is the identity
matrix Im , being generally r D m. Indeed, although matrix Rt can be specified
freely, it is often composed of a selection from the first r columns of the identity
matrix Im .
Initial conditions have to be defined for the state vector at the first time point, ˛1 .
It is generally assumed to be generated as ˛1  N.a1 ; P1 /, and independently on
the observation and state disturbances, It and "t . The mean vector a1 and covariance
matrix P1 can be treated as given and known in almost all stationary processes for
the state vector. For nonstationary processes and in presence of regression effects in
the state vector, the associated elements in the initial mean vector a1 can be treated as
unknown and estimated. For an extensive discussion of initialization in state space
analysis, we refer to Durbin and Koopman [24].
By appropriate choices of ˛t ; It , and "t , of the matrices zt ;  t , Rt , and of the
scalar I2t , a wide range of different structural time series models can be introduced.
Consider the basic structural time series model yt D Tt C St C It with the trend
component specified through Eqs. (1.10) and (1.11), and St as in (1.12) in presence
of quarterly data (s D 4). A state vector of five elements and a disturbance vector
1.2 Part I: Seasonal Adjustment Methods 19

of four elements are required, and they are given by

0 1
Tt
Bˇ C 0 1
B t C "Tt
B C
˛t D B S1;t C ; "t D @
t A :
B C
@ S2;t A !t
S3;t

The state space formulation of the basic decomposition model is given by

Eqs. (1.16) and (1.17) with the system matrices
0 1 0 1
10 0 0 0 1
B0 1 0 0 0 C B0C
B C B C
B C B C
 t D B0 0 1 1 1 C ; zt D B 1 C :
B C B C
@0 0 1 0 0 A @0A
00 0 1 0 0

Formal proofs of the Kalman filter can be found in Anderson and Moore [1], Harvey
[32], and Durbin and Koopman [24].

1.2.3.1 Regression Component

The model (1.9) may provide a good description of the time series, although it
may sometimes be necessary to include additional components. Seasonal economic
time series are often affected by trading day effects and holiday effects which can
influence the dynamic behavior of the series. Hence, a set of explanatory variables
need to be included in the model to capture specific (dynamic) variations in the time
series, as well as outliers and breaks. Therefore, Koopman and Ooms [40] suggest
to extend model (1.9) as follows:

yt D Tt C St C x0t ı C It ; It  N.0; I2 /;

for t D 1; : : : ; n, and where xt is a K-dimensional vector of predetermined covariates

and ı is a K  1 vector of regression coefficients, that can be allowed to change over
time.

atC1jt D  atjt ; PtC1jt D  t Ptjt  0t C Rt ˙ "t RTt :

Figure 1.5 shows the original values of the time series US Unemployment Rate
for Males (16 years and over) observed from January 1992 to December 2013
together with the final seasonally adjusted series obtained with the STAMP software.
Figure 1.6 exhibits the corresponding seasonal component estimated using the
default option (additive decomposition).
20 1 Introduction

12
Original series
STAMP seasonally adjusted series
10
8
6
4

1995 2000 2005 2010

Fig. 1.5 Original and STAMP (default option) seasonally adjusted US Unemployment Rate for
Males (16 years and over) observed from January 1992 to December 2013
0.8
0.4
0.0
-0.4

1995 2000 2005 2010

Fig. 1.6 Seasonal effects of the US Unemployment Rate for Males (16 years and over) estimated
using the STAMP software (default option)
1.3 Part II: Real Time Trend-Cycle Estimation 21

1.3 Part II: Real Time Trend-Cycle Estimation

The second part of this books deals with recent developments on trend-cycle
estimation on real time. It comprises Chaps. 7–11. The basic approach to the
analysis of current economic conditions, known as recession and recovery analysis,
is that of assessing the real time trend-cycle of major socioeconomic indicators
(leading, coincident, and lagging) using percentage changes, based on seasonally
adjusted units, calculated for months and quarters in chronological sequence. The
main goal is to evaluate the behavior of the economic indicators during incomplete
phases by comparing current contractions or expansions with corresponding phases
in the past. This is done by measuring changes of single time series (mostly
seasonally adjusted) from their standing at cyclical turning points with past changes
over a series of increasing spans. This differs from business cycle studies where
cyclical fluctuations are measured around a long-term trend to estimate complete
business cycles. The real time trend corresponds to an incomplete business cycle
and is strongly related to what is currently happening on the business cycle
stage.
In recent years, statistical agencies have shown an interest in providing trend-
cycle or smoothed seasonally adjusted graphs to facilitate recession and recovery
analysis. Among other reasons, this interest originated from the recent crisis and
major economic and financial changes of global nature which have introduced more
variability in the data. The USA entered in recession in December 2007 until June
2009, and this has produced a chain reaction all over the world, with great impact in
Europe and China. There are no evidence of a fast recovery as in previous recession:
the economic growth is sluggish and with high levels of unemployment. It has
become difficult to determine the direction of the short-term trend (or trend-cycle)
as traditionally done by looking at month to month (quarter to quarter) changes
of seasonally adjusted values, particularly to assess the upcoming of a true turning
point. Failure in providing reliable real time trend-cycle estimates could give rise
to dangerous drift of the adopted policies. Therefore, a consistent prediction is of
fundamental importance. It can be done by means of either univariate parametric
models or nonparametric techniques. Since the majority of the statistical agencies
use seasonally adjusted software, such as the Census II-X11 method and its variants
X11/12ARIMA and X13, this book deals with the nonparametric techniques.
The linear filter developed by Henderson [34] is the most frequently applied
and has the property that fitted to exact cubic functions will reproduce their
values, and fitted to stochastic cubic polynomials it will give smoother results than
those estimated by ordinary least squares. The properties and limitations of the
Henderson filters have been extensively discussed by many authors, among them,
[10, 14, 15, 20, 21, 25, 29, 36, 41, 43], and Dagum and Bianconcini [16] who were
the first to represent the Henderson filter using Reproducing Kernel Hilbert Space
(RKHS) methodology.
22 1 Introduction

At the beginning and end of the sample period, the Henderson filter of length, say,
2m C 1 cannot be applied to the m data points, hence only asymmetric filters can
be used. The estimates of the real time trend are then subject to revisions produced
by the innovations brought by the new data entering in the estimation and the time-
varying nature of the asymmetric filters, in the sense of being different for each of
the m data points.
The asymmetric filters applied to the first and last m observations associated with
the Henderson filter were developed by Musgrave [45] on the basis of minimizing
the mean squared revision between the final estimates, obtained with the symmetric
Henderson weights, and preliminary estimates from the asymmetric weights, subject
to the constraint that the sum of these weights is equal to one. The assumption made
is that at the end of the series, the seasonally adjusted values do not follow a cubic
polynomial, but a linear trend-cycle plus a purely random irregular. Several authors
have studied the statistical properties and limitations of the Musgrave filters, among
others, [4, 20, 21, 23, 30, 42, 49], and Dagum and Bianconcini [16, 17] were the
first to introduce an RKHS representation of the asymmetric filters developed by
Musgrave [45].
Dagum [15] developed a nonlinear smoother to improve on the classical Hen-
derson filter. The NonLinear Dagum Filter (NLDF) results from applying the
13-term symmetric Henderson filter (H13) to seasonally adjusted series where
outliers and extreme observations have been replaced and which have been extended
with extrapolations from an ARIMA model. The main purpose of the ARIMA
extrapolations is to reduce the size of the revisions of the most recent estimates
whereas that of extreme values replacement is to reduce the number of unwanted
ripples produced by H13. An unwanted ripple is a 10-month cycle (identified by the
presence of high power at ! D 0:10 in the frequency domain) which, due to its
periodicity, often leads to the wrong identification of a true turning point. In fact, it
falls in the neighborhood between the fundamental seasonal frequency and its first
harmonic. On the other hand, a high frequency cycle is generally assumed to be
part of the noise pertaining to the frequency band 0:10  ! < 0:50. The problem
of the unwanted ripples is specific of H13 when applied to seasonally adjusted
series. Studies by Dagum, Chhab, and Morry [15] and Chhab et al. [9] showed the
superior performance of the NLDF respect to both structural and ARIMA standard
parametric trend-cycle models applied to series with different degrees of signal-
to-noise ratios. In another study, the good performance of the NLDF is shown
relative to nonparametric smoothers, namely: locally weighted regression (LOESS),
Gaussian kernel, cubic smoothing spline, and supersmoother [19].
Dagum and Luati developed in 2009 what they called a Cascade Linear Filter
(CLF), and distinguished between the Symmetric (SLF) and the Asymmetric Linear
Filter (ALF).
A linear filter offers many advantages over a nonlinear one. For one, its
application is direct and hence does not require knowledge of ARIMA model
identification. Furthermore, linear filtering preserves the crucial additive constraint
by which the trend of an aggregated variable should be equal to the algebraic
addition of its component trends, thus avoiding the problem of direct versus indirect
1.3 Part II: Real Time Trend-Cycle Estimation 23

adjustment. Finally, the properties of a linear filter concerning signal passing and
noise suppression can always be compared to those of other linear filters by means
of spectral analysis.
These authors studied the properties of the new CLF relative to 13-term
Henderson filter (H13) and have shown that the CLF is an optimal trend-cycle
estimator among a variety of second and higher order kernels restricted to be of
the same length. The theoretical properties were presented by means of spectral
analysis, whereas the empirical properties were evaluated on a large sample of real
time series pertaining to various socioeconomic areas and with different degrees
of variability. It should be noticed that the theoretical properties of CLF cannot be
compared with those of NLDF since the latter is data dependent.
From another perspective Dagum and Bianconcini in 2008 and 2013 were the
first to introduce an RKHS representation of the symmetric Henderson and the
asymmetric filters developed by Musgrave.
An RKHS is a Hilbert space characterized by a kernel that reproduces, via
an inner product, every function of the space or, equivalently, a Hilbert space of
real valued functions with the property that every point evaluation functional is
bounded and linear. Parzen [47] was the first to introduce an RKHS approach in time
series analysis applying the famous Loéve theorem by which there is an isometric
isomorphism between the closed linear span of a second order stationary stochastic
process and the RKHS determined by its covariance function. Parzen demonstrated
that the RKHS approach provides a unified framework to three fundamental
problems related with: (1) least squares estimation; (2) minimum variance unbiased
estimation of regression parameters; and (3) identification of unknown signals
perturbed by noise. Parzen’s approach is parametric, and basically consists of
estimating the unknown signal by generalized least squares in terms of the inner
product between the observations and the covariance function. A nonparametric
approach of the RKHS methodology was developed by De Boor and Lynch [22]
in the context of cubic spline approximation. Later, Kimeldorf and Wahba[38]
exploited both developments and treated the general spline smoothing problem from
a RKHS stochastic equivalence perspective. These authors proved that minimum
norm interpolation and smoothing problems with quadratic constraints imply an
equivalent Gaussian stochastic process.
In this book we show how nonparametric estimators can be transformed into
kernel functions of order two, that are probability densities, and from which
corresponding hierarchies of estimators are derived. The density function provides
the “initial weighting shape" from which the higher order kernels inherit their
properties. This kernel representation enables the comparison of estimators based
on different smoothing criteria, and has important consequences in the derivation
of the asymmetric filters which can be applied to the most recent observations. In
particular, those obtained by means of RKHS are shown to have superior properties
from the view point of signal passing, noise suppression, and revisions relative to
the classical ones.
24 1 Introduction

The RKHS approach presented in this book is strictly nonparametric. It makes

use of the fundamental theoretical result due to Berlinet [3] according to which
a kernel estimator of order p can always be decomposed into the product of a
reproducing kernel Rp1 , belonging to the space of polynomials of degree at most
p  1, and a probability density function f0 with finite moments up to order 2p.
Hence, the weighted least squares estimation of the nonstationary mean uses weights
derived from the density function f0 from which the reproducing kernel is defined
and not from the covariance function.
In the RKHS framework, given the density function, once the length of the
symmetric filter is chosen, let us say, 2m C 1, the statistical properties of the
asymmetric filters are strongly affected by the bandwidth parameter of the kernel
function from which the weights are derived.
Applied to real data, the kernel acts as a locally weighted average or linear filter
that for each target point t gives the estimate

X
n
yO t D wt;i yi ; t D 1; 2; : : : ; n (1.18)
iD1

where wt;i denotes the weights to be applied to the observations yi to get the estimate
yO t for each point in time t; t D 1; : : : ; n.
Once a symmetric span 2mC1 of the neighborhood has been selected, the weights
for the observations corresponding to points falling out of the neighborhood of any
target point are null or approximately null. Hence, the estimates of the n2m central
observations are obtained by applying 2mC1 symmetric weights to the observations
neighboring the target point. That is,

X
m
yO t D wj ytCj t D m C 1; : : : ; n  m:
jDm

The weights wj ; j D m; : : : ; m depend on the shape of the nonparametric

estimator KpC1 and on the value of a bandwidth parameter b fixed to ensure a
neighborhood amplitude equal to 2m C 1, such that
 
j
KpC1 b
wj D P  : (1.19)
m j
jDm KpC1 b

To derive the Henderson kernel hierarchy by means of the RKHS methodology,

the density corresponding to the Henderson ideal weighting function, Wj , and its
orthonormal polynomials have to be determined. The triweight density function
proposed by Loader [43] gives very poor results when the Henderson smoother
spans are of short or medium lengths, as in most application cases, ranging from
5 to 23 terms. Hence, Dagum and Bianconcini [16] derived the exact density
function corresponding to t Wj . The exact probability density corresponding to the
1.3 Part II: Real Time Trend-Cycle Estimation 25

Henderson’s ideal weighting penalty function is given by

.m C 1/
f0H .t/ D W..m C 1/t/; t 2 Œ1; 1
k
R .mC1
where k D .mC1/ W. j/dj and j D .m C 1/t.
These authors also found that the biweight density function gives almost
equivalent results to those obtained with the exact density function without the need
to be calculated any time that the Henderson smoother length changes. Another
important advantage is that the biweight density function belongs to the well-known
Beta distribution family, that is,
 
r
f .t/ D .1  jtjr /s IŒ1;1 .t/;
2B.s C 1; 1r/
R1
where B.a; b/ D 0 ta1 .1  t/b1 dt with a; b > 0 is the Beta function.
The orthonormal polynomials associated to the biweight function are the Jacobi
polynomials, for which explicit expressions for computation are available and their
properties have been widely studied in literature.
For the m first and last filters we discuss time-varying bandwidth parameters
since the asymmetric filters are time-varying. Three specific criteria of bandwidth
selection are chosen based on the minimization of
1. the distance between the transfer functions of asymmetric and symmetric filters,
2. the distance between the gain functions of asymmetric and symmetric filters, and
3. the phase shift function over the domain of the signal.
We deal only with the reduction of revisions due to filter changes that depend on
how close the asymmetric filters are respect to the symmetric one [14, 15] and do not
consider revisions introduced by the innovations in the new data. Another important
aspect discussed deals with the capability of the asymmetric filters to signal the
upcoming of a true turning point that depends on the time delay for its identification.
This is obtained by calculating the number of months (quarters) it takes for the
last trend-cycle estimate to signal a true turning point in the same position of the
final trend-cycle data. An optimal asymmetric filter should have a time path that
converges fast and monotonically to the final estimate as new observations are added
to the series.
It is shown in Chap. 10 that, applied to a set of US leading, coincident, and
lagging indicators, the new set of asymmetric kernel filters reduced by one half the
size of the revisions and by one third the time delay to detect the June 2009 turning
point relative to the Musgrave filters.
26 1 Introduction

References

1. Anderson, B. D. O., & Moore, J. B. (1979). Optimal filtering. Englewood Cliffs: Prentice-Hall.
2. Bell, W. H., & Hillmer, S. C. (1984). Issues involved with the seasonal adjustment of economic
time series. Journal of Business and Economic Statistics, 2, 291–320.
3. Berlinet, A. (1993). Hierarchies of higher order kernels. Probability Theory and Related Fields,
94, 489–504.
4. Bianconcini, S., & Quenneville, B. (2010). Real time analysis based on reproducing kernel
Henderson filters. Estudios de Economia Aplicada, 28(3), 553–574.
5. Bobbit, L., & Otto, M. C. (1990). Effects of forecasts on the revisions of seasonally
adjusted values using the X11 seasonal adjustment procedure. In Proceedings of the American
Statistical Association Business and Economic Statistics Session (pp. 449–453).
6. Box, G. E. P., & Tiao, G. C. (1975). Intervention analysis with applications to economics and
environmental problems. Journal of the American Statistical Association, 70, 70–79.
7. Burman, J. P. (1980). Seasonal adjustment by signal extraction. Journal of the Royal Statistical
Society Series A, 143, 321–337.
8. Chen, C., & Liu, L. M. (1993). Joint estimation of model parameters and outlier effects in time
series. Journal of the American Statistical Association, 88, 284–297.
9. Chhab, N., Morry, M., & Dagum, E. B. (1999). Results on alternative trend-cycle estimators
for current economic analysis. Estadistica, 49–51, 231–257.
10. Cholette, P. A. (1981). A comparison of various trend-cycle estimators. In O. D. Anderson &
M. R. Perryman (Eds.), Time series analysis (pp. 77–87). Amsterdam: North Holland.
11. Commandeur, J. J. F., & Koopman, S. J. (2007). An introduction to state space time series
analysis. Oxford: Oxford University Press.
12. Dagum, E. B. (1975). Seasonal factor forecasts from ARIMA models. In 40th Session
International Statistical Institute, Warsaw (Vol. 3, pp. 203–216).
13. Dagum, E. B. (1978). Modelling, forecasting and seasonally adjusting economic time series
with the X-11 ARIMA method. The Statistician, 27(3), 203–215.
14. Dagum, E. B. (1988). The X-11 ARIMA/88 seasonal adjustment method-foundations and user’s
manual. Ottawa, ON, Canada: Time Series Research and Analysis Centre, Statistics Canada.
15. Dagum, E. B. (1996). A new method to reduce unwanted ripples and revisions in trend-cycle
estimates from X11ARIMA. Survey Methodology, 22, 77–83.
16. Dagum, E. B., & Bianconcini, S. (2008). The Henderson smoother in reproducing kernel
Hilbert space. Journal of Business and Economic Statistics, 26(4), 536–545.
17. Dagum, E. B., & Bianconcini, S. (2013). A unified probabilistic view of nonparametric
predictors via reproducing kernel Hilbert spaces. Econometric Reviews, 32(7), 848–867.
18. Dagum, E. B., Chhab, N., & Chiu, K. (1996). Derivation and analysis of the X11ARIMA and
census X11 linear filters. Journal of Official Statistics, 12(4), 329–347.
19. Dagum, E. B., & Luati, A. (2000). Predictive performance of some nonparametric linear and
nonlinear smoothers for noisy data. Statistica, LX(4), 635–654.
20. Dagum, E. B., & Luati, A. (2009). A cascade linear filter to reduce revisions and turning points
for real time trend-cycle estimation. Econometric Reviews, 28(1–3), 40–59.
21. Dagum, E. B., & Luati, A. (2012). Asymmetric filters for trend-cycle estimation. In W. R.
Bell, S. H. Holan, & T. S. McElroy (Eds.), Economic time series: Modeling and seasonality
(pp. 213–230). Boca Raton: Chapman & Hall.
22. De Boor, C., & Lynch, R. (1966). On splines and their minimum properties. Journal of
Mathematics and Mechanics, 15, 953–969.
23. Doherty, M. (2001). The surrogate Henderson filters in X-11. Australian and New Zealand
Journal of Statistics, 43(4), 385–392.
24. Durbin, J., & Koopman, S. J. (2001). Time series analysis by state space methods. Oxford:
Oxford University Press.
References 27

25. Findley, D. F., & Martin, D. E. K. (2006). Frequency domain analysis of SEATS and X-11/X-
12-ARIMA seasonal adjustment filters for short and moderate length time series. Journal of
Official Statistics, 22, 1–34.
26. Findley, D. F., Monsell, B. C., Bell, W. R., Otto, M. C., & Chen, B. C. (1998). New capabilities
and methods of the X12ARIMA seasonal adjustment program. Journal of Business and
Economic Statistics, 16(2), 127–152.
27. Gomez, V., & Maravall, A. (1994). Estimation, prediction and interpolation for nonstationary
series with the Kalman filter. Journal of the American Statistical Association, 89, 611–624.
28. Gomez, V., & Maravall, A. (1996). Program TRAMO and SEATS: Instructions for Users.
Working paper 9628. Service de Estudios, Banco de Espana.
29. Gray, A., & Thomson, P. (1996). Design of moving-average trend filters using fidelity and
smoothness criteria. In P. M. Robinson & M. Rosenblatt (Eds.), Time series analysis (in
memory of E.J. Hannan) Vol. II (Vol. 115, pp. 205–219). Springer Lecture notes in statistics.
New York: Springer.
30. Gray, A. G., & Thomson, P. J. (2002). On a family of finite moving-average trend filters for the
ends of series. Journal of Forecasting, 21, 125–149.
31. Hannan, E. J., & Rissanen, J. (1982). Recursive estimation of mixed autoregressive moving
average order. Biometrika, 69, 81–94.
32. Harvey, A. C. (1989). Forecasting, structural time series models and the Kalman filter.
Cambridge: Cambridge University Press.
33. Harvey, A. C., Koopman, S. J., & Penzer, J. (1998). Messy time series: A unified approach. In
T. B. Fomby & R. Carter Hill (Eds.), Advances in econometrics (Vol. 13, pp. 103–143). New
York, NY, USA: JAI Press.
34. Henderson, R. (1916). Note on graduation by adjusted average. Transaction of Actuarial
Society of America, 17, 43–48.
35. Hillmer, S. C., & Tiao, G. C. (1982). An ARIMA model-based approach to seasonal
adjustment. Journal of the American Statistical Association, 77, 63–70.
36. Kenny, P., & Durbin, J. (1982). Local trend estimation and seasonal adjustment of economic
and social time series. Journal of the Royal Statistical Society A, 145, 1–41.
37. Kim, C. J., & Nelson, C. R. (1999). State space models with regime switching. Cambridge,
MA: MIT Press.
38. Kimeldorf, G. S., & Wahba, G. (1970). A correspondence between Bayesian estimation on
stochastic processes and smoothing by splines. Annals of Mathematical Statistics, 41, 495–
502.
39. Kitagawa, G., & Gersch, W. (1996). Smoothness priors analysis of time series. New York:
Springer.
40. Koopman, S. J., & Ooms, M. (2011). Forecasting economic time series using unobserved
components time series models. In M. Clements & D. F. Hendry (Eds.), Oxford handbook
of economic forecasting. Oxford: Oxford University Press.
41. Ladiray, D., & Quenneville, B. (2001). Seasonal adjustment with the X11 method. Lecture
notes in statistics (Vol. 158). New York: Springer.
42. Laniel, N. (1985). Design Criteria for the 13-term Henderson End Weights. Working paper.
Methodology Branch, Statistics Canada, Ottawa.
43. Loader, C. (1999). Local regression and likelihood. New York: Springer.
44. Macaulay, F. R. (1931). The smoothing of time series. New York: National Bureau of Economic
Research.
45. Musgrave, J. (1964). A Set of End Weights to End All End Weights. Working paper. U.S.
Bureau of Census, Washington, DC.
46. Newbold, P., & Granger, C. W. J. (1974). Experience with forecasting univariate time series
and the combination of forecasts (with discussion). Journal of the Royal Statistical Society.
Series A, 137, 131–165.
47. Parzen, E. (1959). Statistical Inference on Time Series by Hilbert Space Methods. Technical
Report No. 53, Statistics Department, Stanford University, Stanford, CA.
48. Persons, W. M. (1919). Indices of business conditions. Review of Economic Statistics, 1, 5–107.
28 1 Introduction

49. Quenneville, B., Ladiray, D., & Lefrancois, B. (2003). A note on Musgrave asymmetrical trend-
cycle filters. International Journal of Forecasting, 19(4), 727–734.
50. Reid, D. J. (1975). A review of short-term projection techniques. In H. A. Gorden (Ed.),
Practical aspects of forecasting (pp. 8–25). London: Operational Research Society.
51. Shiskin, J., Young, A. H., & Musgrave, J. C. (1967). The X-11 Variant of the Census Method
II Seasonal Adjustment Program. Technical Paper 15 (revised). US Department of Commerce,
Bureau of the Census, Washington, DC.
52. Tiao, G. C., & Tsay, R. S. (1983). Consistency properties of least squares estimates of
autoregressive parameters in Arma models. The Annals of Statistics, 11, 856–871.
53. Tsay, R. S. (1984). Regression models with time series errors. Journal of the American
Statistical Association, 79, 118–124.
54. Young, A. H. (1968). A linear approximation to the Census and BLS seasonal adjustment
methods. Journal of the American Statistical Association, 63, 445–457.
Chapter 2
Time Series Components

Abstract An important objective in time series analysis is the decomposition of a

series into a set of unobservable (latent) components that can be associated with
different types of temporal variations. This chapter introduces the definitions and
assumptions made on these unobservable components that are: (1) a long-term
tendency or secular trend, (2) cyclical movements superimposed upon the long-
term trend. These cycles appear to reach their peaks during periods of economic
prosperity and their troughs during periods of depressions, their rise and fall
constituting the business cycle, (3) seasonal variations that represent the composite
effect of climatic and institutional events which repeat more or less regularly
each year, and (4) the irregular component. When the series result from the daily
accumulation of activities, they can also be affected by other variations associated
with the composition of the calendar. The two most important are trading day
variations, due to the fact that the activity in some days of the week is more
important than others, and moving holidays the date of which change in consecutive
months from year to year, e.g., Easter.

A time series consists of a set of observations ordered in time on a given

phenomenon (target variable). Usually the measurements are equally spaced, e.g.,
by year, quarter, month, week, and day. The most important property of a time series
is that the ordered observations are dependent through time, and the nature of this
dependence is of interest in itself.
Formally, a time series is defined as a sample realization of a set of random
variables indexed in time, that is, a stochastic process denoted by fY1 ; : : : ; Yn g or
simply fYt g. In this regard, an observed time series is denoted by fy1 ; : : : ; yn g,
where the sub-index indicates the time to which the observation pertains, or in
compact form fyt g. The first observed value y1 can be interpreted as the realization
of the random variable Y1 , which can also be written as Y.!; t D 1/ where !
denotes the event belonging to a probabilistic space. Similarly, y2 is the realization
of Y2 , and so on. The random variables Y1 ; Y2 ; : : : ; Yn can be characterized by
different probability distribution. For socioeconomic time series the probability
space is continuous, and the time measurements are discrete. The frequency of
measurements is said to be high when it is daily, weekly, or monthly and to be
low when the observations are quarterly or yearly.

© Springer International Publishing Switzerland 2016 29

E. Bee Dagum, S. Bianconcini, Seasonal Adjustment Methods and Real Time
Trend-Cycle Estimation, Statistics for Social and Behavioral Sciences,
DOI 10.1007/978-3-319-31822-6_2
30 2 Time Series Components

2.1 Time Series Decomposition Models

An important objective in time series analysis is the decomposition of a series

into a set of non-observable (latent) components that can be associated with
different types of temporal variations. The idea of time series decomposition is
very old and was used for the calculation of planetary orbits by seventeenth
century astronomers. Persons [37] was the first to state explicitly the assumptions of
unobserved components. As Persons saw it, time series was composed of four types
of fluctuations:
1. a long-term tendency or secular trend;
2. cyclical movements superimposed upon the long-term trend. These cycles appear
to reach their peaks during periods of industrial prosperity and their troughs
during periods of depressions, their rise and fall constituting the business cycle;
3. a seasonal movement within each year, the shape of which depends on the nature
of the series;
4. residual variations due to changes impacting individual variables or other major
events, such as wars and national catastrophes affecting a number of variables.
Traditionally, the four variations have been assumed to be mutually independent
from one another and specified by means of an additive decomposition model:

yt D Tt C Ct C St C It ; t D 1; : : : n (2.1)

where yt denotes the observed series at time t, Tt the long-term trend, Ct the business
cycle, St seasonality, and It the irregulars.
If there is dependence among the latent components, this relationship is specified
through a multiplicative model

yt D Tt  Ct  St  It ; t D 1; : : : n (2.2)

where now St and It are expressed in proportion to the trend-cycle Tt  Ct . In some

cases, mixed additive–multiplicative models are used.
Whether a latent component is present or not in a given time series depends on
the nature of the phenomenon and on the frequency of measurement. For example,
seasonality is due to the fact that some months or quarters of a year are more
important in terms of activity or level. Because this component is specified to cancel
out over 12 consecutive months or 4 consecutive quarters, or more generally over
365.25 consecutive days, yearly series cannot contain seasonality.
Flow series can be affected by other variations associated with the composition of
the calendar. The most important are the trading day variations, which are due to the
fact that the activity in some days of the week is more important than others. Months
with five of the most important days register an excess of activity (ceteris paribus)
in comparison to months with four such days. Conversely, months with five of the
least important days register a short-fall of activity. The length-of-month variation is
2.1 Time Series Decomposition Models 31

usually assigned to the seasonal component. The trading day component is usually
considered as negligible in quarterly series and even more so in yearly data.
Another important calendar variation is the moving holiday or moving-festival
component. That component is associated with holidays which change date from
year to year, e.g., Easter, causing a displacement of activity from 1 month to the
previous or the following month. For example, an early date of Easter in March or
early April can cause an important excess of activity in March and a corresponding
short-fall in April, in variables associated with imports, exports, and tourism.
Under models (2.1) and (2.2), the trading day and moving festival components (if
present) are implicitly part of the irregular. In 1965, Young developed a procedure
to estimate trading day variations which was incorporated in the X11 seasonal
adjustment method [39] and its subsequent versions, the X11ARIMA [17, 18] and
X12ARIMA [21] methods. The latter two versions also include models to estimate
moving holidays, such as Easter.
If the new components are present, the additive decomposition model becomes

yt D Tt C Ct C St C Dt C Ht C It (2.3)

where Dt and Ht denote the trading day and moving holiday components, respec-
tively. Similarly, the multiplicative decomposition model becomes

yt D Tt  Ct  St  Dt  Ht  It (2.4)

where the components St ; Dt ; Ht , and It are proportional to the trend-cycle.

Decomposition models (2.3) and (2.4) are traditionally used by seasonal adjust-
ment methods. Other less used decomposition models are the log-additive and
the mixed models (additive and multiplicative) where, for example, the systematic
relationship among the components is multiplicative, but the irregulars are additive.
Seasonal adjustment actually entails the estimation of all the time series com-
ponents and the removal of seasonality, trading day, and holiday effects from
the observed series. The rationale is that these components which are relatively
predictable conceal the current stage of the business cycle which is critical for policy
and decision making.
There is another kind of time series decomposition often used for modeling and
forecasting univariate ARIMA time series:

yt D t C et (2.5)

where t and et are referred to as the signal and the noise, according to the electrical
engineering terminology. The signal t comprises all the systematic components of
models (2.1) and (2.4), i.e., Tt ; Ct ; St ; Dt , and Ht .
Model (2.5) is classical in signal extraction where the problem is to find the best
estimates of the signal t given the observations yt corrupted by noise et . The best
estimates are usually defined as minimizing the mean square error.
32 2 Time Series Components

Finally, given its fundamental role in time series modeling, we summarize the
well-known decomposition theorem due to Herman Wold [41].
A stochastic process fYt g is second order stationary or weakly stationary if the
first two moments are not time dependent, that is, the mean and the variance are
constant, and the autocovariance function depends only on the time lag and not on
the time origin:

E.Yt / D  (2.6)

E.Yt  /2 D Y2 < 1; (2.7)

EŒ.Yt  /.YtCk  / D k (2.8)

where k D 0; 1; 2; : : : , denotes the time lag.

Wold proved that any stochastic process fYt g, stationary to the second order, can
be decomposed into two mutually uncorrelated processes fZt g and fVt g, such that

Yt D Zt C Vt (2.9)

where
1
X 1
X
2
Zt D j atj ; 0 D 1; j <1 (2.10)
jD0 jD1

with fat g  WN.0; a2 /.

The component fZt g is a convergent infinite linear combination of the at ’s,
assumed to follow a white noise (WN) process with zero mean, constant variance
a2 , and zero autocovariance. Model (2.10) is known as an infinite moving average,
MA.1/, and the at ’s are the innovations. The component fZt g is called the non-
deterministic or purely linear component since only one realization of the process is
not sufficient to determine future values ZtC` ; ` > 0, without error.
The component fVt g can be represented by
1
X  
Vt D  C ˛j sin.j t/ C ˇj cos.j t/ ;  <  < ; (2.11)
jD1

where  is the constant mean of process fYt g and f˛j g; fˇj g are mutually uncorre-
lated white noise processes. The process fVt g is called deterministic because it can
be predicted in the future without error from a single realization by means of an
infinite linear combination of past values.
Wold theorem demonstrates that the property of stationarity is strongly related
to that of linearity. It provides a justification for autoregressive moving average
(ARMA) models [5] and some extensions, such as the autoregressive integrated
moving average (ARIMA) and regression-ARIMA models (regARIMA).
2.2 The Secular or Long-Term Trend 33

2.2 The Secular or Long-Term Trend

The concept of trend is used in economics and other sciences to represent long-
term smooth variations. The causes of these variations are often associated with
structural phenomena such as population growth, technological progress, capital
accumulation, and new practices of business and economic organization. For most
economic time series, the trends evolve smoothly and gradually, whether in a
deterministic or stochastic manner. When there is a sudden change of level and/or
slope this is referred to as a structural change. However, it should be noticed that
series at a higher levels of aggregation are less susceptible to structural changes. For
example, a technological change is more likely to produce a structural change for
some firms than for the whole industry.
The identification and estimation of the secular or long-term trend have posed
serious challenges to statisticians. The problem is not of statistical or mathematical
character but originates from the fact that the trend is a latent (non-observable)
component and its definition as a long-term smooth movement is statistically vague.
The concept of a long-period is relative, since a trend estimated for a given series
may turn out to be just a long business cycle as more years of data become available.
To avoid this problem statisticians have used two simple solutions. One is to estimate
the trend and the business cycles together, calling it the trend-cycle. The other
solution is to estimate the trend over the whole series, and to refer to it as the longest
non-periodic variation.
It should be kept in mind that many systems of time series are redefined every
15 years or so in order to maintain relevance. Hence, the concept of long-term trend
loses importance. For example, the Canadian system of Retail and Wholesale Trade
series was redefined in 1989 to adopt the 1980 Standard Industrial Classification
(SIC), and in 2003 to conform to the North American Industrial Classification
System (NAICS), following the North American Free Trade Agreement. The
following examples illustrate the need for such reclassifications. The 1970 Standard
Industrial Classification (SIC) considered computers as business machines, e.g.,
cash registers and desk calculators. The 1980 SIC rectified the situation by creating
a class for computers and other goods and services. The last few decades witnessed
the birth of new industries involved in photonics (lasers), bio-engineering, nano-
technology, and electronic commerce. In the process, new professions emerged, and
classification systems had to keep up with these new realities.
There is a large number of deterministic and stochastic models which have
been proposed for trend estimation (see Dagum and Dagum [15]). Deterministic
models are based on the assumption that the trend can be well approximated by
mathematical functions of time, whereas stochastic trends models assume that the
trend can be better modeled by applying finite differences of low order together with
autoregressive components and moving average errors.
34 2 Time Series Components

2.2.1 Deterministic Trend Models

The most common representation of a deterministic trend is by means of polynomial

functions. The observed time series is assumed to have a deterministic nonstationary
mean, i.e., a mean dependent on time. A classical model is the regression error
model where the observed data is treated as the sum of the trend and a random
component, such that

yt D Tt C et ; (2.12)

where Tt denotes the trend and fet g is assumed to follow a stationary process, often
white noise. The polynomial trend can be written as

Tt D ˛0 C ˛1 t C    C ˛p tp (2.13)

where generally p  3. The trend is said to be of a deterministic character

because the observed series is affected by random shocks which are assumed to
be uncorrelated with the systematic part.
Besides polynomial of time, three very well-known growth functions have
been widely applied in population and economic studies, namely the modified
exponential, the Gompertz, and the logistic function. Historically, the first growth
model for time series was proposed by Malthus [34] in the context of population
growth. He stated two time path processes, one for the supply of food and the other
for population. According to Malthus, the supply of food followed an arithmetic
progression and the population a geometric one.
For illustrative purposes, Fig. 2.1 shows a deterministic cubic trend fitted to the
annual values of the Real Gross Domestic Product (billions of dollars) for the period
starting from 1929 to 2013.

2.2.2 Stochastic Trends

Stochastic models are appropriate when the trend is assumed to follow a nonstation-
ary stochastic process. The nonstationarity is modeled with finite differences of low
order (cf. [24] and [35]).
A typical stochastic trend model often used in structural time series modeling is
the so-called random walk with constant drift. In the classical notation, the model is

Tt D Tt1 C ˇ C aTt ; t D 1; 2; : : : ; nI aTt  N.0; T2 / (2.14)

Tt D ˇ C aTt
2.2 The Secular or Long-Term Trend 35

Cubic trend
15000
10000
5000 Annual Real Gross Domestic product

1940 1960 1980 2000

years

Fig. 2.1 Cubic trend fitted to the annual values of the Real Gross Domestic Product (billions of
dollars)

where Tt denotes the trend, ˇ a constant drift, and faTt g is a normal white noise
process. Solving the difference equation (2.14) and assuming aT0 D 0, we obtain
1
X
Tt D ˇt C 1 aTt D ˇt C aTtj ; t D 1; : : : ; n; (2.15)
jD0

which show that a random walk with constant drift consists of a linear deterministic
trend plus a nonstationary infinite moving average.
Another type of stochastic trend belongs to the ARIMA (p; d; q) class, where p
is the order of the autoregressive polynomial, q is the order of the moving average
polynomial, and d the order of the finite difference operator .1  B/. The backshift
operator B is such that Bn Tt D Ttn . The ARIMA(p; d; q) model is written as

p .B/.1  B/d Tt D q .B/aTt ; aTt  N.0; T2 /; (2.16)

where Tt denotes the trend, p .B/ the autoregressive polynomial in B of order p,

q stands for the moving average polynomial in B of order q, and fat g denotes the
innovations assumed to follow a normal white noise process. For example, with
p D 1; d D 2, and q D 0, model (2.16) becomes

.1  1 B/.1  B/2 Tt D aTt (2.17)

36 2 Time Series Components

which means that after applying first order differences twice, the transformed series
can be modeled by an autoregressive process of order one.

2.3 The Business Cycle

The business cycle is a quasi-periodic oscillation characterized by periods of

expansion and contraction of the economy, lasting on average from 3 to 5 years.
Because most time series are too short for the identification of a trend, the cycle
and the trend are estimated jointly and referred to as the trend-cycle. As a result
the concept of trend loses importance. The trend-cycle is considered a fundamental
component, reflecting the underlying socioeconomic conditions, as opposed to
seasonal, trading day, and irregular fluctuations.
The proper identification of cycles in the economy requires a definition of
contraction and expansion. The definition used in capitalistic countries to produce
the chronology of cycles is based on fluctuations found in the aggregate economic
activity. A cycle consists of an expansion phase simultaneously present in many
economic activities, followed by a recession phase and by a recovery which
develops into the next expansion phase. This sequence is recurrent but not strictly
periodic. Business cycles vary in intensity and duration. For example, in the USA,
the 1981 recession was very acute but of short duration, whereas the 2008 recession
was severe and of long duration. The latter crisis was precipitated by the sub-prime
loan problem in the financial sector and this has caused an increased interest in the
linkage between financial and real economic activities.
The analysis of business and economic cycles has preoccupied economists and
statisticians for a long time. In 1862, the French economist Clement Juglar identified
the presence of economic cycles 8–11 years long, although he was cautious not
to claim any strict rigid regularity. The period going from 1950/51 till 1973 was
characterized by high worldwide growth, and at least the problem of depression
was declared dead; first in the late 1960s, when the Phillip curve was seen as
being able to steer the economy which was followed by stagflation in the 1970s,
which discredited the theory. Secondly in the early 2000s, following the stability
and growth in the 1980 and 1990 in what came to be known as The Great
Moderation. This phrase was sometimes used to describe the perceived end to
economic volatility created by twentieth century banking systems. The term was
coined by Harvard economists: James Stock and Mark Watson in their article written
in 2002, “Has the Business Cycle Changed and Why.” The validity of this concept
as a permanent shift has been questioned by the economic and financial crisis that
started at the end of 2007. In the mid-1980s major economic variables such as
GDP, industrial production, monthly payroll employment, and the unemployment
rate began a decline in volatility (see [4]). Stock and Watson [40] viewed the causes
of the moderation to be “improved policy, identifiable good luck in the form of
productivity and commodity price shocks, and other unknown forms of good luck.”
The greater predictability in economic and financial performance had caused firms
2.3 The Business Cycle 37

to hold less capital and to be less concerned about liquidity positions. This, in turn,
is thought to have been a factor in encouraging increased debt levels and a reduction
in risk premium required by investors.
The period of the Great Moderation ranges between 1987 and 2007, and it
is characterized by predictable policy, low inflation, and modest business cycles.
However, at the same time various regions in the world have experienced prolonged
depressions, most dramatically the economic crisis in former Eastern Bloc countries
following the end of the Soviet Union in 1991. For several of these countries the
period 1989–2010 has been an ongoing depression, with real income still lower
than in 1989. In economics a depression is a more severe downturn than a recession,
which is seen by economists as part of a normal business cycle.
Considered a rare and extreme form of recession, a depression is characterized
by its length, and by abnormally large increases in unemployment, falls in the
availability of credit, shrinking output and investment, numerous bankruptcies,
significantly reduced amounts of trade and commerce, especially international,
as well as highly volatile relative currency value fluctuations most often due to
devaluations.
In 1946, economists Arthur F. Burns and Wesley C. Mitchell [9] provided the new
standard definition of business cycles in their book “Measuring Business Cycles”:
“Business cycles are a type of fluctuation found in the aggregate economic activity
of nations that organize their work mainly in business enterprises: a cycle consists of
expansions occurring at about the same time in many economic activities, followed
by similarly general recessions, contractions, and revival which merge into the
expansion phase of the next cycle; in duration business cycles vary from more than
one year to ten to twelve years; they are not divisible into shorter cycles of similar
characteristics with amplitudes approximating their own.”
In 1954 Schumpeter [38] stated that an economic cycle has four stages : (1)
expansion (increases in production and prices, low interests rates); (2) crisis (stock
exchanges crash and multiple bankruptcies of firms occur); (3) recession (drops
in prices and in output, high interests rates); and (4) recovery (stocks recover
because of the fall in prices and incomes). In this model, recovery and prosperity are
associated with increases in productivity, consumer confidence, aggregate demand,
and prices.
In a broad sense, there have been two ways by which economic and business
cycles have been studied, one analyzing complete cycles and the other, studying the
behavior of the economic indicators during incomplete phases by comparing current
contractions or expansions with corresponding phases in the past in order to assess
current economic conditions.
Because completed economic and business cycles are not directly observable, for
their identification and estimation, it is needed to remove from seasonally adjusted
series a long-term trend. This was originally done by means of a long moving
average using the Bry and Boschan method [7] adopted by the National Bureau of
Economic Research (NBER). Later, other methods were developed, such as those
of Hodrick and Prescott [28], Baxter and King [2], and Buttherworth [11] filters.
38 2 Time Series Components

The index of Leading Economic Indicators (LEI) is intended to predict future

economic activity. Typically, three consecutive monthly changes in the same
direction suggest a turning point in the economy. For example, consecutive negative
readings would indicate a possible recession. In the USA, LEI is a composite of
the following 11 leading indicators: average workweek (manufacturing), initial
unemployment claims, new orders for consumer goods, vendor performance,
plant and equipment orders, building permits, change in unfilled durable orders,
sensitive material prices, stock prices (S&P 500), real M2, and index of consumer
expectations.
On the other hand, the Index of Coincident Indicators includes nonagricultural
employment, index of industrial production, personal income, and manufacturing
and trade sales.
Until recently the statistical analysis of macroeconomic fluctuations was domi-
nated by linear time series methods. Over the past 15 years, however, economists
have increasingly applied tractable parametric nonlinear time series models to
business cycle data; most prominent in this set of models are the classes of
threshold autoregressive (TAR) models, Markov switching autoregressive (MSAR),
and smooth transition autoregressive (STAR) models.
The basic approach to the analysis of current economic conditions (known as
recession and recovery analysis, see [36]) is that of assessing the short-term trend
of major economic indicators (leading, coincident, and lagging) using percentage
changes, based on seasonally adjusted units and calculated for months and quarters
in chronological sequence. The main goal is to evaluate the behavior of the
economic indicators during incomplete phases by comparing current contractions
or expansions with corresponding phases in the past. This is done by measuring
changes of single time series (mostly seasonally adjusted) from their standing
at cyclical turning points with past changes over a series of increasing spans.
In recent years, statistical agencies have shown an interest in providing further
smoothed seasonally adjusted data (where most of the noise is suppressed) and
trend-cycle estimates, to facilitate recession and recovery analysis. Among other
reasons, this interest originated from major economic and financial changes of
global nature which have introduced more variability in the data and, consequently,
in the seasonally adjusted values, making very difficult to determine the direction of
the short-term trend which includes the impact of the trend jointly with that of the
business cycle, for an early detection of a turning point.
There are two approaches for current economic analysis modeling, the parametric
one, that makes use of filters based on models, such as ARIMA models (see, among
several others, Maravall [35], Kaiser and Maravall [29] and [30]) or state space
models (see, e.g., Harvey [24] and Harvey and Trimbur [25]).
The other approach is nonparametric and based on digital filtering techniques.
For example, the estimation of the trend-cycle with the Census Method II-X11, and
its variants X11ARIMA and X12ARIMA, is done by the application of linear filters
due to Henderson [26]. The Henderson filters are applied to seasonally adjusted
data where the irregulars have been replaced taking into account the extreme values.
2.4 The Seasonal Variations 39

Stochastic linear approaches, parametric and nonparametric, are discussed in detail

in Chap. 7.

2.3.1 Deterministic and Stochastic Models for the Business

Cycle

Similarly to the trend, the models for cyclical variations can be deterministic or
stochastic. Deterministic models often consist of sine and cosine functions of
different amplitude and periodicities. For example, denoting the cycle by ct , a
deterministic model is
2
X  
Ct D ˛j cos.j t/ C ˇj sin.j t/ ; (2.18)
jD1

where 1 D 2=60 and 2 D 2=40. Model (2.18) takes into consideration two
dominant cycles found in the European and American economies, those of 60 and
40 months, respectively.
Stochastic models of the ARIMA type, involving autoregressive models of order
2 with complex roots, have also been used to model the trend-cycle. For example,

Ct D 1 Ct1 C 2 Ct2 C aCt ; aCt  N.0; C2 / (2.19)

where Ct denotes the cycle, faCt g is assumed Gaussian white noise, and the
following conditions apply to the parameters: 1 C 2 < 1, 2  1 < 1, and
1 < 1 < 0 (see [5]).

2.4 The Seasonal Variations

Seasonality originates from climate seasons and conventional events of religious,

social, and civic nature, which repeat regularly from year to year.
The climatic seasons influence trade, agriculture, the consumption patterns of
energy, fishing, mining, and related activities. For example, the consumption of
heating oil increases in winter, and the consumption of electricity increases in the
summer months because of air conditioning.
Institutional seasons such as Christmas, Easter, civic holidays, the school and
academic year have a large impact on retail trade and on the consumption of certain
goods and services, namely travel by plane, hotel occupancy, and consumption of
gasoline.
The four main causes of seasonality are attributed to the weather, composition
of the calendar, major institutional deadlines, and expectations. Seasonality is
40 2 Time Series Components

largely exogenous to the economic system but can be partially offset by human
intervention. For example, seasonality in money supply can be controlled by central
bank decisions on interest rates. In other cases, the seasonal effects can be offset
by international and inter-regional trade. To some extent seasonality can evolve
through technological and institutional changes. For example, the developments
of appropriate construction materials and techniques made it possible to continue
building in winter. The development of new crops, which better resist cold and
dry weather, has influenced the seasonal pattern. The partial or total replacement
of some crops by chemical substitutes, e.g., substitute of sugar, vanilla, and other
flavors, reduces seasonality in the economy.
As for institutional change, the extension of the academic year to include some
summer months affected the seasonal pattern of unemployment for the population
of 15–25 years of age. Similarly, the practice of spreading holidays over the whole
year impacted on seasonality.
The changing industrial mix of an economy also transforms the seasonal pattern,
because some industries are more seasonal than others. In particular, economies
which diversify and depend less on “primary” industries (e.g., fishing and agricul-
ture) typically become less seasonal.
Another important characteristic of seasonality is the feasibility of its identifi-
cation even if it is a latent (not directly observable) variable. The identification
and estimation of seasonality, however, is not done independently on the other
components affecting the time series under study, that is, trend, business cycle,
trading day variations, moving holidays, and irregulars. The seasonal variations can
be distinguished from the trend by their oscillatory character, from the business
cycle by having annual periodicity, and from the irregulars by being systematic.
Seasonality entails large costs to society and businesses. One cost is the necessity
to build warehouses to store inventories of goods to be sold as consumers require
them, for example, grain elevators. Another cost is the under-use and over-use
of the factors of production: capital and labor. Capital in the form of unused
equipment, buildings, and land during part of the year has to be financed regardless.
For example, this is the case in farming, food processing, tourism, and electrical
generation. The cold climate increases the cost of buildings and infrastructure, e.g.,
roads, transportation systems, water and sewage systems, schools, and hospitals; not
to mention the damage to the same caused by the action of ice. The labor force is
over-used during the peak seasons of agriculture and construction for example; and,
under-used in trough seasons sometimes leading to social problems.
In order to determine whether a series contains seasonality, it is sufficient to
identify at least 1 month (or quarter) which tends to be systematically higher or
lower than other month.
For illustrative purposes, Fig. 2.2 exhibits the seasonal variations affecting the
US New Orders for Durable Goods (NODG) from February 1992 till December
2013 that are characterized by three main peaks and two main troughs. The peaks
are observed in March, June and, in recent years, also in December when orders are
highest, whereas the troughs correspond to the months of January and July.
2.4 The Seasonal Variations 41

110
100
95
90
85

1995 2000 2005 2010

Fig. 2.2 Seasonal pattern of US New Orders for Durable Goods (January 1992–December 2013)

The seasonal pattern measures the relative importance of the months of the year.
The constant 100 % represents an average month or a nonseasonal month. The peak
month in March 1993 is 108.6 or 8.6 % larger than on an average month; the trough
month is July with 83.2 in 1992, almost 17 % lower than on an average month.
The seasonal amplitude, the difference between the peak and trough months of the
seasonal pattern, has changed from 25.4 % in 1992 to 19.3 % in 2013. It is apparent
that the seasonal amplitude is decreasing since 2005.

2.4.1 Seasonal Adjustment Methods

The seasonal adjustment methods developed so far are based on univariate time
series decomposition models with no causal explanation. It is difficult to classify
existing methods into mutually exclusive categories. However, it is possible to group
the majority of seasonal adjustment methods into two main classes: one based on
moving averages or linear filters and the other on explicit models with a small
number of parameters for each component. These methods will be extensively
discussed in the following chapters and only a brief summary is provided in this
section.
42 2 Time Series Components

2.4.1.1 Moving Average Methods

The best known and most often applied seasonal adjustment methods are based on
moving averages or smoothing linear filters applied sequentially by adding (and
subtracting) one observation at a time. These methods assume that the time series
components change through time in a stochastic manner. Given a time series, yt ; t D
1; : : : ; n, for any t far removed from both ends, say mC1  t  nm, the seasonally
adjusted value yat is obtained by application of a symmetric moving average W.B/,
as follows:

X
m
yat D W.B/yt D wj ytj (2.20)
jDm

where the weights wj are symmetric, that is, wj D wj , and the length of the average
is 2m C 1.
For current and recent data (n  m < t  n) a symmetric linear filter cannot be
applied, and therefore truncated asymmetric filters are used. For example, for the
last available observation yn , the seasonally adjusted value is given by

X
m
yan D W0 .B/yn D w0;j ynj : (2.21)
jD0

The asymmetric filters are time-varying in the sense that different filters are
applied for the m C 1 first and last observations. The end estimates are revised
as new observations are added because of: (1) the new innovations and (2) the
differences between the symmetric and asymmetric filters. The estimates obtained
with symmetric filters are often called “final.”
The development of electronic computers contributed to major improvements
in seasonal adjustment based on moving averages and facilitated their massive
application. In 1954, Julius Shiskin of the US Bureau of Census developed a
software called Method I, based mainly on the works of Macauley [33] already
being used by the US Federal Reserve Board. Census Method I was followed by
Census Method II and eleven more experimental versions (X1 , X2,: : :, X11). The
best known and widely applied was the Census Method II-X11 variant developed by
Shiskin et al. [39], but produced poor seasonally adjusted data at the end of the series
which is of crucial importance to assess the direction of the short-term trend and the
identification of turning points in the economy. Estela Bee Dagum developed in
1978 [16] a variant she called the X11ARIMA to correct for this serious limitation.
2.4 The Seasonal Variations 43

The X11ARIMA method consists of:

(i) modeling the original series with an ARIMA model of the Box and Jenkins
type,
(ii) extending the original series 1–3 years with forecasts from the ARIMA model
that fits and extrapolates well according to well-defined acceptance criteria,
and
(iii) estimating each component with moving averages that are symmetric for
middle observations and asymmetric for both end years. The latter are obtained
via the convolution of Census II-X11 variant and the ARIMA model extrapo-
lations.
For flow series, deterministic components such as trading day variations and
Easter holiday effects are estimated with dummy variable regression models and
removed from the series, so that only the remainder is subject to steps (i)–(iii)
above. The X11ARIMA was extended again by Dagum in 1988 [18] and, in
1998, David Findley et al. [21] developed X12ARIMA that offers a regARIMA
option to estimate deterministic components, such as trading day variations and
moving holiday effects, simultaneously with the ARIMA model for extrapolation.
It also includes new diagnostic tests and spectral techniques to assess the goodness
of the results. The X12ARIMA method is today the one most often applied by
statistical agencies in the world. The US Bureau of Census continued research on the
development of seasonal adjustment methods and recently produced a beta version
called X13ARIMA-SEATS which enables the estimation of the seasonal component
either via linear filters as those available in X12ARIMA or based on an ARIMA
decomposition model.
All the seasonal adjustment methods based on moving averages discussed in
this book are nonlinear. Hence, the seasonally adjusted total of aggregated series
is not equal to the algebraic sum of the seasonally adjusted series that enter into the
aggregation. The main causes of nonlinearity are
1. a multiplicative decomposition model for the unobserved components,
2. the identification and replacement of extreme values,
3. the ARIMA extrapolations, and
4. automatic selection of moving average length for the estimation of the trend-
cycle and seasonality.
The properties of the combined linear filters applied to estimate the various
components were originally calculated by Young [43] for the standard option of
Census II-X11 variant. Later, Dagum et al. [19] calculated and analyzed all possible
filter combination of Census II-X11 and X11ARIMA. Cleveland and Tiao [14] and
Burridge and Wallis [10] found ARIMA models that approximated well some of the
linear filters used for the trend-cycle and seasonal component of Census II-X11.
44 2 Time Series Components

2.4.1.2 Model-Based Seasonal Adjustment Methods

The best known model-based methods are: (1) the regression methods with global
or locally deterministic models for each component, and (2) stochastic model-based
methods that use ARIMA models.

2.4.1.2.1 Regression Methods

Seasonal adjustment by regression methods is based on the assumption that the

systematic components of time series can be closely approximated by simple
function of time over the entire span of the raw series. In general, two types
of mathematical functions are considered. One is a polynomial of fairly low
degree to represent the trend component; the other, linear combinations of sine
and cosine functions, with different periodicity and fixed amplitude and phase,
to represent business cycles and seasonality. To overcome the limitation of using
global deterministic representations for the trend, cycle, and seasonality, regression
methods were extended to incorporate stochastic representations by means of
local polynomials (spline functions) for successive short segments of the series
and introducing changing seasonal amplitudes. A major breakthrough in this
direction was made by Akaike [1] who introduced prior constraints to the degree of
smoothness of the various components and solved the problem with the introduction
of a Bayesian model. Another important contribution is the regression method with
locally deterministic models estimated by the LOESS (locally weighted regression)
smoother developed by Cleveland et al. [13].
The simplest and often studied seasonal adjustment regression method assumes
that the generating process of seasonality can be represented by strictly periodic
functions of time of annual periodicity. The problem is to estimate s seasonal
coefficients (s being the seasonal periodicity, e.g., 12 for monthly series and 4
for quarterly series) subject to the seasonal constraint that they sum to zero. This
regression seasonal model can be written as

yt D S t C et ; t D 1; : : : ; n
X
s X
s
St D j djt subject to j D 0 (2.22)
jD1 jD1

here the dj ’s are dummy variables taking a value of unity in season j and a value of
zero otherwise; fet g  WN.0; e2 / is a white noise process with zero mean, constant
variance, and non-autocorrelated. The seasonal parameters j ’s can be interpreted
as the expectation of Yt in each season and represent the seasonal effects of the
series under question. Since the parameters are constant, the seasonal pattern repeats
exactly over the years, that is,

St D Sts : (2.23)
2.4 The Seasonal Variations 45

Equation (2.22), often included in econometric models, assumes that the effects
of the seasonal variations are deterministic because they can be predicted with
no error. However, for most economic and social time series, seasonality is not
deterministic but changes gradually in a stochastic manner. One representation of
stochastic seasonality is to assume in regression (2.22) that the j ’s are random
variables instead of constant. Hence, the relationship given in Eq. (2.23) becomes

St D Sts C !t ; 8t > s (2.24)

where f!t g  WN.0; !2 / and E.!t et / D 0; 8t. The stochastic seasonal balance
P
constraint is here given by s1jD0 Stj D !t with expected value equal to zero.
Equation (2.24) assumes seasonality to be generated by a nonstationary stochas-
tic process. In fact, it is a random walk process and can be made stationary by
seasonal differencing .1  Bs /, where B denotes the backward shift operator. In
reduced form, the corresponding regression model is a linear combination of white
noise processes of maximum lag s. Since .1  Bs / D .1  B/.1 C B P C    C Bs1 /,
s1 j
model-based seasonal adjustment methods attribute only .B/ D jD0 B to the
seasonal component leaving .1  B/ as part of a nonstationary trend. Thus, the
corresponding seasonal stochastic model can be written as

.B/St D !t : (2.25)

Equation (2.25) represents a seasonality that evolves in an erratic manner and it was
included in seasonal adjustment methods based on structural model decomposition
(see, e.g., [23] and [31]). This seasonal behavior is seldom found in observed time
series where seasonality evolves smoothly through time. To represent this latter, [27]
introduced a major modification

.B/St D s .B/bt ; (2.26)

where the left-hand side of Eq. (2.26) follows an invertible moving average process
of maximum order s  1, denoted by MA.s  1/. Equation (2.26) is discussed
extensively in [3] and can be generalized easily by replacing the MA part with
stationary invertible autoregressive moving average (ARMA) processes of the [5]
type. Depending on the order of the ARMA process a large variety of evolving
stochastic seasonal variations can be modeled.
Other seasonal models used in regression methods, where the components are
assumed to be local polynomials of time, are based on a smoothness criterion. For
example, given yt D SP
P t C et ; t D 1; : : : ; n, restrictions are imposed on St , such that
n 2 2
tD1 .S t  S ts / and t Œ.B/St be small. The solution is given by minimizing a
weighted linear combination of both criteria as follows:

X
n X
n X
n
.yt  St /2 C ˛ .St  Sts /2 C ˇ ŒS.B/St 2 (2.27)
tD1 tD1 tD1
46 2 Time Series Components

where the ill-posed problem of choosing the ˛ and ˇ parameters was solved by
Akaike [1] using a Bayesian model.

2.4.1.2.2 Stochastic Model-Based Methods

Stochastic model-based methods were mainly developed during the 1980s following
two different approaches. One, originally known as seasonal adjustment by SIGEX
(Signal Extraction), was developed by Burman [8] and as ARIMA model-based
seasonal adjustment by Hillmer and Tiao [27], largely discussed by Bell and Hillmer
[3]. The other is referred to as structural model decomposition method (see, e.g.,
[23] and [31]). The main difference between these two approaches is that in the
latter simple ARIMA models are directly specified for each unobserved component,
whereas in the former an overall ARIMA model is obtained from observable data
and, by imposing certain conditions, models for each component are derived. Since
the components are unknown, to obtain a unique decomposition Hillmer and Tiao
proposed a canonical decomposition which has the properties of maximizing the
variance of the irregulars and minimizing the variance of the other components.
ARIMA models identification and estimation are sensitive to outliers or extreme
values and cannot deal with deterministic components, such as trading days and
moving holidays. Therefore, further developments were made by combining dummy
variables regression models with ARIMA models to deal with these cases. In
this regard, Gomez and Maravall [22] developed at the Bank of Spain a seasonal
adjustment software called TRAMO-SEATS which is currently applied mainly
by European statistical agencies. TRAMO stands for “Time Series Regression
with ARIMA noise, Missing observations and Outliers” and SEATS for “Signal
Extraction in ARIMA Time Series.” First, TRAMO estimates via regression dummy
variables and direct ARIMA modeling, the deterministic components which are
after removed from the input data. In a second round, SEATS estimates the season-
ality and trend-cycle components from an ARIMA model fitted to the modified data
where the deterministic components are removed. SEATS uses the filters derived
from the linearized ARIMA model that describes the behavior of the time series. It
should be mentioned that Eurostat, with the collaboration of the Bank of Belgium,
the US Bureau of the Census, the Bank of Spain, and the European Central Bank,
is developing an interface of TRAMO-SEATS and X12ARIMA called Demetra+,
and more recently JDemetra+. At the Bank of Spain and Eurostat web site it is also
possible to find a considerable number of papers relevant to TRAMO-SEATS as
well as in the European Statistical System (ESS) Guidelines.
On the other hand, the structural model decomposition method starts directly
with an observation equation (sometimes called measurement equation) that relates
the observed time series to the unobserved components. Simple ARIMA or stochas-
tic trigonometric models are assumed for each component. The structural model
is cast in a state space form and estimated with Kalman filtering and smoothing.
Koopman et al. [32] developed STAMP, that stands for Structural Time series
Analyzer, Modeler and Predictor, and includes several types of models for each
2.5 Calendar Variations 47

component. STAMP is set up in an easy-to-use form which enables the user to

concentrate on model selection and interpretation. STAMP 8 is an integrated part of
the OxMetrics modular software system for time series data analysis, forecasting,
financial econometric modeling, and the statistical analysis of cross-section and
panel data. STAMP is not used officially by statistical agencies, but mainly by
econometricians and academics.

2.5 Calendar Variations

The main calendar variations are moving holidays and trading days.

2.5.1 The Moving Holiday Component

The moving holiday or moving festival component is attributed to calendar varia-

tions, namely due to the fact that some holidays change date in successive years.
For example, Easter can fall between March 23 and April 25. The Chinese New
Year date depends on the lunar calendar. Ramadan falls 11 days earlier from year to
year. In the Moslem world, Israel and in the Far East, there are many such festivals.
For example, Malaysia contends with as many as 11 moving festivals, due to its
religious and ethnic diversity. These festivals affect the time series and may cause a
displacement of activity from 1 month to the previous or the following month. For
example, an early date of Easter in March or early April can cause an important
excess of activity in March and a corresponding short-fall in April, in variables
associated with imports, exports, and tourism. When Easter falls late in April (e.g.,
beyond the 10-th), the effect is captured by the seasonal factor of April. In the
long run, Easter falls in April 11 times out of 14. Some of these festivals have a
positive impact on certain variables, for example, air traffic, sales of gasoline, hotel
occupancy, restaurant activity, sales of flowers and chocolate (in the case of Easter).
The impact may be negative on other industries or sectors which close or reduce
their activity during these festivals.
The festival effect may affect only the day of the festival itself, or a number of
days preceding and/or following the festival. In the case of Easter, travelers tend
to leave a few days before and return after Easter, which affects air traffic and
hotel occupancy, etc., for a number of days. Purchases of flowers and other highly
perishable goods, on the other hand, are tightly clustered immediately before the
Easter date.
The effect of moving festivals can be seen as a seasonal effect dependent on the
date(s) of the festival. For illustrative purposes, Fig. 2.3 displays the Easter effect
on US Imports of Goods from Canada observed from January 1985 to December
2013. In this particular case, the Easter effect is rather mild. In some of the years,
48 2 Time Series Components

102
100
99
98
97

1985 1990 1995 2000 2005 2010 2015

Fig. 2.3 Moving Holiday component of US Imports of Goods from Canada

Table 2.1 Dates of Easter April 7 1985 April 23 2000, no effect

and presence of effect in
March 30 1986 April 15 2001, no effect
March
April 19 1987 no effect March 31 2002
April 4 1988 April 20 2003, no effect
March 26 1989 April 11 2004, no effect
April 15, 1990 no effect March 27 2005
March 31 1991 April 16 2006, no effect
April 19 1992, no effect April 8 2007, no effect
April 11 1993, no effect April12 2009, no effect
April 3 1994 April 4 2010
April 16 1995, no effect April 24 2011, no effect
April 7 1996 April 8 2012, no effect
March 30 1997 March 31 2013
April 12 1998, no effect April 20 2014, no effect
April 4 1999

the effect is absent because Easter fell too late in April. The dates of Easter appear
in Table 2.1.
In the case illustrated, the effect is felt 1 day before Easter and on Easter Sunday
but not after Easter. This is evidenced by years 1985, 1986, 1989, 1991, 1997,
2002, 2005, and 2013, where Easters falls early in April or in March. Imports are
substantially affected by Easter, because customs do not operate from Good Friday
2.5 Calendar Variations 49

to Easter Monday. Easter can also significantly affect quarterly series, by displacing
activity from the second to the first quarter.
Generally, festival effects are difficult to estimate, because the nature and
the shape of the effect are often not well-known. Furthermore, there are few
observations, i.e., one occurrence per year.

2.5.2 The Trading Day Component

Time series may be affected by other variations associated with the composition of
the calendar. The most important calendar variations are the trading day variations,
which are due to the fact that the activities of some days of the week are more
important than others. Trading day variations imply the existence of a daily pattern
analogous to the seasonal pattern. However, these daily factors are usually referred
to as daily coefficients.
Depending on the socioeconomic variable considered, the activity of some days
may be 60 % more important than an average day and other days, 80 % less
important.
If the most important days of the week appear five times in a month (instead
of four), the month registers an excess of activity ceteris paribus. If the least
important days appear five times, the month records a short-fall. As a result, the
monthly trading day component can cause variations of +8 % or 8 % (say) between
neighboring months and also between same months of neighboring years. The
trading day component is usually considered as negligible and very difficult to
estimate in quarterly series.
Figure 2.4 displays the monthly trading day component obtained from the
following daily pattern: 101.55, 102.26, 100.75, 99.99, 97.85, 97.73, and 99.95 for
Monday to Sunday (in percentage), respectively. The daily pattern indicates that
Tuesday, Friday, and Saturday are approximately 2.2 % more important than an
average day (100 %); and that Wednesday, Thursday, and Sunday are less important.
For the multiplicative, the log-additive and the additive time series decomposition
models, the monthly trading day component is respectively obtained in the following
manner:
X X
Dt D d =nt D .2800 C d /=nt (2.28)
 2t  2t5
! " ! #
X X
Dt D exp d =nt D exp d =nt (2.29)
 2t  2t5
X X
Dt D d D d (2.30)
 2t  2t5
50 2 Time Series Components

102
101
100
99
98

1985 1990 1995 2000 2005 2010 2015

Fig. 2.4 Trading day estimates of US Imports of Goods from Canada

where nt is the number of days in month t, t5 is the set of days of the week that
appears five times in month t, and d are the daily coefficients in the month. Setting
nt equal to the number of days in month t implies that the length-of-month effect
is captured by the multiplicative seasonal factors, except for Februaries. To adjust
Februaries for the length-of-month, the seasonal factors of that month are multiplied
by 29/28.25 and 28/28.25 for the leap and non-leap years, respectively.
The other option is to set nt equal to 30.4375, so that the multiplicative trading
day component also accounts for the length-of-month effect. The number 2800 in
Eq. (2.28) is the sum of the first 28 days of the months expressed in percentage.
The monthly trading day estimates of the US Imports of Goods from Canada
shown in Fig. 2.4 were obtained with the log-additive model (2.29). They display
a drop of 5.45 % between February and March 1996 and an increase of 6.52 %
between January and February 2000.
One can identify several instances where the change between same-months
is significant. Indeed, same-month year-ago comparisons are never valid in the
presence of trading day variations, not even as a rule of thumb. Furthermore, it
is apparent that the monthly trading day factors in the figure are identical for quite a
few months. Indeed for a given set of daily coefficients, there are only 22 different
monthly values for the trading day component, for a given set of daily coefficients:
seven values for 31-day months (depending on which day the month starts), seven
for 30-day months, seven for 29-day months, and one for 28-day months. In other
words, there are at most 22 possible arrangements of days in monthly data.
2.5 Calendar Variations 51

Many goods and services are affected by daily patterns of activity, which
entail higher costs for producers, namely through the need of higher inventories,
equipment, and staff on certain days of the week. For example, there is evidence that
consumers buy more gasoline on certain days of the week, namely on Thursdays,
Fridays, Saturdays, and holidays, which results in line-ups and shortages at the
pumps. In order to cope with the problem, gasoline retailers raise their price on
those days to promote sales on other days. Furthermore, the elasticity of demand for
gasoline is low. In other words, to reduce consumption by a small percentage, prices
must be raised by a disproportionate percentage, which upsets some consumers.
On the other hand, consumers can buy their gasoline on other days. The alternative
for retailers is to acquire larger inventories, larger tanks, more pumps, and larger
fleets of tanker trucks, all of which imply higher costs and translate into much
higher prices. In other words, there are savings associated with more uniform daily
patterns; and costs, with scattered daily patterns. A similar consumer behavior
prevails for the purchases of food, which probably results in more expensive
prices, namely through higher inventories, larger refrigerators, more numerous cash
registers, and more staff, than otherwise necessary. Deaths also occur more often
on certain days of the week. Car accidents, drowning, skiing, and other sporting
accidents tend to occur on weekend days and on holidays. In principle, stock series
pertaining to 1 day display a particular kind of trading day variations. Among other
things, inventories must anticipate the activity (flow) of the following day(s). For
such stock series, the monthly trading day factor coincides with the daily weight of
the day.

2.5.2.1 A Classical Model for Trading Day Variation

A classical deterministic model for trading day variations was developed by Young
[42],

y t D Dt C e t ; t D 1; : : : ; n (2.31)
7
X
Dt D ˛j Njt (2.32)
jD1

P
where et  WN.0; e2 /; 7jD1 ˛j D 0; ˛j ; j D 1; : : : ; 7, denote the effects of the 7
days of the week, Monday to Sunday, and Njt is the number of times day j is present
P
in month t. Hence, the length of the month is Nt D 7jD1 Njt , and the cumulative
P
monthly effect is given by (2.32). Adding and subtracting ˛N D . 7jD1 ˛j /=7 to
Eq. (2.32) yields

7
X
Dt D ˛N
N tC .˛j  ˛/N
N jt : (2.33)
jD1
52 2 Time Series Components

Hence, the cumulative effect is givenP by the length

of the month plus the net effect
due to the days of the week. Since 7jD1 ˛j  ˛N D 0, model (2.31) takes into
account the effect of the days present five times in the month. Model (2.33) can then
be written as
6
X
Dt D ˛N
N tC .˛j  ˛/.N
N jt  N7t /; (2.34)
jD1

P
with the effect of Sunday being ˛7 D  6jD1 ˛j .
Deterministic models for trading day variations assume that the daily activity
coefficients are constant over the whole range of the series. Stochastic model for
trading day variations has been rarely proposed. Dagum et al. [20] developed a
model where the daily coefficients change over time according to a stochastic
difference equation.

2.6 The Irregular Component

The irregular component in any decomposition model represents variations related

to unpredictable events of all kinds. Most irregular values have a stable pattern,
but some extreme values or outliers may be present. Outliers can often be traced
to identifiable causes, for example, strikes, droughts, floods, and data processing
errors. Some outliers are the result of displacement of activity from one month to
the other.
Figure 2.5 shows the irregular component of US New Orders for Durable
Goods (January 1992–December 2013). It includes extreme values and also outliers,
namely in January 2001, October 2008, and January 2009. The former is identified
as an additive outlier, whereas the latter are identified as level shift outliers.
As illustrated in Fig. 2.5, the values of the irregular component may be very
informative, as they quantify the effect of events known to have happened. Note
that it is much easier to locate outliers in the irregular component than in the raw
series because the presence of seasonality hides the irregular fluctuations.

2.6.1 Redistribution of Outliers and Strikes

Some events can cause displacements of activity from one month to the next months,
or vice versa. This phenomenon is referred to as redistribution of outliers. We also
deal with the strike effect under this headline. The outliers must be modeled and
temporally removed from the series in order to reliably estimate the systematic
components, namely the seasonal and trading day components. Events such as
2.6 The Irregular Component 53

1.15
1.10
1.05
1.00
0.95

1995 2000 2005 2010

Fig. 2.5 Irregular component of US New Orders for Durable Goods (January 1992–December
2013)

major snow storms and power blackouts usually postpone activity to the next month,
without much longer term effect.

2.6.2 Models for the Irregulars and Outliers

The irregulars are most commonly assumed to follow a white noise process fIt g
defined by

E.It / D 0; E.It2 / D I2 < 1; E.It Itk / D 0 if k¤0

If I2 is assumed constant (homoscedastic condition), fIt g is referred to as white

noise in the strict sense.
If I2 is finite but not constant (heteroscedastic condition), fIt g is called white
noise in the weak sense.
For inferential purposes, the irregular component is often assumed to be normally
identically distributed and not correlated, which implies independence. Hence, It 
NID.0; I2 /.
There are different models proposed for the presence of outliers depending on
how they impact the series under question. If the effect is transitory, the outlier is
said to be additive, and if permanent, to be multiplicative.
54 2 Time Series Components

Box and Tiao [6] introduced the following intervention model to deal with
different types of outliers:
1
X 1
X
yt D hj xtj C t D hj Bj xt C t D H.B/xt C t ;
jD0 jD0

where the observed series fyt g consists of an input series fxt g considered a
deterministic function of time and a stationary process ft g of zero mean and
P with fxt g. In such a case the mean of fyt g is given by the deterministic
uncorrelated
function 1 jD0 hj xtj . The type of function assumed for fxt g and weights fhj g depend
on the characteristic of the outlier or unusual event and its impact on the series.
If the outlier at time t0 is additive, in the sense that it will not permanently modify
the level of fyt g, that is, E.yt / D 0; 8t  t0 , and also E.yt / ! 0 for t ! t0 , then an
appropriate xt is the impulse function defined by

1; t D t0 ;
xt D Pt .t0 / D
0; t ¤ t0 :

If, instead, the outlier at time t0 is multiplicative in the sense that it will modify the
level of fyt g, that is, E.yt / D 0; 8t  t0 , and also E.yt / ! c ¤ 0 for t ! t0 , then an
appropriate xt is the step function defined by

1; t ¤ t0 ;
xt D St .t0 / D
0; t < t0 :

In fact, St .t0 /  St1 .t0 / D .1  B/St .t0 / D Pt .t0 /.

Once the deterministic function is chosen, the weights fht g can follow different
patterns dependent on the impact of the outlier. For example, if the outlier is additive
and present at time t D t0 , an appropriate model is

!; t D t0
H.B/xt D !Pt .t0 / D
0; t ¤ t0

where ! represents the effect of the outlier.

If the outlier is seasonal and additive starting at t D t0 , an appropriate model is

!; t D t0 ; t0 C s; t0 C 2s; : : :
H.B/xt D !Pt .t0 C ks/ D
0; otherwise
2.6 The Irregular Component 55

For additive redistribution of outliers (displacement of activity), during period t D

t0 ; t D t0 C 1; : : : ; t D t0 C k, an appropriate model is
8 P
ˆ
ˆ !0 D  kiD1 !i ; t D t0
ˆ
ˆ
ˆ
< !1 ; t D t0 C 1
Xk
H.B/xt D !0 Pt .t0 / C !i Pt .t0 C i/ D 
ˆ
ˆ
iD1 ˆ
ˆ !k ; t D t0 C k
:̂
0; otherwise

where the weights !0 ; : : : ; !k measure the outlier effects during the k C 1 periods
t0 ; t0 C 1; : : : ; t0 C k. The global effect is such that the sum of all the outlier effects
cancel out.
For a sudden permanent change of level at time t D t0 , a step function can be
used

!; t  t0
H.B/xt D !0 St .t0 / D
0; otherwise

where ! represent the level difference of the series before and after the outlier. Since
.1  B/St .n/ D Pt .n/, we can write

! !; t D t0
H.B/xt D Pt .t0 / D
.1  B/ 0; otherwise
P P1
where !.1B/1 D !.1CBCB2 C   /Pt .t0 / D 1 jD0 B Pt .t0 / D
j
jD0 !Pt .t0 Cj/:
For a sudden transitory change at time t D t0 , an appropriate model is
8
ˆ
ˆ 0; t < t0
! <
!; t D t0
H.B/xt D Pt .t0 / D
.1  ıB/ ˆ ı 2 !; t D t0 C 1
:̂ 2
ı !; t D t0 C 2

where ! denotes the initial effect, and 0 < ı < 1 is the rate of decrease of the initial
effect.
For a gradual permanent level change at time t D t0 , an appropriate model is
8
ˆ
ˆ 0; t < t0
ˆ
ˆ
ˆ
ˆ !; t D t0
ˆ
<
! .1 C ı/!; t D t0 C 1
H.B/xt D St .t0 / D 2
.1  ıB/ ˆ C ı C ı /!; t
ˆ
ˆ
.1 D t0 C 2
ˆ
ˆ 
ˆ
:̂ !
.1ı/ ; t!1
56 2 Time Series Components

The intervention model proposed by Box and Tiao [6] requires the location
and the dynamic pattern of an event to be known. Chen and Liu [12] consider
the estimation problem when both the location and the dynamic pattern are not
known a priori. The proposed iterative procedure is less vulnerable to the spurious
and masking effects during outlier detection and allows to jointly estimate the
model parameters and multiple outlier effects. This procedure is the one actually
implemented in X12ARIMA and TRAMO-SEATS software.

References

1. Akaike, H. (1980). Seasonal adjustment by a Bayesian modelling. Journal of Time Series

Analysis, 1, 1–13.
2. Baxter, M., & King, R. G. (1995). Measuring business cycles approximate band-pass filters for
economic time series. NBER Working Papers 5022, National Bureau of Economic Research,
Inc.
3. Bell, W. H., & Hillmer, S. C. (1984). Issues involved with the seasonal adjustment of economic
time series. Journal of Business and Economic Statistics, 2, 291–320.
4. Bernanke, B. S. (2004). Gradualism. Speech 540, Board of Governors of the Federal Reserve
System (U.S.).
5. Box, G. E. P., & Jenkins, G. M. (1970). Time series analysis: Forecasting and control. San
Francisco, CA, USA: Holden-Day (Second Edition 1976).
6. Box, G. E. P., & Tiao, G. C. (1975). Intervention analysis with applications to economics and
environmental problems. Journal of the American Statistical Association, 70, 70–79.
7. Bry, G., & Boschan, C. (1971). Cyclical analysis of time series: Selected procedures and
computer programs. NBER Technical Paper n. 20.
8. Burman, J. P. (1980). Seasonal adjustment by signal extraction. Journal of the Royal Statistical
Society Series A, 143, 321–337.
9. Burns, A. F., & Wesley, M. C. (1946). Measuring business cycles. New York: National Bureau
of Economic Research.
10. Burridge, P., & Wallis, K. F. (1984). Unobserved components models for seasonal adjustment
filters. Journal of Business and Economic Statistics, 2(4), 350–359.
11. Butterworth, S. (1930). On the theory of filter amplifiers. Experimental Wireless and the Radio
Engineer, 7, 536–541.
12. Chen, C., & Liu, L. M. (1993). Joint estimation of model parameters and outlier effects in time
series. Journal of the American Statistical Association, 88(241), 284–297.
13. Cleveland, R. B., Cleveland, W. S., McRae, J. E., & Terpenning, I. (1990). STL: A seasonal
trend decomposition procedure based on LOESS. Journal of Official Statistics, 6(1), 3–33.
14. Cleveland, W. P., & Tiao, G. C. (1976). Decomposition of seasonal time series: A model for
the census X-11 program. Journal of the American Statistical Association, 71, 581–587.
15. Dagum, C., & Dagum, E. B. (1988). Trend. In S. Kotz & N. L. Johnson (Eds.), Encyclopedia
of statistical sciences (Vol. 9, pp. 321–324). New York: Wiley.
16. Dagum, E. B. (1978). Modelling, forecasting and seasonally adjusting economic time series
with the X11ARIMA method. The Statistician, 27(3), 203–216.
17. Dagum, E. B. (1980). The X11ARIMA seasonal adjustment method. Statistics Canada, Ottawa,
Canada. Catalogue No. 12-564.
18. Dagum, E. B. (1988). The X11ARIMA/88 seasonal adjustment method-foundations and user’s
manual. Ottawa, Canada: Time Series Research and Analysis Centre, Statistics Canada.
19. Dagum, E. B., Chhab, N., & Chiu, K. (1990). Derivation and properties of the X11ARIMA
and census X11 linear filters. Journal of Official Statistics, 12(4), 329–347.
References 57

20. Dagum, E.B., Quenneville, B., & Sutradhar, B. (1992). Trading-day variations multiple
regression models with random parameters. International Statistical Review, 60(1), 57–73.
21. Findley, D. F., Monsell, B. C., Bell, W. R., Otto, M. C., & Chen, B. C. (1998). New capabilities
and methods of the X12ARIMA seasonal adjustment program. Journal of Business and
Economic Statistics, 16(2), 127–152.
22. Gomez, V., & Maravall, A. (1996). Program TRAMO and SEATS. Madrid: Bank of Spain.
23. Harvey, A. C. (1981). Time series models. Oxford, UK: Phillip Allan Publishers Limited.
24. Harvey, A. C. (1985). Trends and cycles in macroeconomic time series. Journal of the Business
and Economic Statistics, 3, 216–227.
25. Harvey, A. C., & Trimbur, T. M. (2003). General model-based filters for extracting cycles and
trends in economic time series. The Review of Economics and Statistics, 85(2), 244–255.
26. Henderson, R. (1916). Note on graduation by adjusted average. Transaction of Actuarial
Society of America, 17, 43–48.
27. Hillmer, S. C., & Tiao, G. C. (1982). An ARIMA-model-based approach to seasonal adjust-
ment. Journal of the American Statistical Association, 77, 63–70.
28. Hodrick, R. J., & Prescott, E. C. (1997). Postwar U.S. business cycles: An empirical
investigation. Journal of Money, Credit and Banking, 29(1), 1–16.
29. Kaiser, R., & Maravall, A. (2002). A complete model-based interpretation of the Hodrick-
Prescott filter: Spuriousness reconsidered. Banco de Espana Working Papers 0208, Banco de
Espana.
30. Kaiser, R., & Maravall, A. (2005). Combining filter design with model-based filtering (with
an application to business-cycle estimation). International Journal of Forecasting, 21(4), 691–
710.
31. Kitagawa, G., & Gersch, W. (1984). A smoothness priors state-space modelling of time series
with trend and seasonality. Journal of the American Statistical Association, 79, 378–389.
32. Koopman, S. J., Harvey, A. C., Doornik, J. A., & Shephard, N. (1998). Structural time series
analysis STAMP (5.0). London: Thomson Business Press.
33. Macaulay, F. R. (1931). The smoothing of time series. New York: National Bureau of Economic
Research.
34. Malthus, T. R. (1798). An essay on the principle of population. Oxford World’s Classics reprint:
xxix Chronology.
35. Maravall, A. (1993). Stochastic and linear trend models and estimators. Journal of Economet-
rics, 56, 5–37.
36. Moore, G. H. (1961). Business cycle indicators. Princeton, NJ: Princeton University Press.
37. Persons, W. M. (1919). Indices of business conditions. Review of Economic Statistics, 1, 5–107.
38. Schumpeter, J. A. (1954). History of economic analysis. New York: Oxford University Press
Inc.
39. Shiskin, J., Young, A. H., & Musgrave, J. C. (1967). The X-11 variant of the Census Method
II seasonal adjustment program. Technical Paper 15 (revised). US Department of Commerce,
Bureau of the Census, Washington, DC.
40. Stock, J. H., & Watson, M. W. (2003). Has the business cycle changed and why?. NBER
macroeconomics annual 2002 (Vol. 17, pp. 159–230). Cambridge: MIT Press.
41. Wold, H. O. (1938). A study in the analysis of stationary time series. Uppsala, Sweden:
Almquist and Wiksell (Second Edition 1954).
42. Young, A. H. (1965). Estimating trading-day variations in monthly economic time series. US
Bureau of the Census, Technical Report No. 12.
43. Young, A. H. (1968). A linear approximation to the Census and BLS seasonal adjustment
methods. Journal of the American Statistical Association, 63, 445–457.
 Part I
Seasonal Adjustment Methods
Chapter 3
Seasonal Adjustment: Meaning, Purpose, and
Methods

Abstract This chapter deals with the causes and characteristics of seasonality.
For gradual changes in seasonality whether of a stochastic or a deterministic type,
various models have been proposed that can be grouped into two broad categories:
(1) models that assume that the generating process of seasonality varies only in
amplitude and (2) models that assume that the generating process varies in both
amplitude and phase. The basic assumptions of both groups of models are studied
with particular reference to the two seasonal adjustment methods officially adopted
by statistical agencies, the X12ARIMA and TRAMO-SEATS. The economic
significance of seasonality and the need for seasonal adjustment are also discussed.
Since seasonality ultimately results mainly from noneconomic forces (climatic and
institutional factors), external to the economic system, its impact on the economy as
a whole cannot be modified in a short period of time. Therefore, it is of interest for
policy making and decision taking to have the seasonal variations removed from the
original series. The main reason for seasonal adjustment is the need of standardizing
socioeconomic series because seasonality affects them with different timing and
intensity.

3.1 Seasonality, Its Causes and Characteristics

A great deal of data in business, economics, and natural sciences occurs in the
form of time series where observations are dependent and where the nature of
this dependence is of interest in itself. The time series is generally compiled for
consecutive and equal periods, such as weeks, months, quarters, and years. From a
statistical point of view, a time series is a sample realization of a stochastic process,
i.e., a process controlled by probability laws. The observations made as the process
continues indicate the way it evolves.
In the analysis of economic time series, Persons [44] was the first to distinguish
four types of evolution, namely: (a) the trend, (b) the cycle, (c) the seasonal
variations, and (d) the irregular fluctuations.
Among all these components, the influence of the seasonal fluctuations in the
human activity has been felt from earlier times. The organization of society, the
means of production and communication, the habits of consumption, and other

© Springer International Publishing Switzerland 2016 61

E. Bee Dagum, S. Bianconcini, Seasonal Adjustment Methods and Real Time
Trend-Cycle Estimation, Statistics for Social and Behavioral Sciences,
DOI 10.1007/978-3-319-31822-6_3
62 3 Seasonal Adjustment: Meaning, Purpose, and Methods

social and religious events have been strongly conditioned by both climatic and
conventional seasons. The seasonal variations in agriculture, the high pre-Easter and
pre-Christmas retail sales, and the low level of winter construction are all general
knowledge. The main causes of seasonality, the climatic and institutional factors
are mainly exogenous to the economic system and cannot be controlled or modified
by the decision makers in the short run. The impact of seasonality in the economic
activity, however, is usually not independent on the stages of the business cycle.
It is well-known, for example, that the seasonal unemployment among adult males
is much larger during periods of recession than during periods of prosperity.
Another main feature of seasonality is that the phenomenon repeats with certain
regularity every year but it may evolve. The latter is mainly due to systematic
intervention of the government and to technological and institutional changes as
well; and it is more the general case than the exception for economic time series.
Therefore, the assumption of stable seasonality, i.e., of seasonality being repre-
sentable by a strictly periodic function, can be used as a good approximation for few
series only. In effect, even in the extreme cases of those activities where seasonality
is mainly caused by climatic conditions, e.g., agriculture, fishing, forestry, the
seasonal variations change, for weather itself measured by the temperature, and
quantity of precipitation changes. There are many other reasons that can produce
temporal shifts in the seasonal effects. A decline in the participation of the primary
sector in the Gross National Product will modify seasonal patterns in the economy
as a whole, as will a change in the geographical distribution of industry in a country
extending over several climatic zones. Changes in technology alter the importance
of climatic factors. Customs and habits change with a different distribution of
income and, thus, modify the demand for certain goods and its corresponding
seasonal pattern.
Once the assumption of stable seasonality is abandoned, new assumptions must
be made regarding the nature of its evolution; if seasonality changes, is it slowly or
rapidly? Is it gradually or abruptly? Is it in a deterministic or a stochastic manner?
Today, the most widely accepted hypothesis is that seasonality moves gradually,
slowly, and in a stochastic manner. For gradual changes in seasonality whether of a
stochastic or a deterministic type, various models have been studied. These models
can be grouped into two broad categories:
1. models that assume the generating process of seasonality varies in amplitude
only and
2. models that assume the generating process varies in both amplitude and phase.
Kuznets [38] and Wald [49] were the first to propose an estimation procedure that
assumed seasonality changes in amplitude while the pattern (i.e., the proportionality
relationship between the seasonal effects of each month) remained constant over
time.
On the other hand, the majority of the methods officially adopted by statistical
agencies belongs to the second group. Seasonality is assumed to change either
in a stochastic or in a deterministic manner. The stochastic approach has been
analyzed among many others by Hannan [23], Hannan et al. [24], Brewer et al. [5],
3.2 The Economic Significance of Seasonality and the Need for Seasonally. . . 63

Cleveland [9], and Cleveland and Tiao [10]. Pierce [45] proposed a mixed model
where seasonality can be in part deterministic and in part stochastic.
A third characteristic of seasonality is that the phenomenon can be distinguished
from other movements (trend, cycle, and irregulars) that influence an economic time
series and can be measured. The seasonal variations are distinguished from trend
by their oscillating character, from the cycle by being confined within the limits
of an annual period, and from the irregulars, by the fact of being systematic. For
example, the causes of seasonality in labor force series are found in the main factors
that influence their demand and supply. The seasonal variations in employment and
unemployment of adult males are strongly conditioned by the stages of the business
cycle and the weather. There is practically no seasonality in the supply of adult
males labor but there is a large amount of seasonal variation in the demand side.
This group is mainly absorbed by the primary sector (construction, agriculture,
mining, fishing, and forestry) where seasonality is mostly climatic with very large
oscillation, and by the industrial sector, where seasonality is mostly induced by the
seasonality in the primary sector but it is also strongly affected by the stages of the
cycle. On the other hand, seasonality for females and young males originates from
the side of the demand as well as of the supply. In fact, females and young males
are mainly employed by the tertiary sector where seasonality tends to be originated
by institutional and religious events (Christmas, Easter, Federal Taxes deadlines).
Similarly, from the viewpoint of the supply, seasonal variations occur because this
group tends to move in and out of the labor force in accordance with the school year.

3.2 The Economic Significance of Seasonality and the Need

for Seasonally Adjusted Series

The seasonal variations affect the economic sectors in different degrees. Nor is it
only the intensity of the primary effects of the seasons that differs for the various
economic activities, producing seasonal swings of varying size or amplitude, the
timing of these effects is also different, creating different patterns as the season
begins in 1 month or another. Thus, for example, in winter occur both the seasonal
low in construction and the seasonal peak in retail trade. The seasonal amplitudes
for both sectors also differ significantly, being much larger for construction than for
retail trade.
Similar to the business cycle, the presence of seasonality in the economy
introduces a disequilibrium in the allocation of the resources, implying an extra cost.
Because of the seasonal variations, the economy has a supply of labor, equipment,
and raw materials in excess to what it would be required if the activity proceeds
at an even rate throughout the year. Since these surpluses may arise also from
the different impacts of the other components (trend, cycle, and irregulars) on the
economic processes, it becomes very difficult to estimate the actual cost imposed by
seasonality. Some attempts of measurement have been made by Kuznets [38], Baron
64 3 Seasonal Adjustment: Meaning, Purpose, and Methods

[3], Dawson et al. [16], and Judek [33] among many others. For instance, Kuznets
[38] suggested two measures of the degree of overcapacity in industry attributable
to seasonal variations: one gives the total difference in activity between each of the
months and the months of peak output; the other, the amount by which the months
below 100 depart from 100. The basic idea is that the seasonal peak months measure
the normal plus the latent capacity of the industry, therefore, an output below the
seasonal peak indicates an inadequate utilization of the maximum possible capacity.
On the other hand, an output below a seasonal index of 100 indicates an inadequate
utilization of the normal capacity of the industry. The same idea is applicable to the
surplus of labor. If the seasonal peak in employment indicates the number of workers
attached to a given industry, then the average monthly departure from the peak
measures the average monthly seasonal unemployment and hence the surplus of
labor that is in the industry because of seasonal swings. If, however, only the annual
average is considered the correct estimate of labor attached to the industry, then only
departures below 100 should be considered as measuring seasonal unemployment
and excess of labor supply.
Because seasonality ultimately results mainly from noneconomic forces (climatic
and institutional factors), external to the economic system, its impact on the
economy as a whole cannot be modified in a short period of time. Therefore, it is of
interest to decision makers to have the seasonal variations removed from the original
series to obtain a seasonally adjusted series. Seasonal adjustment means the removal
of seasonal variations in the original series jointly with trading day variations and
moving holiday effects. The main reason for seasonal adjustment is the need of
standardizing socioeconomic series because seasonality affects them with different
timing and intensity. In this manner, the variations of a seasonally adjusted series are
due to variations only in the trend, the cycle, and the irregulars. It should be noticed,
however, that the removal of seasonal variations from a time series does not indicate
how the series would have evolved as there had been no seasonal variations, rather
it shows more clearly the trend-cycle abstracting from seasonality.
The information given by seasonally adjusted series plays a very important role
in the analysis of current economic conditions, particularly, in determining the stage
of the cycle at which the economy stands. Such knowledge is useful in forecasting
subsequent cyclical movements and provides the basis for decision making to
control the level of the economic activities. It is particularly important around
turning points since, e.g., failure to recognize the downturn in the business cycle
may lead to the adoption of policies to curb expansion when, in fact, a recession is
under way. Analysts who wish to get a picture of the economic situation not distorted
by exogenous variables related to the seasonal variations, may make comparisons
with the same month of the year before. However, apart from other considerations,
the basic problem with same-month year-ago comparisons is that they show only
what has happened a year after, not what was happening during the year not what is
happening currently. To evaluate the current state of the economy, the analyst should
be able to measure cyclical changes for each month over less than 1 year span, e.g.,
to compare May with April (1-month span) or May with February (3-month span).
3.3 Basic Assumptions of Main Seasonal Adjustment Methods 65

Decision making based on the raw data can lead to wrong policies, particularly
if the series is strongly affected by seasonal variations. The average absolute
month percentage change in the seasonal variation can be much greater than the
corresponding changes in the irregular or trend-cycle. Results of several studies of
selected economic US indicators show that the average absolute month-to-month
percentage changes in the seasonal component run between 3 and 7 times the
average absolute percentage changes in the trend-cycle or in the irregulars over the
same time spans (see Shiskin [46]). The measurement of seasonality is also very
useful for short-term decision making. The knowledge of the seasonal pattern of the
economic activities facilitates a better planning of the economic resources during
the periods of peak loads and of inactivity. This knowledge can also be applied for a
better integration of economic activities characterized by opposite seasonal patterns.
Resources which formerly served only one purpose can serve additional purposes
and thus reduce the burden imposed by seasonality to the whole economy.

3.3 Basic Assumptions of Main Seasonal Adjustment

Methods

The majority of the seasonal adjustment methods developed until now is based
on univariate time series models, where the estimation of seasonality is made in
a simple and mechanical manner and not based on a causal explanation of the
phenomenon under study. Very few attempts have been made to follow the latter
approach and none of them reached further than the experimental stage. Among
others, Mendershausen [43] tried to regress the seasonal effects for each month on a
set of exogenous variables (meteorological and social variables) in order to build an
explanatory model for seasonality but his empirical results were not very fruitful.
On the other hand, univariate time series methods of seasonal adjustment try to
estimate the generating mechanism of the observations under the simple assumption
that the series is composed of a systematic part which is a well-determined function
of time and, a random part which obeys a probability law. The feasibility of
this decomposition was proved in a well-known theorem due to Herman Wold
[52]. Wold showed that any second order stationary process fYt g can be uniquely
represented as the sum of two mutually uncorrelated processes

Yt D Zt C Vt ;

where fZt g is a convergent infinite moving average process, and fVt g a deterministic
process. The decomposition is linear and is determined entirely by the second
moments of the process.
In a broad sense, seasonal methods can be classified as deterministic or stochastic
depending on the assumptions made concerning how seasonality evolves through
time. Deterministic methods assume that the seasonal pattern can be predicted with
66 3 Seasonal Adjustment: Meaning, Purpose, and Methods

no error or with variance of the prediction error null. On the contrary, stochastic
methods assume that seasonality can be represented by a stochastic process, a
process governed by a probability law and, consequently, the variance of the
prediction error is not null. The best known seasonal adjustment methods belong
to the following types:
(a) regression methods which assume global or local simple functions of time,
(b) stochastic-model-based methods which assume simple autoregressive inte-
grated moving average (ARIMA) models, and
(c) moving average methods which are based on linear filtering and hence do not
have explicit parametric models.
Only methods (a), which assume global simple functions for each component,
are deterministic; the others are considered stochastic. Moving average or linear
filtering methods are those adopted mostly by statistical agencies to produce
officially seasonally adjusted series.

3.3.1 Regression Methods

The use of regression methods for the decomposition of a time series is old. In
the late 1930s, Fisher [20] and Mendershausen [43] proposed to fit polynomials
by the least squares method to estimate the seasonal effects. To deal with moving
seasonality, Cowden [11] suggested to fit polynomials to each month seasonal
obtained from the ratios (differences) of the original series and a centered twelve
term moving average. Jones [30] and Wald [49] considered the problem of fitting
simultaneously a polynomial of a relatively low degree to estimate the trend plus
twelve constants (one for each month) representing stable seasonality.
In the 1960s, the use of multiple regression techniques for the seasonal adjust-
ment of economic time series was strongly advocated. The main reasons for this
were the widespread interest in the use of monthly and quarterly series for econo-
metric model building and the fast development of electronic computers. In effect,
the former posed the problem of whether to use as input seasonally adjusted data
directly, or to use first, raw data, and then seasonally adjusted. These two alternatives
do not give the same results unless the seasonal adjustment method is strictly
linear (conditioned fulfilled by the regression methods). The electronic computers
reduced significantly the heavy burden imposed by the computations in the least
squares method despite the existence of several simplifying procedures such as
the orthogonal polynomials and the Buys-Ballot table. Important contributions
were made by Hannan [22], Lovell [40], Ladd [39], Jorgenson [31], Henshaw
[28], Stephenson and Farr [48], and Wallis [50]. Their contributions facilitated the
application of least squares methods for the seasonal adjustment of economic time
series.
3.3 Basic Assumptions of Main Seasonal Adjustment Methods 67

The regression methods assume that the systematic part of a time series can be
approximated closely by simple functions of time over the entire span of the series.
In general, two types of functions of time are considered. One is a polynomial
of fairly low degree that fulfills the assumption that the economic phenomenon
moves slowly, smoothly, and progressively through time (the trend). The other type
of function is a linear combination of sines and cosines of different frequencies
representing oscillations, strictly periodic or not, that affect also the total variation
of the series (the cycle and seasonality).
For certain regression functions such as fixed periodic functions, polynomial
time trends, and interaction of the two types, ordinary least squares estimates
are asymptotically efficient if the random part is second order stationary, i.e.,
the mean and variance are constant and the covariance is function only of the
time lag. Therefore, by taking differences to ensure a stationary process, ordinary
least squares are generally adequate. If the relationship among the components is
multiplicative, the standard procedure is to take logarithms and then, differences, in
order to transform the generating mechanism of the series into an additive form with
a stationary random term.
A large number of model specifications have been considered for monthly series.
Young [53] analyzed and compared the seasonal adjustment obtained by regression
methods with the results given by the Census II-X11 variant which belongs to
the category of linear smoothing techniques. The models analyzed have different
degrees of flexibility concerning the pattern of the seasonals and trend-cycle and the
relationship between the components is assumed to be multiplicative.
Model I

log yt D log St C log Ct C log It (3.1)

6
X
log St D Œ˛1;k sin k C ˛2;k cos k t
kD1

2
X  
C ˇ1;k tk sin k t C ˇ2;k tk cos k t ; (3.2)
kD1

2k
where k D 12 ,

5
X
log Ct D 0 C k tk C ı1 sin !t C ı2 cos !t; (3.3)
kD1

where ! D 2 72 .
This model contains a deterministic stable seasonality represented by the sines
and cosines of periodicity 12, 6, 4, 3, 2.4, and 2 plus a deterministic moving
seasonality that is assumed to follow a second degree polynomial through time.
The trend-cycle expression contains a fifth degree polynomial and sines and cosines
with a periodicity of 72 months that take into account a 5 year cycle.
68 3 Seasonal Adjustment: Meaning, Purpose, and Methods

Model II

log yt D log St C log Ct C log It (3.4)

12
X
log St D ˛k dk (3.5)
kD1
10
X
log Ct D k t k : (3.6)
kD1

Instead of using sines and cosines, this model contains a deterministic stable
seasonality represented by 12 dummy variables dk where dk D 1 if t is the k-th
month of the year and equals 0 otherwise. The trend-cycle curve is represented by a
10-th degree polynomial.
Model III

log yt D log St C log Ct C log It (3.7)

12
X 12
X 12
X
log St D ˛k dk C ˇ0;k dk C ˇ1;k dk t (3.8)
kD1 kD1 kD1
5
X
log Ct D k t k : (3.9)
kD1

The seasonal model corresponds to a deterministic stable seasonality plus a

deterministic linearly moving seasonality. The trend-cycle is represented by a fifth
degree polynomial.
Young [53] concluded that for central observations some of the model specifica-
tions for the trend-cycle and the seasonals approached closely the weights applied
by the Census II - X11 program, but not for the observations at both ends. Moreover,
he found that the amount of revision between the current and the historical (“final”)
seasonal factors was significantly larger with the regression models than with
Census II - X11, although the historical factors were obtained from the regression
models. In fact, the results showed that the current factors of the Census II - X11
were better predictors of the historical regression factors than the current regression
factors.
In 1972, Stephenson and Farr [48] developed a flexible regression model for
changing trend and seasonality, but still yielded results that were not, in overall,
superior to those of the Census II - X11 program.
Lovell [41] suggested a simple model of the following form:

12
X 12
X
yt D ˛1;k dk C ˛2;k Œdk C C It (3.10)
kD1 kD1
3.3 Basic Assumptions of Main Seasonal Adjustment Methods 69

subject to

12
X
˛2;k D 0
kD1

where C is a 12 month lagging average of the series (a trend level estimate) and the
dk ’s are monthly dummies. This model assumes a seasonal pattern that is stable plus
a slowly moving trend conditioned to seasonality for the whole span of the series.
The estimate of the trend C is the same for all the series. The seasonally adjusted
data are calculated by adding the mean of the original series to the residuals It .
This model which is extremely simple as compared to other models specifications
for regression methods did not give better results than those obtained by moving
averages procedures. As suggested by Early [17], the revisions in the seasonally
adjusted values were high and not necessarily monotonic for earlier years.
To overcome the limitation of using global deterministic representations for
the trend, cycle, and seasonality, regression methods were extended to incorporate
stochastic representations by means of local polynomials (spline functions) for suc-
cessive short segments of the series and introducing changing seasonal amplitudes.
A major breakthrough in this direction was made by Akaike [1] who introduced
prior constraints to the degree of smoothness of the various components and
solved the problem with the introduction of a Bayesian model. Another important
contribution is the regression method with locally deterministic models estimated
by the LOESS (locally weighted regression) smoother developed by Cleveland
et al. [8].
The simplest and often studied seasonal adjustment regression method assumes
that the generating process of seasonality can be represented by strictly periodic
functions of time of annual periodicity. The problem is to estimate s seasonal
coefficients (s being the seasonal periodicity, e.g., 12 for monthly series, 4 for
quarterly series) subject to the seasonal constraint that they sum to zero. This
regression seasonal model can be written as

yt D S t C I t ; t D 1; : : : ; n
X
s X
s
St D j djt subject to j D 0; (3.11)
jD1 jD1

where the dj ’s are dummy variables taking a value of unity in season j and a value of
zero otherwise; fIt g  WN.0; e2 / is a white noise process with zero mean, constant
variance, and non-autocorrelated. The seasonal parameters j ’s can be interpreted
as the expectation of Yt in each season, and represent the seasonal effects of the
series under question. Since the parameters are constant, the seasonal pattern repeats
exactly over the years, that is,

St D Sts : (3.12)
70 3 Seasonal Adjustment: Meaning, Purpose, and Methods

Equation (3.11), often included in econometric models, assumes that the effects
of the seasonal variations are deterministic because they can be predicted with
no error. However, for most economic and social time series, seasonality is not
deterministic but changes gradually in a stochastic manner. One representation of
stochastic seasonality is to assume in Eq. (3.11) when the j ’s are random variables
instead of constant. Hence, the relationship given in Eq. (3.12) becomes

St D Sts C !t ; 8t > s; (3.13)

where f!t g  WN.0; !2 / and E.!t It / D 0; 8t. The stochastic seasonal balance
P
constraint is here given by s1jD0 Stj D !t , with expected value equal to zero.
Equation (3.13) assumes seasonality to be generated by a nonstationary stochas-
tic process. In fact, it is a random walk process and can be made stationary by
applying seasonal differences .1Bs /, where B denotes the backward shift operator.
In reduced form, the corresponding regression model is a linear combination of
white noise processes of maximum lag s. Since .1Bs/ D .1B/.1CBC  CBs1 /,
model-based seasonal adjustment methods attribute only .1 C B C    C Bs1 / D
P s1 j
jD0 B to the seasonal component, leaving .1  B/ as part of a nonstationary trend.
Thus, the corresponding seasonal stochastic model can be written as

.1 C B C    C Bs1 /St D !t : (3.14)

Equation (3.14) represents a seasonality that evolves in an erratic manner and it was
included in seasonal adjustment methods based on structural model decomposition
(see, e.g., Harvey [25] and Kitagawa and Gersch [36]). This seasonal behavior is
seldom found in observed time series where seasonality evolves smoothly through
time. To represent this latter, Hillmer and Tiao [29] introduced a major modification

.1 C B C    C Bs1 /St D s .B/bt ; bt  WN.0; b2 / (3.15)

where the left-hand side of Eq. (3.15) follows an invertible moving average process
of maximum order s  1, denoted by MA.s  1/. Equation (3.15) is discussed
extensively in Bell and Hillmer [4], and can be generalized easily by replacing
the MA part with stationary invertible autoregressive moving average (ARMA)
processes of the Box and Jenkins type. Depending on the order of the ARMA
process, a large variety of evolving stochastic seasonal variations can be modeled.
Other seasonal models used in regression methods, where the components are
assumed to be local polynomials of time, are based on a smoothness criterion. For
example, given Yt D SP
P t C It ; t D 1; : : : ; n, restrictions are imposed on St , such that
2
n
tD1 .S t  S ts / and t Œ.1 C B C    C B
s1
/St 2 have to be small. The solution is
given by minimizing a weighted linear combination of both criteria as follows:

X
n X
n X
n
.yt  St /2 C ˛ .St  Sts /2 C ˇ Œ.1 C B C    C Bs1 /St 2 (3.16)
tD1 tD1 tD1
3.3 Basic Assumptions of Main Seasonal Adjustment Methods 71

where the ill-posed problem of choosing ˛ and ˇ parameters was solved by

Akaike [1] using a Bayesian model. Mixed seasonal models with deterministic and
stochastic effects are given by Pierce [45].
Although the performance of regression methods might be improved by more
closely tailoring the model specifications to the particular series, it is difficult to
derive regression estimates which, from the standpoint of revisions, are better than
moving averages estimates. The specifications made concerning the trend-cycle
and seasonality in the regression models apply to the whole span of the series.
Because of this, the procedure lacks of flexibility and the most recent estimated
values are always influenced by distant observations. Regression methods assume a
deterministic behavior of the components, in contrast to moving averages techniques
that assume a stochastic pattern. The latter, therefore, tend to follow better cyclical
movements and changing seasonality.

3.3.2 Stochastic Model-Based Methods

Stochastic model-based methods were mainly developed during the 1980s following
two different approaches. One, originally known as seasonal adjustment by SIGEX
(Signal Extraction), was developed by Burman [6] and as ARIMA model-based
seasonal adjustment by Hillmer and Tiao [29], the other is referred to as structural
model decomposition method (see, e.g., Harvey [25] and Kitagawa and Gersch
[36]). The main difference between these two approaches is that in the latter simple
ARIMA models are directly specified for each unobserved component, whereas
in the former an overall ARIMA model is obtained from observable data and, by
imposing certain restrictions, models for each component are derived. Since the
components are unknown, to obtain a unique decomposition Hillmer and Tiao [29]
proposed a canonical decomposition which has the properties of maximizing the
variance of the irregulars and minimizing the variance of the stochastic seasonal
models.
ARIMA models identification and estimation are sensitive to outliers or extreme
values and cannot deal with deterministic components, such as trading days and
moving holidays. Therefore, further developments were made by combining dummy
variables regression models with ARIMA models to deal with these cases. In
this regard, Gomez and Maravall [21] developed at the Bank of Spain a seasonal
adjustment software called TRAMO-SEATS which is currently applied mainly
by European statistical agencies. TRAMO stands for “Time Series Regression
with ARIMA noise, Missing observations and Outliers” and SEATS for “Signal
Extraction in ARIMA Time Series.” Firstly, TRAMO estimates via regression
dummy variables and direct ARIMA modeling, the deterministic components which
are after removed from the input data. In a second round, SEATS estimates the
seasonal and trend-cycle components from an ARIMA model fitted to the data
where the deterministic components are removed. SEATS uses the filters derived
from the ARIMA model that describes the behavior of the linearized time series.
72 3 Seasonal Adjustment: Meaning, Purpose, and Methods

It should be mentioned that Eurostat, with the collaboration of the Bank of Belgium,
the US Bureau of the Census, the Bank of Spain, and the European Central Bank,
has developed an interface of TRAMO-SEATS and X12ARIMA called Demetra+,
and more recently an update version called JDemetra+ . At the Bank of Spain
and Eurostat web site it is also possible to find a considerable number of papers
relevant to TRAMO-SEATS as well as in the European Statistical System (ESS)
Guidelines.
The structural model decomposition method starts directly with an observation
equation, also called measurement equation, that relates the observed time series
to the unobserved components. Simple ARIMA or stochastic trigonometric models
are assumed for each component. The structural model is cast in a state space form
and estimated with Kalman filtering and smoothing. Koopman et al. [37] developed
STAMP, that stands for Structural Time series Analyzer, Modeler and Predictor,
and includes several types of models for each component. STAMP is set up in
an easy-to-use form which enables the user to concentrate on model selection and
interpretation. STAMP 8 is an integrated part of the OxMetrics modular software
system for time series data analysis, forecasting, financial econometric modeling,
and the statistical analysis of cross-section and panel data. STAMP is not used
officially by statistical agencies, but mainly by econometricians and academics.

3.3.3 Linear Smoothing Methods

Linear smoothing filters or moving averages for seasonal adjustment were already
known in the early 1920s but seldom used [18, 35]. One of the main reasons for
this was the fact that the best known seasonal annihilator, the centered 12 months
moving average, was found to be a poor trend-cycle estimator. It cannot follow
closely peaks and troughs of short-term business cycles (of 5 years of periodicity or
less). Moreover, unless the erratic fluctuations are small, it has not enough terms to
smooth the data successfully. Another reason was that moving averages are sensitive
to outliers and, therefore, an “a priori” treatment of extreme values is required.
This sort of limitation lead King [35] and other researchers to use moving medians
instead of moving arithmetic means for the estimation of seasonality.
However, by the end of the 1920s, the development of new smoothing filters and
different techniques of application gave a big push to the use of this procedure for
the seasonal adjustment of economic time series. Joy and Thomas [32] described a
method based on moving averages and applied to one hundred series of the Federal
Reserve Board. According to the authors, the method was able to handle “long-time
changes in seasonal fluctuations” which appear to have occurred in a number of
important industries .
By the same time, Frederick Macaulay [42] developed another method also based
on linear smoothing filters. His book, The Smoothing of Time Series, published
in 1931, became a classic on the subject. Gradually, government agencies and
statistical bureaus started to apply smoothing procedures for the seasonal adjustment
3.3 Basic Assumptions of Main Seasonal Adjustment Methods 73

of their series. The procedure, however, was costly, time consuming, and mainly
subjective, because the adjustments were mostly made by hand.
The basic properties of linear smoothing filters or moving averages were
extensively discussed in the following three classical books: (1) Whittaker and
Robinson [51], (2) Macauley [42], and (3) Kendall and Stuart [34].
Only a summary of the basic principles of smoothing techniques will be
presented here, and given the complexity of the subject, we shall make it as simple
as possible.
The majority of the seasonal adjustment methods officially adopted by statistical
bureaus makes the assumption that although the signal of a time series is a smooth
function of time, it cannot be approximated well by simple mathematical functions
over the entire range. Therefore, they use the statistical technique of smoothing.
A smooth curve is one which does not change its slope in a sudden or erratic manner.
The most commonly used measure of smoothness is based on the smallness of the
sum of squares of the third differences of successive points on the curve [42].
Given a time series, yt ; t D 1; : : : ; n, for any t far removed from both ends, say
m < t < n  m, the seasonally adjusted value yat is obtained by application of a
symmetric moving average W.B/

X
m
yat D W.B/yt D wj ytj ; (3.17)
jDm

where the weights wj are symmetric, that is, wm;j D wm;j , and the length of the
average is 2m C 1.
For current and recent data (n  m < t  n), a symmetric linear filter cannot be
applied, and therefore truncated asymmetric filters are used. For example, for the
last available observation yn , the seasonally adjusted value is given by

X
m
yan D W0 .B/yn D w0;j ynj : (3.18)
jD0

The asymmetric filters are time-varying in the sense that different filters are
applied for the m first and last observations. The end estimates are revised as new
observations are added because of the new innovations and the differences between
the symmetric and asymmetric filters. The estimates obtained with symmetric filters
are often called final.
The development of electronic computers contributed to major improvements
in seasonal adjustment based on moving averages and facilitated their massive
application. In 1954, Julius Shiskin [47] of the US Bureau of Census developed
a software called Method I, based mainly on the works of Macauley [42], already
being used by the US Federal Reserve Board. Census Method I was followed by
Census Method II and eleven more experimental versions (X1 , X2,: : :, X11). The
best known and widely applied was the Census II - X11 variant developed by Shiskin
74 3 Seasonal Adjustment: Meaning, Purpose, and Methods

et al. [47], but produced poor seasonally adjusted data at the end of the series
which is of crucial importance to assess the direction of the short-term trend and the
identification of turning points in the economy. Estela Bee Dagum [12] developed
the X11ARIMA to correct for this serious limitation.
The X11ARIMA method consists of:
(i) modeling the original series with an ARIMA model of the Box and Jenkins
type,
(ii) extending the original series 1–3 years with forecasts from the ARIMA model
that fits and extrapolate well according to well-defined acceptance criteria, and
(iii) estimating each component with moving averages that are symmetric for
middle observations and asymmetric for both end years. The latter are obtained
via de convolution of Census II - X11 variant and the ARIMA model
extrapolations.
For flow series, deterministic components such as trading day variations and
moving holiday effects are estimated with dummy variable regression models, and
removed from the series, so that only the remainder is subject to steps (i)–(iii)
above. The X11ARIMA software was developed by Dagum in 1980 [13] and was
extended again in 1988 [14]. In 1998, Findley et al. [19] developed X12ARIMA
that offers a regARIMA option to estimate deterministic components such as trading
day variations and moving holiday effects, simultaneously with the ARIMA model
for extrapolation. It also includes new diagnostic tests and spectral techniques to
assess the goodness of the results. The X12ARIMA method is today one of the
most often applied by statistical agencies in the world. The US Bureau of Census
continued research on the development of seasonal adjustment methods and recently
produced a beta version called X13ARIMA-SEATS which enables the estimation of
the seasonal component either via linear filters as those available in X12ARIMA or
based on an ARIMA model decomposition.
All the moving average methods here discussed are nonlinear. Hence, the
seasonally adjusted total of aggregated series is not equal to the algebraic sum of
the seasonally adjusted series that enters into the aggregation. The main causes of
nonlinearity are:
(i) a multiplicative decomposition model for the unobserved components,
(ii) the identification and replacement of extreme values,
(iii) the ARIMA extrapolations, and
(iv) automatic selection of moving average length for the estimation of the trend-
cycle and seasonality.
The properties of the combined linear filters applied to estimate the various
components were originally calculated by Young [53] for a standard option. Later,
Dagum et al. [15] calculated and analyzed all possible filter combinations of Census
II - X11 and X11ARIMA. Cleveland and Tiao [10] and Burridge and Wallis [7]
found ARIMA models that approximated well some of the linear filters used for the
trend-cycle and seasonal component of Census II - X11.
3.3 Basic Assumptions of Main Seasonal Adjustment Methods 75

The set of weights of a smoothing linear filter can be obtained following two
different techniques: (a) by fitting a polynomial and (b) by summation formulae.
The process of fitting a polynomial to generate the set of weights of a moving
average requires that the span or length of the average as well as the degree of the
polynomial be chosen in advance. For a given span, say 2m C 1 and a polynomial
of degree p not greater than 2m, the coefficients are estimated by the method of
least squares, and then the polynomial is used to determine the smoothed value in
the middle of the span. The weights are function only of the span of the average
2m C 1 and the degree of the polynomial, p, and do not depend on the values of
the observations. The smoothed value is a weighted average of the observations, the
weights being independent on which part of the series is taken. For a given p, the
variance of the smoothed series decreases with increasing 2m C 1 and, for a given
2m C 1, the variance goes up with increasing p (see, e.g., [2]).
The set of weights for linear smoothing filters based on summation formulae
was mainly developed by actuaries. The basic principle for the summation formulae
is the combination of operations of differencing and summation in such a manner
that when differences above a certain order is neglected, they will reproduce the
functions operated on. The merit of this procedure is that the smoothed values
thus obtained are functions of a large number of observed values whose errors to
a considerable extent cancel out.
The smoothing linear filters developed by actuaries using summation formulae,
have the good properties that when fitted to second or third degree parabolas, will
fall exactly on those parabolas. If fitted to stocastic, non mathematical data, it will
give smoother results than can be obtained from the weights which give the middle
point of a second degree parabola fitted by least squares method. Recognition of the
fact that smoothness of the resulting filtering depends directly on the smoothness of
the weight diagram, led Robert Henderson [26] to develop a formula which makes
the sum of squares of the third differences of the smoothed series a minimum for
any number of terms. To fulfill this requirement is equivalent to minimize the sum
of the squares of the third differences of the set of weights of the filter.
There is a relationship between smoothing by the mid ordinate of third degree
parabolas fitted by least squares and smoothing by the mid ordinate of third degree
parabolas from summation formulae. In the least squares, the assumption made
is that all deviations between observed and fitted values are equally weighted,
and thus the sum of squares of the deviations is made a minimum. In the
summation formulae, the deviations between observed and fitted values are not
equally weighted and if different weights are applied, the sum of squares of the
deviations is made a minimum.
The Henderson’s trend-cycle moving averages give the same results as if
weighted least squares had been applied where the weights assigned to the devi-
ations are those that give the smoothest possible diagram. The latter in the sense
that the sum of squares of the third differences is made a minimum. In smoothing
real data there is always a compromise between: (1) how good the fit should be
(in the sense of minimizing the sum of squares of the deviations between observed
and fitted values); and (2) how smooth the fitted curve should be. The lack of fit is
76 3 Seasonal Adjustment: Meaning, Purpose, and Methods

measured by the sum of squares of the deviations between observed and smoothed
values and the lack of smoothness by the sum of squares of the third differences of
the smoothed curve.
This problem was suggested by Henderson and Sheppard [27], and first analyzed
by Whittaker and Robinson [51], who thus put the theory of smoothing within the
framework of the theory of probability. The degree of fitness was weighted by a
constant k to be determined and the problem was to minimize k times the sum of
squares of the deviations between the observed and fitted curve plus the sum of
squares of the third differences of the fitted (smoothed) curve. That is to minimize:

X
n X
n
k .yt  yO t /2 C ..1  B/3 yO t /2 (3.19)
tD1 tD3

The smoothed curve obtained with the Whittaker procedure is such that each of
its values is equal to the corresponding unadjusted values plus 1=k times a sixth
difference of the smoothed curve. The weights were calculated for the whole range
of the data and not in a moving manner. The set of weights eliminates stable seasonal
fluctuations only if k is very small. For k D 0:0009, more than 95 % of the seasonal
effects corresponding to a 12 month sine curve are eliminated, but cycles of less
than 36 month periodicity are very poorly fitted as shown by Macauley [42]. This
problem can be expressed in a more adequate way by minimizing a loss function
given by a convex combination between the lack of smoothness and the lack of
fit. The principle should be applied in a moving manner in order to make it more
flexible. Hence, we want to minimize

X
n X
n
k .yt  yO t /2 C .1  k/ ..1  B/3 yt /2 ; 0  k  1: (3.20)
tD1 tD3

The larger is k, the more importance is given to closeness of fit as compared to

smoothness; if k D 1, the problem reduces to fitting only and, if k D 0, the problem
reduces to smoothing only.

References

1. Akaike, H. (1980). Seasonal adjustment by a Bayesian modelling. Journal of Time Series

Analysis, 1, 1–13.
2. Anderson, T. W. (1971). The statistical analysis of time series. New York: Wiley.
3. Baron, R.V. (1973). Analysis of seasonality and trends. In Statistical series, W.I. Methodology.
Causes and effects of seasonality. Technical Publication, No. 39. Israel: Central Bureau of
Statistics.
4. Bell, W. H., & Hillmer, S. C. (1984). Issues involved with the seasonal adjustment of economic
time series. Journal of Business and Economic Statistics, 2, 291–320.
References 77

5. Brewer, K. R.W., Hagan, P. J., & Perazzelli, P. (1975). Seasonal adjustment using Box-Jenkins
models. In Proceedings of International Institute of Statistics, Warsaw. 40th Session (pp. 133–
139). Contributed Papers.
6. Burman, J. P. (1980). Seasonal adjustment by signal extraction. Journal of the Royal Statistical
Society Series A, 143, 321–337.
7. Burridge, P., & Wallis, K. F. (1984). Unobserved components models for seasonal adjustment
filters. Journal of Business and Economic Statistics, 2(4), 350–359.
8. Cleveland, R. B., Cleveland, W. S., McRae, J. E., & Terpenning, I. (1990). STL: A seasonal
trend decomposition procedure based on LOESS. Journal of Official Statistics, 6(1), 3–33.
9. Cleveland, W. P. (1972). Analysis and Forecasting of Seasonal Time Series (215 pp.).
Unpublished doctoral thesis, Department of Statistics, University of Wisconsin.
10. Cleveland, W. P., & Tiao, G. C. (1976). Decomposition of seasonal time series: A model for
the census X-11 program. Journal of the American Statistical Association, 71, 581–587.
11. Cowden, D. J. (1942). Moving seasonal indexes. Journal of the American Statistical Associa-
tion, 37, 523–524.
12. Dagum, E. B. (1978). Modelling, forecasting and seasonally adjusting economic time series
with the X11ARIMA method. The Statistician, 27(3), 203–216.
13. Dagum, E. B. (1980). The X11ARIMA seasonal adjustment method. Ottawa, ON, Canada: Time
Series Research and Analysis Center, Statistics Canada.
14. Dagum, E. B. (1988). The X11ARIMA/88 seasonal adjustment method-foundations and user’s
manual. Ottawa, ON, Canada: Time Series Research and Analysis Centre, Statistics Canada.
15. Dagum, E. B., Chhab, N., & Chiu, K. (1990). Derivation and properties of the X11ARIMA
and census X11 linear filters. Journal of Official Statistics, 12(4), 329–347.
16. Dawson, D. A., Denton, F. T., Feaver, C., & Robb, A. (1975). Seasonal patterns in the
Canadian Labour force (108 pp.). Economic Council of Canada, Discussion Paper. No. 38.
17. Early, J. (1976). Comment. Brooking Papers on Economic Activity, 1, pp. 239–243.
18. Falkner, H. D. (1924). The measurement of seasonal variation. Journal of the American
Statistical Association, 19, 167–179.
19. Findley, D. F., Monsell, B. C., Bell, W. R., Otto, M. C., & Chen, B. C. (1998). New capabilities
and methods of the X12ARIMA seasonal adjustment program. Journal of Business and
Economic Statistics, 16(2), 127–152.
20. Fisher, A. (1937). A brief note on seasonal variation. The Journal of Accountancy, 64, 174–199.
21. Gomez, V., & Maravall, A. (1996). Program TRAMO and SEATS. Madrid: Bank of Spain.
22. Hannan, E. J. (1960). The estimation of seasonal variations. Australian Journal of Statistics, 2,
1–15.
23. Hannan, E. J. (1964). The estimation of changing seasonal pattern. Journal of the American
Statistical Association, 59, 1063–1077.
24. Hannan, E. J., Terrell, R. D., & Tuckwell, N. E. (1970). The seasonal adjustment of economic
time series. International Economic Review, 11, 24–52.
25. Harvey, A. C. (1981). Time series models. Oxford, UK: Phillip Allan Publishers Limited.
26. Henderson, R. (1916). Note on graduation by adjusted average. Transaction of Actuarial
Society of America, 17, 43–48.
27. Henderson, R., & Sheppard, H. N. (1919). Graduation of mortality and other tables. Actuarial
Studies, No. 4. New York: Actuarial Society of America.
28. Henshaw, R. C. (1966). Application of the general linear model to seasonal adjustment of
economic time series. Econometrica, 34, 381–395.
29. Hillmer, S. C., & Tiao, G. C. (1982). An ARIMA model-based approach to seasonal
adjustment. Journal of the American Statistical Association, 77, 63–70.
30. Jones, H. L. (1943). Fitting polynomial trends to seasonal data by the method of least squares.
Journal of the American Statistical Association, 38, 453–465.
31. Jorgenson, D. N. (1964). Minimum variance, linear unbiased, seasonal adjustment of economic
time series. Journal of the American Statistical Association, 59, 402–421.
32. Joy, A., & Thomas, N. (1928). The use of moving averages in the measurement of seasonal
variations. Journal of the American Statistical Association, 23, 242–252.
78 3 Seasonal Adjustment: Meaning, Purpose, and Methods

33. Judek, S. (1975). Canada’s seasonal variations in employment and unemployment, 1953–73
(118 pp.). Research Paper No. 75. Canada: University of Ottawa.
34. Kendall, M. G., & Stuart, A. (1966). The advanced theory of statistics (Vol. 3, 552 pp.). New
York: Hafner Publishing Co.
35. King, W. F. (1924). An improved method for measuring the seasonal factor. Journal of the
American Statistical Association, 19, 301–313.
36. Kitagawa, G., & Gersch, W. (1984). A smoothness priors state-space modelling of time series
with trend and seasonality. Journal of the American Statistical Association, 79, 378–389.
37. Koopman, S. J., Harvey, A. C., Doornik, J. A., & Shephard, N. (1998). Structural time series
analysis STAMP (5.0). London: Thomson Business Press.
38. Kuznets, S. (1933). Seasonal variations in industry and trade (455 pp.). New York: National
Bureau of Economic Research.
39. Ladd, G. W. (1964). Regression analysis of seasonal data. Journal of the American Statistical
Association, 59, 402–421.
40. Lovell, M. C. (1963). Seasonal adjustment of economic time series and multiple regression
analysis. Journal of the American Statistical Association, 58, 993–1010.
41. Lovell, M. C. (1976). Least squares seasonally adjusted unemployment data. Brooking Papers
on Economic Activity, l, pp. 225–237.
42. Macaulay, F. R. (1931). The smoothing of time series. New York: National Bureau of Economic
Research.
43. Mendershausen, H. (1939). Eliminating changing seasonals by multiple regression analysis.
The Review of Economic Studies, 21, 71–77.
44. Persons, W. M. (1919). Indices of business conditions. Review of Economic Statistics, 1, 5–107.
45. Pierce, D. (1978). Seasonal adjustment when both deterministic and stochastic seasonality are
present. In A. Zellner (Ed.), Proceedings of the NBER/Bureau of the Census Conference on
Seasonal Analysis of Economic Time Series, Washington, DC (pp. 242–269).
46. Shiskin, J. (1973). Measuring current economic fluctuations. In Annals of economic and social
measurements (Vol. 2, pp. 1–16.). Cambridge: National Bureau of Economic Research.
47. Shiskin, J., Young, A. H., & Musgrave, J. C. (1967). The X-11 variant of the Census Method II
seasonal adjustment program. Technical Paper 15 (revised). Washington, DC: US Department
of Commerce, Bureau of the Census.
48. Stephenson, J. A., & Farr, H. T. (1972). Seasonal adjustment of economic data by application of
the general linear statistical model. Journal of the American Statistical Association, 67, 37–45.
49. Wald, A. (1936). Berechnung und Ausschaltung von Saisonschwankungen. Beitrage Zur
Konjunkturforschung. 9, Vienna: Verlag Julius Springer.
50. Wallis, K. E. (1978). Seasonal adjustment and multiplicative time series analysis. In A. Zellner
(Ed.), Proceedings of the NBER/Bureau of the Census Conference on Seasonal Analysis of
Economic Time Series,Washington, DC (pp. 347–357).
51. Whittaker, E., & Robinson, G. (1924). The calculus of observation: A treatise on numerical
mathematics (397 pp.). London: Blackie and Sons Ltd. (also published by Dover Publications,
New York, 1967).
52. Wold, H. O. (1938). A study in the analysis of stationary time series. Uppsala, Sweden:
Almquist and Wiksell (Second Edition 1954).
53. Young, A. H. (1968). A linear approximation to the Census and BLS seasonal adjustment
methods. Journal of the American Statistical Association, 63, 445–457.
Chapter 4
Linear Filters Seasonal Adjustment Methods:
Census Method II and Its Variants

Abstract The best known and most often applied seasonal adjustment methods are
based on smoothing linear filters or moving averages applied sequentially by adding
(and subtracting) one observation at a time. This chapter discusses with details the
basic properties of the symmetric and asymmetric filters of the Census Method II-
X11 method which belong to this class. It also discusses the basic assumptions of its
two more recent variants, X11ARIMA and X12ARIMA. The latter consists of two
linked parts: the regARIMA model for estimation of the deterministic components
(mainly calendar effects), and the decomposition part of the linearized series for the
stochastic components (trend-cycle, seasonality, and irregulars) performed using the
X11 filters combined with those of the ARIMA model extrapolations. An illustrative
example of the seasonal adjustment with the X12ARIMA software default option is
shown with the US New Orders for Durable Goods series. The illustrative example
concentrates on the most important tables of this software that enable to assess the
quality of the seasonal adjustment.

4.1 Introduction

The best known and most often applied seasonal adjustment methods are based on
smoothing linear filters or moving averages applied sequentially by adding (and
subtracting) one observation at a time. These methods assume that the time series
components change through time in a stochastic manner. Given a time series, yt ; t D
1; : : : ; n, for any t far removed from both ends, say, m < t < n  m, the seasonally
adjusted value yat is obtained by application of a symmetric moving average W.B/
as follows:

X
m X
m
yat D W.B/yt D wj Bj yt D wj ytj ; (4.1)
jDm jDm

where the weights wj are symmetric, that is, wj D wj , and the length of the average
is 2m C 1.
For current and recent data (n  m < t  n) a symmetric linear filter cannot be
applied, and therefore truncated asymmetric filters are used. For example, for the

© Springer International Publishing Switzerland 2016 79

E. Bee Dagum, S. Bianconcini, Seasonal Adjustment Methods and Real Time
Trend-Cycle Estimation, Statistics for Social and Behavioral Sciences,
DOI 10.1007/978-3-319-31822-6_4
80 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

last available observation yn , the seasonally adjusted value is given by

X
m
yan D W0 .B/yn D w0;j ynj : (4.2)
jD0

The asymmetric filters are time-varying in the sense that different filters are
applied to the m first and last observations. The end estimates are revised as
new observations are added because of: (1) the new innovations, and (2) the
differences between the symmetric and asymmetric filters. The estimates obtained
with symmetric filters are often called “final.”
The development of electronic computers contributed to major improvements
in seasonal adjustment based on moving averages and facilitated their massive
application.
In 1954, Julius Shiskin of the US Bureau of Census developed software called
Method I, based mainly on the works of Macauley [22] already being used by the
US Federal Reserve Board. Census Method I was followed by Census Method II and
eleven more experimental versions (X1 , X2, : : :, X11). The best known and widely
applied was the Census Method II-X11 variant developed by Shiskin et al. [26], but
produced poor seasonally adjusted data at the end of the series which is of crucial
importance to assess the direction of the short-term trend and the identification of
turning points in the economy. Dagum [6] developed the X11ARIMA mainly to
correct for this serious limitation and, later Findley et al. [9] developed X12ARIMA
that offers a regARIMA option to estimate deterministic components, such as
trading day variations, moving holiday effects and outliers, simultaneously with the
ARIMA model for extrapolation. It also includes new diagnostic tests and spectral
techniques to assess the goodness of the results. The X12ARIMA method is today
the one most often applied by statistical agencies in the world. The US Bureau of
Census continued research on the development of seasonal adjustment methods,
and recently produced a beta version called X13ARIMA-SEATS which enables the
estimation of the seasonal component either via linear filters as those available in
X12ARIMA or based on an ARIMA model decomposition.
All the seasonal adjustment methods based on moving averages discussed in this
book have nonlinear elements. Hence, the seasonally adjusted total of an aggregated
series is not equal to the algebraic sum of the seasonally adjusted series that enter
into the aggregation. The main causes of nonlinearity are generated by:
1. a multiplicative decomposition model for the unobserved components,
2. the identification and replacement of extreme values,
3. the ARIMA extrapolations, and
4. automatic selection of the moving average length for the estimation of the trend-
cycle and seasonality.
The properties of the combined linear filters applied to estimate the various
components were originally calculated by Young [28] for the standard option of
Census II-X11 variant. Later, Dagum et al. [8] calculated and analyzed all possible
filter combinations of Census II-X11 and X11ARIMA.
4.1 Introduction 81

4.1.1 Main Steps to Produce a Seasonally Adjusted Series

Given the fact that X11ARIMA and X12ARIMA are methods that are based on
the Bureau of the Census Method II-X11 [26], we shall first briefly discuss the
Census II-X11 variant and refer the reader to Ladiray and Quenneville [18] for a
very detailed description.
The Census II-X11 method assumes that the main components of a time series
follow a multiplicative or an additive model, that is,
(multiplicative model) yt D TCt  St  It
(additive model) yt D TCt C St C It
where yt stands for the original series, TCt for the trend-cycle, St for the seasonal,
and It for the irregular.
There are no mixed models in this program, such as yt D TCt  St C It or other
possible combinations. The estimation of the components is made with different
kinds of smoothing linear filters.
The main steps followed to produce monthly seasonally adjusted data are:
1. Compute the ratios (differences) between the original series and a centered 12
months moving average (2  12 m.a.), that is, a 2 months average of 12 months
(average) as a first estimate of the seasonal and irregular components.
2. Apply a weighted 5-term moving average (3  3 m.a.) to the seasonal–irregular
ratios (differences) for each month separately, to obtain an estimate of the
seasonal factors (effects). Compute a 2  12 m.a. of the preliminary factors for
the entire series. To obtain the six missing values at either end of this average,
repeat the first (last) available value six times.
3. Adjust these seasonal factors (differences) to sum to 12 (0.000 approximately)
over any 12-month period by dividing (subtracting) the centered 12 months
average into (from) the factors (differences).
4. Divide (subtract) the seasonal factors (differences) into (from) the seasonal–
irregular ratios (differences) to obtain an estimate of the irregular component.
5. Compute a moving 5-year standard deviation () of the estimates of the
irregular component and test the irregulars in the central year of the 5-year
period against 2:5. Remove values beyond 2:5 as extreme and recompute
the moving 5-year .
6. Assign a zero weight to irregulars beyond 2:5 and a weight of one (full weight)
to irregulars within 1:5. Assign a linearly graduated weight—between 0 and
1 to irregulars between 2:5 and 1:5.
7. For each separate month apply a weighted 5-term moving average (3  3 m.a.)
to the seasonal–irregular ratios (differences) with extreme values replaced by
the corresponding values in step 6, to estimate preliminary seasonal factors
(effects).
8. Repeat step 3.
9. To obtain a preliminary seasonally adjusted series divide (subtract) the series
obtained at step 8 into (from) the original series.
82 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

10. Apply a 13-term Henderson moving average to the seasonally adjusted series
and divide (subtract) the resulting trend-cycle into (from) the original series to
obtain a second estimate of the seasonal–irregular ratios (differences).
11. Apply a weighted 7-term moving average (3  5 m.a.) to the seasonal–irregular
ratios (differences) for each month separately to obtain a second estimate of the
seasonal component.
12. Repeat step 3.
13. Divide (subtract) the series obtained at step 11 into (from) the original series to
obtain a seasonally adjusted series.
The filters are applied sequentially and the thirteen steps are repeated twice.
The Census II-X11 variant produces seasonal factor forecasts for each month
which are given by

1
Sj;tC1 D Sj;t C .Sj;t  Sj;t1 /; j D 1; : : : ; 12;
2
where j is the month and t the current year. The use of seasonal factor forecasts was
very popular in statistical Bureaus till the end of 1970s. In 1975, Dagum proposed
the use of concurrent seasonal factors obtained from data that included the most
recent value and this was adopted by Statistics Canada. Gradually, other statistical
agencies applied concurrent seasonal factors and this is now the standard practice.
The final trend-cycle is obtained by 9-, 13-, or 23-term Henderson moving
averages applied to the final monthly seasonally adjusted series (5-term and 7-term
Henderson filters are used for quarterly data). The selection of the appropriate filter
is made on the basis of a preliminary estimate of the I=C ratio (the ratio of the
average absolute month-to-month change in the irregular to that in the trend-cycle).
A 9-term is applied to less irregular series and a 23-term to highly irregular series.
Using a linear approximation of the Census Method II, Young [28] arrived at the
conclusion that a 145-term moving average was needed to estimate one seasonal
factor with central weights if the trend-cycle component is adjusted with a 13-
term Henderson moving average. The first and last 72 seasonal factors (6 years)
are estimated using sets of asymmetrical end weights. It is important to point out,
however, that the weights given to the most distant observations are very small and,
therefore, the moving average can be very well approximated by taking one half
of the total number of terms plus one. So, if a 145-term moving average is used
to estimate the seasonal factor of the central observation, a good approximation
is obtained with only 73 terms, i.e., 6 years of observations. This means that the
seasonal factor estimates from unadjusted series whose observations end at least
3 years later can be considered “final”, in the sense that they will not change
significantly when new observations are added to the raw data.
4.2 Basic Properties of the Two-Sided Linear Smoothing Filters of Census. . . 83

4.2 Basic Properties of the Two-Sided Linear Smoothing

Filters of Census Method II-X11 Variant

The linear smoothing filters applied by Census Method II-X11 variant to produce
seasonally adjusted data can be classified according to the distribution of their set
of weights in symmetric (two-sided) and asymmetric (one-sided). The symmetric
moving averages are used to estimate the component values that fall in the middle
of the span of the average, say, 2m C 1, and the asymmetric moving averages, to the
m first and last observations.
The sum of the weights of both kinds of filters is one and thus, the mean of
the original series is unchanged in the filtering process. The sum of the weights of
a filter determines the ratio of the mean of the smoothed series to the mean of the
unadjusted series assuming that these means are computed over periods long enough
to ensure stable results. It is very important in filter design that the filter does not
displace in time the components of the output relative to those of the input. In other
words, the filter must not introduce phase shifts.
Symmetric moving averages introduce no time displacement for some of the
components of the original series and a displacement of 180ı for others. A phase
shift of 180ı is interpreted as a reverse in polarity which means that maxima are
turned into minima and vice versa. In other words, peaks (troughs) in the input
are changed into troughs (peaks) in the output. For practical purposes, however,
symmetric moving averages act as though the time displacement is null. This is so
because the sinusoids that will have a phase shift of 180ı in the filtering process are
cycles of short periodicity (annual or less) and moving averages tend to suppress or
significantly reduce their presence in the output. In spectral analysis, the phase is a
dimensionless parameter that measures the displacement of the sinusoid relative to
the time origin. Because of the periodic repetition of the sinusoid, the phase can be
restricted to ˙180ı . The phase is a function of the frequency of the sinusoid, being
the frequency equal to the reciprocal of the length of time or period required for one
complete oscillation.
The next two sections discuss the basic properties of the two-sided and one-
sided filters applied by the Census II-X11 method. The study is based on a spectral
analysis of the corresponding filters.

4.2.1 The Centered 12 Months Moving Average

The centered 12 months moving average used for a preliminary estimate of the
trend-cycle (step 1) reproduces exactly a linear trend and annihilates a stable
seasonal in an additive model. However, if the relationship among the components
is multiplicative, this filter will reproduce only a constant trend and eliminate a
stable seasonality. Unless the cycles are long (5 years or more) and the irregulars
quite small, this filter will not follow closely the peaks and troughs and will not
84 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

smooth the data successfully. If a sine curve of 3-year periodicity and amplitude
100 is input to this filter, the output is a sine curve of equal periodicity but with
amplitude reduced to 82.50. For a 5-year cycle the amplitude is reduced to 93.48
and for a 10-year cycle is 98.33. In
 general,
 any simple moving average of m equal
2
weights fitted to a parabola falls m121 c inside the parabola, see Macauley [22].
Hence, this filter fitted to a second degree parabola yt D a C bt C ct2 will give a
smoothed curve that falls 12.17 units inside the parabola, i.e., yO t D yt C 12:17c. The
bias introduced by this filter will not affect significantly the preliminary estimate
of the seasonal component if the trend is nearly linear or if the direction of the
curvature changes frequently. In the latter case, the bias will be sometimes positive
and sometimes negative, so it may average out in applying the seasonal filters to the
Seasonal–Irregular (SI) ratios for each month separately over several years.

4.2.2 The 9-, 13-, and 23-Term Henderson Moving Averages

The Henderson [14] moving averages are applied to obtain an improved estimate
of the trend-cycle (step 11). These linear smoothing filters were developed by
summation formulae. Their weights minimize the sum of squares of the third
differences of the smoothed curve. The latter is equivalent to say that these filters
reproduce the middle value of a function that is a third degree parabola within the
span of the filter. The Henderson’s moving averages give the same results as would
be obtained by smoothing the middle value of a third degree parabola fitted by
weighted least squares, where the weights given to the deviations are as smooth
as possible.
The fact that the trend-cycle is assumed to follow a cubic over an interval of short
duration (between 1 and 2 years approximately) makes these filters very adequate
for economic time series.
None of the Henderson’s filters used by the Census II-X11 program eliminates
the seasonal component, but since they are applied to already seasonally adjusted
data, this limitation becomes irrelevant. On the other hand, they are extremely good
to pass sines of any periodicity longer than a year. Thus, the 13 months Henderson,
which is the most frequently applied, will not reduce the amplitude of sines of
20 months periodicity or more that stand for trend-cycle variations. Moreover, it
eliminates almost all the irregular variations that can be represented by sines of very
short periodicity, 6 months or less.

4.2.3 The Weighted 5-Term (3  3) and 7-Term (3  5) Moving

Averages

The weighted 5-term moving average is a three-term moving average of a three-

term moving average (3  3 m.a.). Similarly, the weighted 7-term moving average is
4.3 Basic Properties of the One-Sided Linear Smoothing Filters of Census. . . 85

a three-term moving average of a five-term moving average (3  5 m.a.). These two

filters are applied to the seasonal–irregular ratios (or differences) for each month
separately over several years. Their weights are all positive and, consequently, they
reproduce exactly the middle value of a straight line within their spans. This property
enables the Census II-X11 method to estimate a linearly moving seasonality within 5
and 7 years spans. Therefore, these filters can approximate quite adequately gradual
seasonal changes that follow nonlinear patterns over the whole range of the series
(more than 7 years).
The weighted 5-term moving average (3  3 m.a.) is a very flexible filter that
allows for fairly rapid changes in direction. But, since the span of the filter is
short, the irregulars must be small for the SI to be smoothed successfully. In effect,
assuming that sinusoids of 5 years of periodicity or less represent the irregular
variations, given the fact that by definition seasonality is expected to move slowly,
this filter still passes about 20 % of the amplitude of a 5-year sine wave. Sinusoids
of 3 years periodicity or less are completely eliminated.
The weighted 7-term moving average (3  5 m.a.) is less flexible and it is applied
for the final estimate of the seasonal factors. Sinusoids of periodicity shorter than 5
years are completely eliminated. For series the seasonal factors of which are nearly
stable, this program also provides other optional sets of weights which are applied
to longer spans.

4.3 Basic Properties of the One-Sided Linear Smoothing

Filters of Census Method II-X11 Variant

It is inherent to any moving average procedure that the first and last m points of an
original series cannot be smoothed with the same set of symmetric 2m C 1 weights
applied to the middle values. The Census II-X11 method uses one-sided filters to
smooth these end points. Moving averages with asymmetric weights are bounded to
introduce phase shifts for all the components of the original series. This is a very
undesirable characteristic since it may cause, for example, a business cycle in the
smoothed series to turn down before or after it actually does in the unadjusted data.
The asymmetric weights applied in conjunction with the Henderson symmetric
filters, developed by Musgrave [23], no longer estimate a cubic within their
span. In fact, the weight associated with the 13-term Henderson filter, applied to
estimate the last data point, amplifies sinusoids of 3 to 1 year periodicities up to
10 %, passes between 37 and 17 % of the amplitude of sine waves representing
irregular variations, and it has a linearly decreasing phase shift for sinusoids of
long periodicity. The latter means that the smoothed series lags the original series
a number of units of time equal to the slope of the line. For an 18-month cycle, the
phase shift is about 4 %, i.e., a lag of more than half of a month.
The set of end weights used in conjunction with the Henderson filters can
estimate well only a linear trend-cycle within the length of the filters [23].
86 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

Concerning the other two seasonal one-sided filters, they also introduce phase
shifts and can only estimate stable or nearly stable seasonality. There are other
reasons for the poor quality of both the current and seasonal factor forecasts
generated by the Census II-X11 method. As mentioned above, the program repeats
the first and last six preliminary seasonal factors obtained with the 33 m.a. In other
words, for a series ending, say, in December 1977, the seasonal factors from July to
December of 1977 are equal to the corresponding seasonal factors of the previous
year. This operation is repeated any time that the centered 12-term moving average is
applied to estimate the trend-cycle. In the past, when seasonal factor forecasts were
used to produce concurrent seasonal adjustment, the last 6 months were subject to
larger errors than the corresponding factors for the first 6 months because they were
based on observations lagging 2 years with respect to the forecasting year. It was
then preferable to forecast only 6 months instead of 12. By concurrent seasonal
adjustment is meant the adjustment of the original series made with seasonal factors
obtained from a series that includes the last observed point.

4.4 The X11ARIMA Method

4.4.1 General Outline and Basic Assumptions

A detailed description of the X11ARIMA method of seasonal adjustment is given

in [6, 7]. The X11ARIMA is a modified version of the Census Method II-X11
variant that was developed mainly to produce a better current seasonal adjustment
of series with seasonality that changes rapidly in a stochastic manner. The latter
characteristic is often found in main socioeconomic indicators, e.g., retail trade,
imports and exports, unemployment, and so on.
The X11ARIMA method consists of:
(i) modeling the original series with an ARIMA model of the Box and Jenkins
type,
(ii) extending the original series 1–3 years with forecasts from the ARIMA model
that fits and extrapolates well according to well-defined acceptance criteria,
and
(iii) estimating each component with moving averages that are symmetric for
middle observations and asymmetric for both end years. The latter are obtained
via the convolution of Census II-X11 variant and the ARIMA model extrapo-
lations.
For series that result from the accumulation of daily activities, called flow series,
deterministic components such as trading day variations and moving holiday effects
are estimated with dummy variable regression models and removed from the series,
so that only the remainder is subject to steps (i)–(iii) above. The X11ARIMA was
extended again by Dagum [7].
4.4 The X11ARIMA Method 87

An alternative way to cope with the problem of estimating rapidly changing

stochastic seasonality is to introduce more adequate one-sided filters but this
alternative presents the old problem of end weights to be chosen. In effect, it is well
known the trade-off between the phase shift introduced by one-sided filters, and the
power to pass low frequencies without modification. Therefore, it was preferable
to extend the original series with forecasts. The next step in the development of
the X11ARIMA was to decide what kind of forecasting model should be used
for the original series. The selection was made in accordance with the following
requirements:
1. the forecasting model must belong to the “simplest” class in terms of its
description of the real world. No explanatory variables must be involved,
the series should be described simply by its past values and lagged random
disturbances,
2. it should be efficient for short-term forecasting, one or two years at most,
3. it must generate forecasts that are robust to small changes in parameter values,
4. it must be parsimonious in the number of parameters,
5. it must generate optimum forecasts, in the sense of minimum mean square error
forecasts, and
6. it must generate forecasts that follow well the intra-year movement although they
could miss the level.
All these requirements were fulfilled by the ARIMA models except for point
5, that is satisfied for only the one-step ahead forecasts. ARIMA models of the
Box and Jenkins type were chosen which were previously very effective for
forecasting a large number of series [24, 25]. In the Box and Jenkins notation,
the general multiplicative ARIMA model, for a series with seasonality of period
s, is defined to be of order .p; d; q/.P; D; Q/s, where d and D are the orders of the
regular and seasonal difference operator, respectively, p and P are the orders of
the autoregressive polynomials in the regular and seasonal lag operator B and Bs ,
respectively, whereas q and Q are the orders of the moving average polynomials in
the regular and seasonal lag operator B and Bs , respectively.
The main advantages of X11ARIMA over the Census II-X11 variant are:
1. The availability of a statistical model that describes the stochastic behavior of the
original series. The expected value and the variance of the original series can be
calculated and thus, confidence intervals can be constructed for the observations.
This is very important to detect outliers at the end of the series.
2. The one-step forecast from these models is a minimum mean square error
forecast and can be used as a benchmark for preliminary figures.
3. The total error in the seasonal factor forecasts, and in the concurrent seasonal
factors, is significantly reduced for all months. Generally, a reduction of about
30 % in the bias and of 20 % in the absolute values of the total error has been
found for Canadian and American series.
4. If concurrent seasonal factors are applied to obtain current seasonally adjusted
data, there is no need to revise the series more than twice. For many series, just
88 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

one revision will give seasonal factors that are “final” in a statistical sense. There
are several reasons for the significant reduction of the error in the seasonal factor
forecasts and concurrent seasonal factors. The X11ARIMA produces seasonal
factor forecasts from the combination of two filters:
a. the filters of the autoregressive integrated moving averages (ARIMA) models
used to forecast the raw data; and
b. the filters that Census Method II-X11 variant applies for concurrent seasonal
adjustment.
In this manner, the seasonal factor forecasts are obtained from the forecasted
raw values with filters that even though are still one-sided, they are closer to the
two-sided filters as compared to the forecasting function of the Census II-X11
variant.
5. The trend-cycle estimate for the last observation is made with the symmetric
weights of the Henderson moving average which can reproduce a cubic within
their time span. This is relevant for years with turning points because the
Census II-X11 applies the asymmetric weights of the Henderson filters which
can adequately estimate only a linear trend.
6. Finally, by adding one more year of preliminary data (with no extremes since
they are forecasts) a better estimate of the variance of the irregulars is obtained.
The latter allows a significant improvement in the identification and replacement
of outliers which, as we already know, can severely distort the estimates obtained
with linear moving filters.
For concurrent seasonal factors, the same observations are valid except that the
seasonal filters are closer to the two-sided filters than those corresponding to the
seasonal factor forecasts. For this reason, the number of revisions in the seasonal
factor estimates is also significantly reduced.

4.4.2 The Forecasting Filters of ARIMA Models and Their

Properties

We will here analyze the forecasting filters of the ARIMA models. The ARIMA
models applied to seasonal series belong to the general multiplicative Box and
Jenkins type, that is,

p .B/˚P .Bs /.1  B/d .1  Bs /D yt D q .B/Q .Bs /at ; (4.3)

where s denotes the periodicity of the seasonal component (equal to 12 for monthly
series and 4 for quarterly data), B is the backshift operator, such that Byt D yt1 ,
p .B/ D .1  1 B      p Bp / is the nonseasonal Autoregressive (AR) operator of
order p, ˚P .Bs / D .1  ˚1 Bs      ˚P BPs / is the seasonal AR operator of order P,
q .B/ D .11 B  q Bq / is the nonseasonal Moving Average (MA) operator of
4.4 The X11ARIMA Method 89

order q, Q .Bs / D .11 Bs   Q BQs / is the seasonal MA operator of order Q,

and the at ’s are i.i.d. with mean zero and constant variance a2 . The .1B/d .1Bs /D
term implies nonseasonal differencing of order d and seasonal differencing of order
D. If d D D D 0 (no differencing), it is common to replace yt in (4.3) by deviations
from its mean.
The general multiplicative model (4.3) is said to be of order .p; d; q/.P; D; Q/s. Its
forecasting function can be expressed in different forms. For computational purpose,
the difference equation form is the most useful. Thus, at time tCl the ARIMA model
may be written as

ytCl D 1 ytCl1 CC f ytClf  atCl  1 atCl1      g atClg (4.4)

where f D pCsPCdCsD and g D qCsQ; .B/ D p .B/˚P .Bs /.1B/d .1Bs /D is

the general autoregressive operator and .B/ D q .B/Q .Bs / is the general moving
average operator.
Standing at origin t, to make a forecast l steps ahead of ytCl , denoted by yO t .l/, the
conditional expectation of (4.4) is taken at time t with the following assumptions:

Et .ytCl / D ytCl ; if l < 0I Et .ytCl / D yO t .l/; if l>0 (4.5)

Et .atCl / D atCl ; if l < 0I Et .atCl / D 0; if l > 0; (4.6)

where Et .ytCl / is the conditional expectation of ytCl taken at origin t. Thus, the
forecasts yO t .l/ for each lead time are computed from previous observed y’s, previous
forecast of y’s, and current and previous random shocks a’s. The unknown a’s
are replaced by zeros. In other words, they are weighted averages of previous
observations and forecasts made at previous lead times from the same time origin.
In general, if the moving average operator .B/ D q .B/Q .Bs / is of degree
qCsQ, the forecast equations for yO t .1/; yO t .2/; : : : ; yO t .qCsQ/ will depend directly on
the a’s, but forecasts at longer lead times will not. The latter will receive indirectly
the impact of the a’s by means of the previous forecasts. In fact, yO t .q C sQ C 1/ will
depend on the q C sQ previous yO t which in turn will depend on the a’s.
In the case of X11ARIMA, the lead time of the forecasts for monthly series is
often equal to 12 and the identification of the ARIMA models has a Q generally
equal to 1. Therefore, the forecasts depend directly on the a’s.
Finally, it is important to keep in mind that it is the general autoregressive
operator .B/ defined above that determines whether the forecasting function is to
be a polynomial, a mixture of sines and cosines, a mixture of exponentials, or some
combinations of these functions. The one-step ARIMA forecasts are minimum mean
square error forecasts and are easily updated as new raw values become available.
The ARIMA models for X11ARIMA must fulfill the double condition of (1)
fitting the data well, and (2) generating acceptable forecasts, where by acceptable
it is understood forecasts that have a mean absolute error smaller than 5 % for well
behaved series (e.g., employment males 20 years and over) and smaller than 12 %
for more irregular series (e.g., unemployment females 16–19 years).
90 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

In the automated version of this method, the user has the choice of: (a) submitting
his/her own ARIMA model, or (b) letting the program to select automatically an
ARIMA model that passes the above guidelines.
The automatic selection is made from three ARIMA models that have been
successfully chosen from a set of 12 ARIMA models testing out of sample
extrapolated values for the 4 last years. The sample consisted of 174 monthly and
quarterly economic time series of 15 years of length. The three models passed the
guidelines in about 80 % of all the cases. The 12 models originally tested were:

(1) (1, 1, 1) (1, 1, 1)s (7) log (2, 1, 1) (0, 1, 2)s

(2) (2, 1, 2) (0, 1, 1)s (8) log (0, 1, 2) (1, 1, 2)s
(3) (2, 0, 1) (0, 1, 2)s (9) log (0, 1, 1) (0, 1, 1)s
(4) (2, 0, 1) (1, 2, 1)s (10) log (0, 1, 1) (0, 2, 2)s
(5) (2, 0, 0) (0, 1, 1)s (11) log (0, 2, 2) (0, 1, 1)s
(6) (0, 0, 2) (0, 1, 1)s (12) log (0, 2, 2) (0, 2, 2)s

The automated option chooses for a multiplicative decomposition from the follow-
ing ARIMA models:

(1) log(0,1,1)(0,1,1)s
(2) log(0,2,2)(0,1,1)s
(3) (2,1,2)(0,1,1)s

and without the log transformation for an additive decomposition. It prints a

message indicating the failure of the three models, in which case the user can either
identify his own model or try again the automatic option submitting the same series
with prior modifications for outliers. The latter is strongly recommended because
the guidelines for acceptance or rejection of an ARIMA model are highly sensitive
to outliers.
The ARIMA models are robust and once well identified, they are generally
adequate for several years. It is advisable, however, to check the adequacy of the
model when an extra year of data becomes available to ensure that the most recent
movements of the series are taken into account.
Because the correct identification and estimation of ARIMA models have to
be done on series not affected by trading day variations and/or moving holidays,
these effects are estimated in a first round of the software without the ARIMA
identification. Once trading day variations and/or moving holidays are estimated,
they are removed from the original series and a second run is done using the above
three automated ARIMA models for extrapolation. The ARIMA model that gives
the smallest forecasting error for the last 3 years of data and passes a chi-square
probability test for the null hypothesis of randomness of the residual is chosen.
4.4 The X11ARIMA Method 91

4.4.3 Other Main Improvements Incorporated

into the Automated Version of X11ARIMA

A set of new statistical tests, tables, and graphs has been incorporated into the
automated version of the X11ARIMA. These tests are used to assess the quality of
the original series and the reliability of the seasonal adjustment. A brief description
of these main improvements follows.

4.4.3.1 An F-Test for the Presence of Seasonality in Table B1

This test is based on a one-way analysis of variance on the seasonal–irregular (SI)

ratios (differences), similar to the one already available in Census Method II-X11
variant for the presence of stable seasonality in Table D8. It differs only in that the
estimate of the trend-cycle is made directly from the original series, by a centered
12-term moving average.
The estimate of the trend-cycle is removed from the original series by dividing
into (subtracting from) the raw data for a multiplicative (additive) model.
The value of the F-ratio is printed in the computer program Table B1. It is the
quotient of two variances: (1) the “between months or quarters” variance, which
is due to seasonality, and (2) the “residual” variance, which is due to the irregular
component.
Because several of the basic assumptions in the F-test are likely to be violated,
the value of the F-ratio to be used, for rejecting the null hypothesis of no significant
seasonality being present, is not the one obtained from the tabulated F-distribution.
From experimentation with a large number of real and simulated series, it was
concluded that, for a monthly series of about 10 years, an F-value less than 10
indicates that there is not enough seasonality to justify using the Census II-X11
filters.

4.4.3.2 A Test for the Presence of Moving Seasonality in Table D8

The moving seasonality test is based on a two-way analysis of variance performed

on the SI ratios (differences) from Table D8 and was developed by Higginson [15].
It tests for the presence of moving seasonality characterized by gradual changes in
the seasonal amplitude but not in the phase.
The total variance of the SI ratios (differences) is considered as the sum of:
1. The “between months or quarters” variance, denoted by m2 , which mainly
measures the magnitude of the seasonality. It is equal to the sum of squares of the
difference between the average for each month of the SI and the overall average,
corrected by the corresponding degrees of freedom.
2. The “between years” variance, denoted by y2 , which mainly measures the
year-to-year movement of seasonality. It is equal to the sum of squares of the
92 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

differences between the annual average of the SI, for each year, and the average
of the SI for the whole table, corrected by the corresponding degrees of freedom.
3. The “residual” variance, denoted by r2 , equal to the total variance minus the sum
of the “between months or quarters” variance and the “between years” variance.
The F-ratio for the presence of moving seasonality is the quotient between the
“between years” variance and the “residual” variance.
To calculate the variances in an additive model, the (S C I) data are taken
in absolute value, otherwise the annual average is always equal to zero. For a
multiplicative model, the SI ratios are replaced by absolute deviations from 100,
i.e., by j SI  100 j. Contrary to the previous test, for which a high value of the F-
statistic is a good indication of the presence of identifiable seasonality, a high value
of the F-statistic corresponding to moving seasonality is a bad sign. The F-test,
printed in Table D8, indicates whether moving seasonality is present or not.
The presence of moving seasonality could be taken as an indication of residual
seasonality when Census II-X11 is used, because this program forces stable or
nearly stable seasonality on the first and last 3 years of data. This is partially
corrected by the ARIMA option.

4.4.3.3 A Test for the Presence of Identifiable Seasonality in Table D8

This test combines the previous test for the presence of moving seasonality with the
Census II-X11 test for the presence of stable seasonality and another nonparametric
test for the presence of stable seasonality, the Kruskal–Wallis chi-squared test,
whose value is also printed in Table D8.
The main purpose of this test is to determine whether the seasonality of the
series is “identifiable” or not. For example, if there is little stable seasonality and
most of the process is dominated by highly moving seasonality, the chances are that
the seasonal factors will not be well estimated because they will not be properly
identified by the Census II-X11 program.
This test basically consists of combining the F-values obtained from the two
previous tests as follows:
1. If the test for the presence of stable seasonality, denoted by FS , fails at the 1 %
level of significance, the null hypothesis (that seasonality is not identifiable) is
accepted.
2. If step 1 passes but the test for the presence of moving seasonality, indicated by
FM , fails at the 5 % level of significance, then this FM value is combined with the
FS value from step 1 to give

7 3FM
T1 D and T2 D ;
FM  FS FS
4.4 The X11ARIMA Method 93

and an average of these two statistics T1 and T2 , TN D .T1 C T2 /=2, is taken. If TN

is greater than or equal to one, the null hypothesis of identifiable seasonality not
being present is accepted.
3. If the FM -test passes, but either TN < 1 or the Kruskal–Wallis test fails at the
1 % level, then the program prints out that identifiable seasonality is “probably”
present.
4. If the FS , FM , and the Kruskal–Wallis chi-square value pass, then the null
hypothesis of identifiable seasonality not being present is rejected, and the
program prints “identifiable seasonality present.”
This test has been developed by Lothian and Morry [20], and the program
automatically prints out the messages mentioned above at the end of Table D8.

4.4.3.4 A Test for the Presence of Residual Seasonality in Table D11

This is an F-test applied to the values of Table D11 and calculated for the whole
length of the series, and for only the last 3 years. The effect of the trend is removed
by a first order difference of lag three, that is, (Oyt -Oyt3 ) where yO t are the values of
Table D11. Two F-ratios are printed at the end of the table, as well as a message
saying whether residual seasonality is present or not for the last 3 years.

4.4.3.5 A Test for the Randomness of the Residuals

The Census II-X11 program uses the Average Duration of Run (ADR) statistic to
test for autocorrelation in the final estimated residuals obtained from Table D13.
This nonparametric test was developed by Wallis and Moore [27] and is based
on the number of turning points. It is only efficient for testing the randomness of
the residuals against the alternative hypothesis that the errors follow a first order
autoregressive process of the form It D It1 C et , where  is the autocorrelation
coefficient and fet g is a white noise process. This test, however, is not efficient
for detecting the existence of periodicity in the residuals, which can occur when
relatively long series are seasonally adjusted, or when the relative variation of the
seasonal component is small compared to that of the irregulars.
To test the independence of the residuals against the alternative hypothesis of
being a periodic process, the normalized cumulative periodogram as developed by
Box and Jenkins [3] has been incorporated.
The normalized cumulative periodogram values are given in a table and also as a
graph. By visual inspection, it is possible to determine whether or not components
with certain periodicity are present in the residuals. If the residuals are a sample
realization of a purely random process, and if the size of the sample tends to infinity,
then the normalized cumulative periodogram tends to coincide with the diagonal of
the square in which it is drawn.
94 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

Deviations of the periodogram from those expected, assuming purely random

residuals, can be assessed by using the Kolmogorov–Smirnov test. This test is useful
to determine the nature of hidden periodicities left in the residuals, whether of
seasonal or cyclical character, and complements the information provided by the
test for the presence of residual seasonality.

4.4.3.6 A New Table D11A

This new table, D11A, produces a modified seasonally adjusted series, where the
annual totals of the seasonally adjusted values are made equal to those of the raw
data. The discrepancy between the annual totals is distributed over the seasonally
adjusted values of Table D11 in such a way as to minimize the distortion on
the month-to-month or quarter-to-quarter movements of the originally seasonally
adjusted series. The procedure is based on a quadratic minimization of the first
differences of the annual discrepancies, expressed as differences or ratios. This
procedure was originally developed by Hout [16] and improved by Cholette [5].

4.4.3.7 A Set of Guidelines Summarized in One Statistic that Helps

to Assess the Reliability of the Seasonal Adjustment

The Statistics Canada Census II-X11 version, as developed in 1975, had two
statistics called Q1 and Q2 that provided an indication of the size and nature of
the irregular and seasonal components, respectively. A description of these statistics
and their basic assumptions are discussed by Hout and De Fontenay [17].
Considerable research has been carried out since the first set of guidelines was
developed and there is now only one Q statistic, which results from a combination
of several other measures developed by Lothian and Morry [21]. Most of these
are obtained from the summary measures printed in Table F2 of the X11ARIMA
computer program. Their values vary between 0 and 3, and only values less than
1 are considered as acceptable. The set of statistics to produce the final Q statistic
is:
1. the relative contribution of the irregular component over spans of 3 months, as
obtained from Table F2, denoted by M1 ;
2. the relative contribution of the irregular component to the stationary portion of
the variance, denoted by M2 ;
3. the value of the I=C ratio between the average absolute month-to-month (or
quarter-to-quarter) percentage change in the irregular and that in the trend-
cycle, from Table D7 and denoted by M3 ;
4. the value of the average duration of run for the irregulars from Table F2 and
denoted by M4 ;
5. the MCD or QCD (the number of months or quarters it takes the average
absolute change in the trend-cycle to dominate that in the irregular), from Table
F2 and denoted by M5 ;
4.5 The X12ARIMA Method 95

6. the global I=S moving seasonality ratio, obtained as an average of the monthly
moving seasonality ratios from Table D9A and denoted by M6 (it is the ratio of
the average absolute year-to-year percentage change in the irregular factors to
that in the seasonal factors.);
7. the amount of stable seasonality in relation to the amount of moving seasonality
from the tests of Table D8 and denoted by M7 ;
8. a measure of the variation of the seasonal component for the whole series from
Table F2 and denoted by M8 ;
9. the average linear movement of the seasonal component for the whole series,
denoted by M9 ;
10. same as 8, but calculated for recent years only and denoted by M10 ; and
11. same as 9, but calculated for recent years only and denoted by M11 .
Furthermore, there are two new tables that show the outliers, and a D16 table
that shows the total effect of the trading day factors and the seasonal effects. Several
other modifications were introduced to correct for some limitations of the Census
II-X11 variant [7].

4.5 The X12ARIMA Method

4.5.1 General Outline

The X12ARIMA is today the most often applied seasonal adjustment method by
statistical agencies. It was developed by Findley et al. [9] and is an enhanced version
of the X11ARIMA method.
The major modifications concern: (1) extending the automatic identification and
estimation of ARIMA models for the extrapolation option to many more than the
three models available in X11ARIMA, and (2) estimating trading day variations,
moving holidays, and outliers in what is called regARIMA. The latter consists
of regression models with ARIMA (AutoRegressive Integrated Moving Average)
errors. More precisely, they are models in which the mean function of the time series
(or its logs) is described by a linear combination of regressors, and the covariance
structure of the series is that of an ARIMA process. If no regressors are used,
indicating that the mean is assumed to be zero, the regARIMA model reduces to
an ARIMA model.
Whether or not special problems requiring the use of regressors are present in
the series to be adjusted, a fundamentally important use of regARIMA models is
to extend the series with forecasts (and backcasts) in order to improve the seasonal
adjustments of the most recent (and the earliest) data. Doing this reduces problems
inherent in the trend estimation and asymmetric seasonal averaging processes of the
type used by the Census II-X11 method near the ends of the series. The provision
of this extension was the most important improvement offered by the X11ARIMA
program. Its theoretical and empirical benefits have been documented in many
publications, such as Dagum [7], Bobbit and Otto [2], and references therein.
96 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

The X12ARIMA method has all the seasonal adjustment capabilities of the
X11ARIMA variant. The same seasonal and trend moving averages are available,
and the program still offers the Census II-X11 calendar and holiday adjustment
routines incorporated in X11ARIMA. Several new options have been included,
namely:
1. sliding spans diagnostic procedures, developed by Findley et al. [10];
2. capability to produce the revision history of a given seasonal adjustment;
3. a new Henderson trend filter routine which allows the user to choose any odd
number for the length of the Henderson symmetric filter;
4. new options for seasonal filters;
5. several new outlier detection options for the irregular component of the seasonal
adjustment;
6. a new table of trading day factors by type of day; and
7. a pseudo-additive seasonal adjustment mode.
The modeling module is designed for regARIMA model building with sea-
sonal socioeconomic time series. To this end, several categories of predetermined
regression variables are available, including trend constants or overall means, fixed
seasonal effects, trading day effects, holiday effects, additive outliers, level shifts,
temporary change outliers, and ramp effects. User-specified regression variables can
also be included in the models.
The specification of a regARIMA model requires specification of both the regres-
sion variables to be included in the model and the type of ARIMA model for the
regression errors (i.e., the order .p; d; q/.P; D; Q/s ). Specification of the regression
variables depends on user knowledge about the series being modeled. Identification
of the ARIMA model for the regression errors follows well-established procedures
based on examination of various sample autocorrelation and partial autocorrelation
functions produced by the X12ARIMA program.
Once a regARIMA model has been specified, X12ARIMA estimates its param-
eters by maximum likelihood using an Iterative Generalized Least Squares (IGLS)
algorithm. Diagnostic checking involves examination of residuals from the fitted
model for signs of model inadequacy. X12ARIMA produces several standard
residual diagnostics for model checking, as well as provides sophisticated methods
for detecting additive outliers and level shifts. Finally, X12ARIMA can produce
point forecasts, forecast standard errors, and prediction intervals from the fitted
regARIMA model. In addition to these modeling features, X12ARIMA has an
automatic model selection procedure based mainly on the automatic model selection
procedure of TRAMO [12]. There are also options that use AICC, a version of
Akaike’s AIC that adjusts for the length of the series being modeled, to determine if
user-specified regression variables (such as trading day or Easter regressors) should
be included into a particular series. Also, histories can be generated for likelihood
statistics (such as AICC) and forecasts to facilitate comparisons between competing
models.
4.5 The X12ARIMA Method 97

4.5.2 The General RegARIMA Model

ARIMA models are frequently used for seasonal time series. A general multi-
plicative seasonal ARIMA model for a time series yt ; t D 1; : : : ; n, is specified
in Eq. (4.3). A useful extension of ARIMA models results from the use of a
time-varying mean function modeled via linear regression effects. More explicitly,
suppose we write a linear regression equation for the time series yt as

X
r
yt D ˇi xit C zt (4.7)
iD1

where xit ; i D 1; : : : ; r, are regression variables observed concurrently with yt , being

the Pˇi ; i D 1; : : : ; r are the corresponding regression parameters, such that zt D
yt  riD1 ˇi xit , the time series of regression errors, is assumed to follow the ARIMA
model defined in (4.3). Modeling zt as ARIMA addresses the fundamental problem
of applying standard regression methodology to time series data, that is, the standard
regression assumes that the regression errors (zt in (4.7)) are uncorrelated over time.
In fact, for time series data, the errors in (4.7) will usually be autocorrelated, and,
moreover, will often require differencing. Assuming that zt is uncorrelated will
typically lead to invalid results.
The expressions (4.3) and (4.7) taken together define the general regARIMA
model allowed by the X12ARIMA program, that can be written in a single equation
as
!
Xr
p .B/˚P .B /.1  B/ .1  B / yt 
s d s D
ˇi xit D q .B/Q .Bs /at : (4.8)
iD1

The regARIMA model (4.8) can be thought of either as generalising P the pure
ARIMA model (4.3) to allow for a regression mean function ( riD1 ˇi xit ), or as
generalising the regression model (4.7) to allow the errors zt to follow the ARIMA
model (4.3). In any case, notice that the regARIMA model implies that first the
regression effects are subtracted from yt to get the zero mean series zt , then the
error series zt is differenced to get a stationary series, say wt , assumed to follow a
stationary ARMA model, .B/˚.Bs /wt D .B/.Bs /at .
Another way to write the regARIMA model (4.8) is
X
.1  B/d .1  Bs /D yt D ˇi .1  B/d .1  Bs /D xit C wt ; (4.9)
i

where wt follows the stationary ARMA model just given before. Equation (4.9)
emphasizes that the regression variables xit in the regARIMA model, as well as in
the series yt , are differenced by .1  B/d .1  Bs /D .
Notice that the regARIMA model as written in (4.8) assumes that the regression
variables xit ; i D 1; : : : ; r, affect the dependent series yt only at concurrent time
98 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

points, i.e., model (4.8) does not explicitly provide for lagged regression effects such
as ˇi xi;t1 . Lagged effects can be included by the X12ARIMA program, however,
by reading appropriate user-defined lagged regression variables.
The X12ARIMA program provides additional flexibility in the specification
of the ARIMA part of the regARIMA model by permitting (1) more than two
multiplicative ARIMA factors, (2) missing lags within the AR and MA polynomials,
(3) the fixing of individual AR and MA parameters at user-specified values when the
model is estimated, and (4) inclusion of a trend constant, which is a nonzero overall
mean for the differenced series .1  B/d .1  Bs /D yt .
The most basic regression variable is the constant term. If the ARIMA model
does not involve differencing, this is the usual regression intercept, which, if there
are no other regression variables in the model, represents the mean of the series.
If the ARIMA model does involve differencing, X12ARIMA uses a regression
variable such that, when it is differenced according to the ARIMA model [see
Eq. (4.9)], a column of ones is produced. The corresponding parameter is then
called a trend constant, since it provides for a polynomial trend of the same degree
as the order of differencing in the model. For example, with d > 0 and D D 0,
the (undifferenced) trend constant regression variable is proportional to td . Notice
that the lower order polynomial terms, tj for 0  j < d, are not included among
the regression variables because they would be differenced to zero by .1  B/d ,
hence their coefficients cannot be estimated. With or without the trend constant,
the model (4.8) (or 4.9) implicitly allows for these lower order polynomial terms
through the differencing. If seasonal differencing is requested (D > 0), the nature
of the undifferenced trend constant regression variable is more complicated, though
the trend constant can be thought of as allowing for a polynomial of degree d C sD.
Without a trend constant, model (4.8) implicitly allows for a polynomial of degree
d C sD  1.
Fixed seasonal effects in a monthly series can be modeled using 12 indicator
variables, one for each calendar month. Since these 12 variables always add to
one, they are confounded with an overall level effect. This leads to one of two
singularity problems: collinearity with the usual constant term in a model with no
differencing; or a singularity in a model with differencing since the 12 variables,
when differenced, always sum to 0. One appropriate reparameterization instead uses
11 contrasts in the 12 indicator variables. An alternative reparameterization uses
11 variables taken from the Fourier (trigonometric) series representation of a fixed
monthly pattern. X12ARIMA allows either of these options, and also allows spec-
ifying the trigonometric terms only for selected frequencies. For quarterly series,
or for series with other seasonal periods, X12ARIMA constructs the appropriate
versions of these variables. Notice that these variables cannot be used in a model
with seasonal differencing, as they would all be differenced to zero.
Trading day effects occur when a series is affected by the different day-of-the-
week compositions of the same calendar month in different years. Trading day
effects can be modeled with seven variables that represent (no. of Mondays), . . . ,
(no. of Sundays) in month t. Bell and Hillmer [1] proposed a better parameterization
of the same effects using six variables defined as (no. of Mondays/  .no. of
4.5 The X12ARIMA Method 99

Sundays), . . . , (no. of Saturdays/  .no. of Sundays), along with a seventh variable

for Length of Month (LOM) or its deseasonalized version, the leap-year regressor
(lpyear). In X12ARIMA the six variables are called the tdnolpyear variables.
Instead of using a seventh regressor, a simpler and often better way to handle
multiplicative leap-year effects is to rescale the February values of the original
time series before transformation to m N Feb yt =mt , where yt is the original time series
before transformation, mt is the length of month t (28 or 29), and m N Feb D 28:25 is
the average length of February. If the regARIMA model includes seasonal effects,
these can account for the length-of-month effect except in Februaries, so the trading
day model only has to deal with the leap-year effect. When this is done, only the
tdnolpyear variables need to be included in the model. X12ARIMA allows explicit
choice of either approach, as well as an option (td) that makes a default choice
of how to handle length-of-month effects. When the time series being modeled
represents the aggregation of some daily series (typically unobserved) over calendar
months they are called monthly flow series. If the series instead represents the value
of some daily series at the end of the month, called a monthly stock series, then
different regression variables are appropriate. Trading day effects at the end-of-
month stock series can be modeled using seven indicator variables for the day of the
week that the months end on. Since the sum of these variables is always one, this
leads to a singularity problem, so six variables are used instead. X12ARIMA also
allows specification of regression variables appropriate for stock series defined as of
some other day of the month. Trading day effects in quarterly series are relatively
rare because the calendar composition of quarters does not vary as much over time,
on a percentage basis, as that of months do. Trading day variables are not provided
for flow time series with seasonal periods other than monthly or quarterly, or for
stock series other than monthly.
X12ARIMA also provides a simplified model for trading day variation of
monthly or quarterly series that uses only one regressor, a weekday–weekend
variable:
5
Dt D .no. of Weekdays/  .no. of Saturdays and Sundays/:
2
The underlying assumption for this model is that all weekdays (Monday–Friday)
have identical effects, and Saturday and Sunday have identical effects.
Holiday effects in a monthly flow series arise from holidays whose dates vary
over time if: (1) the activity measured by the series regularly increases or decreases
around the date of the holiday, and (2) this affects 2 (or more) months depending on
the date the holiday occurs each year. Effects of holidays with a fixed date, such as
Christmas, are indistinguishable from fixed seasonal effects. Easter effects are the
most frequently found holiday effects in American and European economic time
series, since the date of Easter Sunday varies between March 22 and April 25. Labor
Day and Thanksgiving also are found in American and Canadian time series. The
basic model used by X12ARIMA for Easter and Labor Day effects assumes that
the level of activity changes on the k-th day before the holiday for a specified k,
100 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

and remains at the new level until the day before the holiday. For Thanksgiving the
model used assumes that the level of activity changes on the day that is a specified
number of days before or after Thanksgiving and remains at the new level until
December 24. The regression variable constructed for the holiday effect is, for a
given month t, the proportion of the affected time period that falls in month t.
Actually, these regressors are deseasonalized by subtracting their long-run monthly
means. Essentially the same Easter effect variable applies also to quarterly flow time
series, but Labor Day and Thanksgiving effects are not present in quarterly series.
X12ARIMA provides four other types of regression variables to deal with abrupt
changes in the level of a series of a temporary or permanent nature: additive outliers
(AO), level shifts (LS), temporary changes (TC), and ramps. AOs affect only one
observation in the time series, LSs increase or decrease all observations from a
certain time point onward by some constant amount, TCs allow for an abrupt
increase or decrease in the level of the series that returns to its previous level
exponentially rapidly, and ramps allow for a linear increase or decrease in the level
of the series over a specified time interval. LS regression variables are defined as
1 and then 0, in preference to an equivalent 0 and then 1, to make the overall
level of the regression mean function of any forecasts consistent with the most
recent level of the time series. Similar considerations dictate the definition of ramp
variables. Often, however, large seasonal movements make it difficult to identify
where such changes in level have occurred. Identifying the location and nature of
potential outliers is the object of the outlier detection methodology implemented .
This methodology can be used to detect AOs, TCs, and LSs (not ramps); any that are
detected are automatically added to the model as regression variables. Prespecified
AOs, LSs, TCs, and ramps are actually simple forms of interventions as discussed
by Box and Tiao [4]. While X12ARIMA does not provide the full range of dynamic
intervention effects discussed by Box and Tiao [4], often a short sequence of suitably
chosen AO, LS, TC, and/or ramp variables can produce reasonable approximations
to more complex dynamic intervention effects, although with one or two additional
parameters.

4.6 Illustrative Example: X12ARIMA Seasonal Adjustment

of the US NODG Series

The X12ARIMA seasonal adjustment program is an enhanced version of the

X11ARIMA method [7]. The enhancements include a more self-explanatory and
versatile user interface and a variety of new diagnostics to help the user detect
and remedy any inadequacies in the seasonal and calendar adjustments obtained
under the program options selected. As described in the previous sections, the
program also includes a variety of new options to overcome adjustment problems
and thereby enlarge the range of economic time series that can be adequately
seasonally adjusted.
4.6 Illustrative Example: X12ARIMA Seasonal Adjustment of the US NODG. . . 101

The main source of these new options is an extensive set of time series model
building facilities for fitting what are called regARIMA models, that is, regression
models with ARIMA errors. A fundamental use of regARIMA models is to
extend the series with forecasts (and backcasts) in order to improve the seasonal
adjustments of the most recent (and the earliest) data. Doing this mitigates problems
inherent in the trend estimation and asymmetric seasonal averaging processes of
the type used by the Census Method II-X11 near the ends of the series. The
introduction of the ARIMA extension was a major improvement introduced by
Estela Bee Dagum in the X11ARIMA. Its benefits, both theoretical and empirical,
have been documented in many publications, see Dagum [7], Bobbitt and Otto [2],
and references therein.
X12ARIMA is available as an executable program for computers running DOS
(version 3.0 or higher), Sun 4 UNIX workstations, and VAX/VMS computers.
When released, the X12ARIMA program is in the public domain, and can be
copied or transferred. The corresponding documentation and examples are available
online at http://www.census.gov/srd/www/x12a/. X12ARIMA is also implemented
in OxMetrics, in the open source software Gretl, and in the Eurostat software
Demetra+, and its recent extension JDemetra+. Since the 1990s, in the European
countries, Eurostat has been playing a leading role in the promotion, development,
and maintenance of source available software solution for seasonal adjustment of
time series in line with established best practices. In this regard, Eurostat developed
Demetra to provide a convenient and flexible tool for seasonal adjustment using
either TRAMO/SEATS and X12ARIMA methods. Subsequently, Demetra+ was
developed by the National Bank of Belgium, but due to constraints implied by the
use of FORTRAN components, Eurostat created a new seasonal adjustment software
which is platform independent and extensible, that is, JDemetra+ [13]. It is highly
influenced by the output of TRAMO/SEATS, structural models, and X12ARIMA.
JDemetra+ runs on operating systems that support the Java VM (Virtual Machine),
such as Microsoft Windows XP SP3/Vista SP1/Windows 7, Ubuntu 9.10, various
other Linux distributions, Solaris OS version 11 Express (SPARC and x86/x64
Platform Edition), and Macintosh OS X 10.6 Intel.
In this section, we make use of the X12ARIMA program as developed by the
US Bureau of Census to perform the seasonal adjustment of the US New Orders for
Durable Goods (NODG) series, observed from February 1992 to December 2013.
We do not provide a detailed discussion of all the tables and results produced by the
software, but we focus on the most important outputs for the interpretation of the
seasonal adjustment of the NODG series.

4.6.1 Input: Specification File

To apply X12ARIMA to any particular time series, an input file, called specification
file, must be created. This ASCII (or text) file contains a set of specifications that
X12ARIMA reads to obtain the information it needs about the time series data,
102 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

the time series model to be used, the analysis to be performed, and the output
desired. This specification file has the extension .spc. It is common practice to allow
the software to test automatically for the logarithmic or level specification, for the
presence of trading days effects as well as Easter and leap-year effects. Furthermore,
an automatic detection of several types of outliers [Additive Outliers (AO) and
Level Shift (LS)] is generally required by the software. The most appropriate
ARIMA(p; d; q)(P; D; Q/s model can be automatically selected by the software.
Based on these considerations, we have written the specification file presented in
Table 4.1. In detail, we have selected the following options:
• series describes the time series data (data), start date (start), seasonal period
(period), and the series title (title);
• transform illustrates that we allow the software to automatically select the most
appropriate transformation (level or log) for the data;
• regression specifies the regression variables used to form the regression part of
the regARIMA model. In particular, we allow the estimation of the trading day
effects (td), Easter effect over a span of 9 days before and after Easter (easter[9]),
as well as of the leap-year effect if the log transformation is selected;
• outlier allows for an automatic detection of additive outliers and/or level shifts
using the estimated model;
• pickmdl specifies an automatic selection procedure based on X11ARIMA [7]. It
differs from the option automdl according to which an automatic model selection
procedure is performed based on TRAMO [12]. The latter specification is imple-
mented by default in both Gretl and JDemetra+. However, it should be noticed
that we have found that this option generally selects less parsimonious models
than needed. The main reason be the fact that it is automatically used together
with the simplified trading day model with only one regressor. This generally
produces wrong estimates that are left in the residuals and must be picked up
by the ARIMA model which is then more complex than necessary. Using this
option the model chosen for the log transformed data of the US New Orders for
Durable Goods time series is the ARIMA(1,1,2)(0,1,1)12, but using the trading

Table 4.1 Specification file series{

to perform the X12ARIMA
period=12
seasonal adjustment of the
NODG series title=“nodg”
start=
data=(. . . )
}
transform{function=auto}
regression{variables = (td easter[9])}
outlier{}
pickmdl{}
check{print=all}
x11{save=( d11 d12 d13 ) print=alltables}
4.6 Illustrative Example: X12ARIMA Seasonal Adjustment of the US NODG. . . 103

day with six regressors and the automatic selected ARIMA(0,1,1)(0,1,1)12 we

obtained better results. As discussed by the main X11ARIMA guidelines [7] and
by Fischer and Planas [11], the airline model, ARIMA(0,1,1)(0,1,1)s, is generally
found to be appropriate in the seasonal adjustment of the major economic and
social indicators. In this regard, we recommend to use the trading day model
based on six regressors together with the pickmdl option in X12ARIMA;
• check produces statistics useful to perform a diagnostic checking of the estimated
model. Based on the print=all specification, the software provides summary
statistics of the residuals, corresponding sample ACF and PACF with associated
standard errors, Ljung–Box Q statistics, normality test statistics, and a spectral
plot of the model residuals, as well as an histogram of the standardized residuals;
• x11 describes the seasonal adjustment options, including mode of adjustment,
seasonal and trend filters, and Easter holiday adjustment option, and some
seasonal adjustment diagnostics. Based on the selected option save=(d11 d12
d13) the seasonally adjusted series, the estimated trend-cycle, and irregular
components are saved in the workspace. Furthermore, based on print=alltables,
all the tables produced by the X12ARIMA software are displayed in the output.

4.6.2 Testing for the Presence of Identifiable Seasonality

Before going through the X12ARIMA decomposition, it is necessary to check if the

original series presents identifiable seasonality. In this regard, Table D8 provides
combined seasonality tests that are useful to check for the presence of seasonality in
the original series. This test combines the test for the presence of moving seasonality
with the F-test for the presence of stable seasonality and the Kruskal–Wallis chi-
squared test, another nonparametric test for the presence of stable seasonality. The
main purpose of this test is to determine whether the seasonality of the series is
identifiable or not. For example, if there is little stable seasonality and most of the
process is dominated by rapidly moving seasonals, chances are that the seasonals
will not be accurately estimated for they will not be properly identified by the
X12ARIMA method. Table 4.2 illustrates the output for the NODG series.
The nonparametric Friedman test is used to determine if stable seasonality
is present in the series. This test uses a preliminary estimate of the unmodified
Seasonal–Irregular (SI) component. Indeed, the seasonal component includes the
intra-year variation that repeats each year (stable seasonality) or evolves from year
to year (moving seasonality). A high test statistic and low p-value indicate that a
significant amount of variation in the SI ratios is due to months (or quarters), which
in turn is evidence of seasonality. If the p-value is lower than 1 %, the null hypothesis
of no seasonal effect is rejected. Conversely, a small value of the F-test and great
p-value (close to 1.0) is evidence that variation due to months or quarters could be
determined by the random error and the null hypothesis of no month/quarter effect
is not rejected. As shown in Table 4.2, in our specific case, the null hypothesis is
rejected, such that seasonality is present.
104 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

Table 4.2 Combined seasonality test performed on the NODG series

D8.A F-tests for seasonality

Test for the presence of seasonality assuming stability.

Sum of squares Dgrs. freedom Mean square F-value
Between months 9305:6408 11 845:96735 111:033  
Residual 1912:3856 251 7:61907
Total 11;218:0264 262
**Seasonality present at the 0.1 % level.

Nonparametric test for the presence of seasonality assuming stability

Kruskal–Wallis statistic Dgrs. freedom Probability level
213:9990 11 0:000 %
Seasonality present at the 1 % level.

Moving Seasonality Test

Sum of squares Dgrs. freedom Mean square F-value
Between years 60:1151 20 3:005756 0:495
Error 1335:2196 220 6:069180
No evidence of moving seasonality at the 5 % level.

COMBINED TEST FOR THE PRESENCE OF IDENTIFIABLE SEASONALITY

IDENTIFIABLE SEASONALITY PRESENT

Kruskal–Wallis test is the second test for stable seasonality provided by the
software. The test is calculated for the final estimation of the unmodified SI
component from which k samples are derived (k D 12 for monthly series and
k D 4 for quarterly series) of size n1 ; n2 ; : : : ; nk , respectively. The test is based
on the statistic:

12 X Sjk
WD  3.n C 1/;
n.n C 1/ jD1 nj

where Sj is the sum of the ranks of the observations from the j-th sample within
P
the whole sample of n D kjD1 nj observations. Under the null hypothesis, the test
statistic follows a chi-square distribution with k  1 degrees of freedom. Table 4.2
shows that, for the NODG series, the outcome of the test is stable seasonality
present.
A more detailed description of these two tests is provided in Sect. 4.4.3, and
we refer the reader to that section for the theoretical description of the moving
seasonality test as well as of the combined seasonality test. As shown in Table 4.2,
4.6 Illustrative Example: X12ARIMA Seasonal Adjustment of the US NODG. . . 105

all the tests confirm the presence of stable seasonality and the absence of moving
seasonality in the US NODG series.

4.6.3 Pre-processing

As discussed in the previous sections, the X12ARIMA method consists of two

linked parts: the regARIMA model for estimation of the deterministic components,
and the decomposition part of the linearized series for the stochastic components
performed using the Census II-X11 filters combined with those of the ARIMA
extrapolation. Part A of the X12ARIMA output provides the results for the user
selected options to correct a priori the series by introducing adjustment factors. For
the NODG series, these latter include correction for the effect of Easter holiday, of
changes in the level of the series (e.g., effect of a strike), and also introduce seven
weights for each day of the week to take into account the variations due to the
trading day composition of the months.
The first information provided by the software is about the automatic selection
between the level and logarithmic transformation for the data as illustrated in
Table 4.3. In particular, likelihood statistics for model fit to untransformed and
log transformed series are displayed. It can be noticed that for the NODG series
the logarithmic transformation is applied and a multiplicative seasonal adjustment
model is selected.

4.6.3.1 Regressors

Based on our specification, the X12ARIMA provides the estimated coefficients

of the trading days variables to evaluate their significant effect on the NODG
series. Table 4.4 illustrates that the variables related to Wednesday and Sunday
(derived) are characterized by coefficients that are significantly different from one.
Specifically, Fig. 4.1 displays the monthly trading days component that ranges from
98.7 to 103.99 over the observed period February 1992–December 2013, being on
average equal to 100. The Easter effect is also evaluated in the model specification,
and the value of the estimated corresponding coefficient is reported with its standard
error and t-statistic. As shown in Table 4.4, there is a significant effect lasting over
9 days around Easter. In particular, Fig. 4.2 displays that the Easter effect is felt
9 days before Easter and on Easter Sunday but not after Easter. This is evidenced
by years 1997, 1999, 2002, 2005, 2008, 2010, and 2013, where Easters fall very
early in April or in March. Indeed, orders are generally affected by Easter, because
customs do not operate from Good Friday to Easter Monday.
The software automatically detects the presence of three different outliers: an
additive outlier at June 2000, and two level shift outliers on October 2008 and
January 2009, respectively.
106 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

Table 4.3 Automatic selection between level and logarithmic transformation of the NODG series
Likelihood statistics for model fit to untransformed series.

Likelihood statistics
Number of observations (nobs) 263
Effective number of observations (nefobs) 250
Number of parameters estimated (np) 11
AIC 5258.6176
AICC (F-corrected-AIC) 5259.7268
Hannan–Quinn 5274.2077
BIC 5297.3537
Likelihood statistics for model fit to log transformed series.

Likelihood statistics
Number of observations (nobs) 263
Effective number of observations (nefobs) 250
Number of parameters estimated (np) 10
Log likelihood 428.0586
Transformation Adjustment 3029.8684
Adjusted Log likelihood (L) 2601.8098
AIC 5223.6195
AICC (F-corrected-AIC) 5224.5400
Hannan–Quinn 5237.7924
BIC 5258.8342
AICC (with aicdiffD 2:00) prefers log transformation
Multiplicative seasonal adjustment will be performed.

4.6.3.2 ARIMA Model

For the log transformed NODG series under study, the following ARIMA(0,1,1)
(0,1,1)12 (airline) model has been automatically selected and estimated by the
software:

.1  B/.1  B12 /Yt D .1  0:5357B/.1  0:7018B12/at ;

where at is assumed to follow a white noise process with zero mean and estimated
variance equal to 0.0124. All the coefficients result significantly different from zero
and satisfy the invertibility conditions (Table 4.5).
Similar to X11ARIMA the ARIMA model to be used in the context of the
X12ARIMA method must fulfill the double condition of fitting the data well
and generating “acceptable” projections for the last 3 years of observed data. By
“acceptable” projection is meant a projection with a mean absolute error smaller
than 5 % for well-behaved series (e.g., unemployment adult males) and smaller than
12 % for highly irregular series (e.g., unemployment teenage males). In this regard,
4.6 Illustrative Example: X12ARIMA Seasonal Adjustment of the US NODG. . . 107

Table 4.4 Estimates of calendar effects and outliers detection

Regression model
Parameter Standard
Variable estimate error t-Value
Trading day
Mon 0:0048 0:00427 1:12
Tue 0:0076 0:00426 1:78
Wed 0:0129 0:00424 3:05
Thu 0:0034 0:00423 0:81
Fri 0:0059 0:00428 1:38
Sat 0:0025 0:00425 0:59
*Sun (derived) 0:0101 0:00427 2:37

Easter [9] 0:0250 0:00898 2:79

Automatically identified outliers

AO2000.Jun 0:1547 0:02904 5:32
LS2008.Oct 0:1808 0:02803 6:45
LS2009.Jan 0:2358 0:02808 8:40
*For full trading day and stable seasonal effects, the derived
parameter estimate is obtained indirectly as minus the sum of
the directly estimated parameters that define the effect.

Chi-squared tests for groups of regressors

Regression effect df Chi-square p-value
Trading day 6 28:88 0:00

it is recommended the identification of the ARIMA model be performed on data in

which extreme values have been previously modified. This recommendation is even
more relevant if the outliers fall in the most recent years, in order to avoid the rejec-
tion of good models simply because the outliers have inflated the absolute average
extrapolation error above the acceptance level of the guidelines. It can be noticed in
Table 4.6 that these requirements are satisfied by the estimated ARIMA model.
To determine whether or not a model fits the data well, the Portmanteau test of
fit developed by Box and Pierce [3] with the variance correction for small samples
introduced by Ljung and Box [19] is used. As shown in Table 4.6, the null hypothesis
of randomness of the residuals is tested and not rejected at a 10 % level.

4.6.3.3 Residuals

Beside the Portmanteau test previously introduced, a more complete analysis of the
residuals of the estimated ARIMA model has been carried out by specifying the
option check in the specification file. As suggested by the Portmanteau test, the
residuals satisfy the assumption of being generated by a purely random process
108 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

102
101
100
99
98

1995 2000 2005 2010

Fig. 4.1 Trading day estimates of the US New Orders for Durable Goods
101.5
100.5
99.5
98.5

1995 2000 2005 2010

Fig. 4.2 Moving holiday component of the US NODG

4.6 Illustrative Example: X12ARIMA Seasonal Adjustment of the US NODG. . . 109

Table 4.5 Estimated ARIMA(0,1,1)(0,1,1)12 model

ARIMA model: (0 1 1)(0 1 1)
Nonseasonal differences: 1
Seasonal differences: 1
Standard
Parameter Estimate errors
Nonseasonal MA
Lag 1 0.5357 0.05229

Seasonal MA
Lag 12 0.7018 0.04671

Variance 0.12400E02
SE of Var 0.11091E03

Likelihood statistics
Number of observations (nobs) 263
Effective number of observations (nefobs) 250
Number of parameters estimated (np) 13
Log likelihood 477.6115
Transformation Adjustment 3029.8684
Adjusted Log likelihood (L) 2552.2569
AIC 5130.5139
AICC (F-corrected-AIC) 5132.0562
Hannan–Quinn 5148.9386
BIC 5176.2928

Table 4.6 Average absolute percentage error in forecasts and goodness of fit for the estimated
ARIMA(0,1,1)(0,1,1)12 model
Average absolute percentage error in within-sample forecasts:
Last year: 3.27 Last 1 year: 6.93
Last 2 years: 2.09 Last 3 years: 4.10

Chi-square Probability: 17.34 % (Q D 28.0679, 22 df)

with constant mean and variance. However, to verify this assumption in more
detail, several statistics are computed on these residuals. Furthermore, the normality
assumption on the residual distribution is not rejected.
110 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

4.6.4 Decomposition

The Decomposition part includes tables with results from consecutive iterations
of the Census II-X11 filters and quality measures. As discussed in Sects. 4.4
and 4.5, to estimate the different components of the series, while taking into account
the possible presence of extreme observations, X12ARIMA proceeds iteratively
through four processing stages, denoted as A, B, C, and D, plus two stages, E,
and Quality measures, that propose statistics and charts and are not part of the
decomposition per se. Stage A has been widely discussed in the previous section.
On stages B, C, and D the basic algorithm is used. Specifically,
B. First automatic correction of the series. This stage consists in a first estimation
and downweighting of the extreme observations and a first estimation of the
working day effects. This stage is performed by applying the basic algorithm
detailed earlier. These operations lead to Table B20, adjustment values for
extreme observations, used to correct the unadjusted series and result in the
series from Table C1.
C. Second automatic correction of the series. Still applying basic algorithm once
again, this part leads to a more precise estimation of replacement values of the
extreme observations (Table C20). The series, finally “cleaned up”, is shown in
Table D1 of the printouts.
D. Seasonal adjustment. This part, at which the basic algorithm is applied for the
last time, is that of the seasonal adjustment per se, as it leads to final estimates:
• of the seasonal component (Table D10);
• of the seasonally adjusted series (Table D11);
• of the trend-cycle component (Table D12); and
• of the irregular component (Table D13).
X12ARIMA selects the filters automatically, taking into account the global
moving seasonality ratio, which is computed on preliminary estimates of the
irregular component and of the seasonal. For the NODG series, a 3  5 m.a.
has been selected to estimate the seasonal component, whereas the trend-cycle
has been estimated using a 13-term Henderson filter, being the signal-to-noise
ratio equal to 2.47. The seasonally adjusted NODG series is shown in Fig. 4.3
together with the original series, and the irregular component in Fig. 4.4.
E. Components modified for large extreme values. Part E includes: (1) components
modified for large extreme values, (2) a comparison between the annual totals of
the raw time series and seasonally adjusted time series, (3) changes in the final
seasonally adjusted series, (4) changes in the final trend-cycle, and (5) robust
estimation of the final seasonally adjusted series. The results from part E are
used in the subsequent part to calculate the quality measures.
F. Quality measures. The final part contains statistics for judging the quality of
the seasonal adjustment, as described in Sects. 4.4 and 4.5. The M statistics
are used to assess the quality of the seasonal adjustment. They account for the
relative contribution of the various components (irregular, trend-cycle, seasonal,
4.6 Illustrative Example: X12ARIMA Seasonal Adjustment of the US NODG. . . 111

NODG
Seasonally adjusted series
200000
100000

1995 2000 2005 2010

Fig. 4.3 NODG original and seasonally adjusted series

115
110
105
100
95

1995 2000 2005 2010

Fig. 4.4 Irregular component for the US NODG series

preliminary factors, trading days, and holidays effects) to the variance of the
stationary part of the original time series. These statistics vary between 0 and 3,
but only values smaller than 1 are acceptable. Table 4.7 illustrates these statistics
for the components estimated in the X12ARIMA seasonal adjustment of the
NODG series. It can be noticed that all the measures range from 0 to 3, indicating
a good seasonal adjustment for the series.
 112

Table 4.7 Diagnostic checking on the residuals of the estimated ARIMA(0,1,1)(0,1,1)12 model
F3. Monitoring and quality assessment statistics
The measures below are between 0 and 3; acceptance region from 0 to 1.

1. The relative contribution of the irregular over 3 months span (from Table F2.B). M1 D 0.827
2. The relative contribution of the irregular component to the stationary portion of the variance (from Table F 2.F). M2 D 0.939
3. The amount of month-to-month change in the irregular component as compared to the amount of month-to-month
change in the trend-cycle (from Table F2.H). M3 D 0.736
4. The amount of autocorrelation in the irregular as described by the average duration of run (Table F 2.D). M4 D 0.285
5. The number of months it takes the change in the trend-cycle to surpass the amount of change in the irregular (from Table F 2.E). M5 D 0.625
6. The amount of year-to-year change in the irregular as compared to the amount of year-to-year change in the seasonal (from Table F 2.H). M6 D 0.295
7. The amount of moving seasonality present relative to the amount of stable seasonality (from Table F 2.I). M7 D 0.195
8. The size of the fluctuations in the seasonal component throughout the whole series. M8 D 0.518
9. The average linear movement in the seasonal component throughout the whole series. M9 D 0.236
10. Same as 8, calculated for recent years only. M10 D 0.454
11. Same as 9, calculated for recent years only. M11 D 0.332

ACCEPTED at the level 0.50

Q (without M2) D 0.45 ACCEPTED.
4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants
References 113

A composite indicator calculated from M statistics is the Q statistic, defined

as follows:
10M1 C 12M2 C 10M3 C 8M4 C 11M5 C 10M6 C 18M7 C 7M8 C 7M9 C 4M10 C 4M11
QD
100

Q without M2 (also called Q2 ) is the Q statistic without the M2 statistics. If time

series does not cover at least 6 years, statistics M8 ; M9 ; M10 , and M11 cannot be
calculated. In this case, the Q statistics is computed as:

14M1 C 15M2 C 10M3 C 8M4 C 11M5 C 10M6 C 32M7 C 0M8 C 0M9 C 0M10 C 0M11
QD :
100

The model has satisfactory quality if the Q statistic is less than 1. In this regard,
it can be noticed that the X12ARIMA seasonal adjustment of the NODG series
is satisfactory.

References

1. Bell, W. H., & Hillmer, S. C. (1984). Issues involved with the seasonal adjustment of economic
time series. Journal of Business and Economic Statistics, 2, 291–320.
2. Bobbit, L., & Otto, M. C. (1990). Effects of forecasts on the revisions of seasonally
adjusted values using the X11 seasonal adjustment procedure. In Proceedings of the American
Statistical Association Business and Economic Statistics Session (pp. 449–453).
3. Box, G. E. P., & Jenkins, G. M. (1970). Time series analysis: Forecasting and control. Holden-
Day: San Francisco (2nd ed., 1976).
4. Box, G. E. P., & Tiao, G. C. (1975). Intervention analysis with applications to economics and
environmental problems. Journal of the American Statistical Association, 70, 70–79.
5. Cholette, P. A. (1979). A comparison and assessment of various adjustment methods of sub-
annual series to yearly benchmarks. Research paper. Seasonal Adjustment and Time Series
Staff, Statistics Canada, Ottawa.
6. Dagum, E. B. (1978). Modelling, forecasting and seasonally adjusting economic time series
with the X-11 ARIMA method. The Statistician, 27(3), 203–216.
7. Dagum, E. B. (1988). The X-11 ARIMA/88 seasonal adjustment method-foundations and user’s
manual. Statistics Canada, Ottawa: Time Series Research and Analysis Centre.
8. Dagum, E. B., Chhab, N., & Chiu, K. (1990). Derivation and properties of the X11ARIMA
and census X11 linear filters. Journal of Official Statistics, 12(4), 329–347.
9. Findley, D. F., Monsell, B. C., Bell, W. R., Otto, M.C., & Chen B. C. (1998). New capabilities
and methods of the X12ARIMA seasonal adjustment program. Journal of Business and
Economic Statistics, 16(2), 127–152.
10. Findley, D. F., Monsell, B. C., Shulman, H. B., & Pugh, M. G. (1990). Sliding spans diagnostics
for seasonal and related adjustments. Journal of the American Statistical Association, 85, 345–
355.
11. Fischer, B., & Planas, C. (2000). Large scale fitting of regression models with ARIMA errors.
Journal of Official Statistics, 16(2), 173–184.
12. Gomez, V., & Maravall, A. (1996). Program TRAMO and SEATS. Madrid: Bank of Spain.
13. Grudkowska, S. (2015). JDemetra+ Reference manual. Warsaw: National Bank of Poland.
14. Henderson, R. (1916). Note on graduation by adjusted average. Transaction of Actuarial
Society of America, 17, 43–48.
114 4 Linear Filters Seasonal Adjustment Methods: Census Method II and Its Variants

15. Higginson, J. (1975). An F-test for the presence of moving seasonality when using Census
Method II X-11 variant. Research paper, Seasonal Adjustment and Time Series Staff, Statistics
Canada, Ottawa.
16. Huot, G. (1975). Quadratic minimization adjustment of monthly or quarterly series to annual
totals. Research paper. Seasonal Adjustment and Time Series Staff, Statistics Canada, Ottawa.
17. Huot, G., & De Fontenay, A. (1973). General seasonal adjustment guidelines. Part 1, Statistics
Canada’s version of the U.S. Bureau of the Census monthly multiplicative and additive X-
11 program. Research paper. Seasonal Adjustment and Time Series Staff, Statistics Canada,
Ottawa
18. Ladiray, D., & Quenneville, B. (2001). Seasonal adjustment with the X11 method. Lecture
notes in statistics (Vol. 158). New York: Springer.
19. Ljung, G., & Box, G. (1979). The likelihood function of stationary autoregressive-moving
average models. Biometrika, 66, 265–270.
20. Lothian, J., & Morry, M. (1978a). A test for the presence of identifiable seasonality when
using the X-11 program. Research paper. Seasonal Adjustment and Time Series Analysis Staff,
Statistics Canada, Ottawa.
21. Lothian, J., & Morry, M. (1978b). A set of guidelines to assess the reliability of seasonal
adjustment by the X-11 program. Research paper. Seasonal Adjustment and Time Series Staff,
Statistics Canada, Ottawa.
22. Macaulay, F. R. (1931). The smoothing of time series. New York: National Bureau of Economic
Research.
23. Musgrave, J. (1964). A set of end weights to end all end weights. Working paper. U.S. Bureau
of Census, Washington, DC.
24. Newbold, P., & Granger, C.W. J. (1974). Experience with forecasting univariate time series and
the combination of forecasts (with discussion). Journal of the Royal Statistical Society. Series
A, 137, 131–165.
25. Reid, D. J. (1975). A review of short-term projection techniques. In H. A. Gorden (Ed.),
Practical aspects of forecasting (pp. 8–25). London: Operational Research Society.
26. Shiskin, J., Young, A. H. & Musgrave, J. C. (1967). The X-11 variant of the Census Method
II seasonal adjustment program. Technical Paper 15 (revised). US Department of Commerce,
Bureau of the Census, Washington, DC.
27. Wallis, W. A., & Moore, G. H. (1941). A significance test for time series analysis. Journal of
the American Statistical Association, 36, 401–409.
28. Young, A. H. (1968). A linear approximation to the Census and BLS seasonal adjustment
methods. Journal of the American Statistical Association, 63, 445–457.
Chapter 5
Seasonal Adjustment Based on ARIMA Model
Decomposition: TRAMO-SEATS

Abstract TRAMO-SEATS is a seasonal adjustment method based on ARIMA

modeling. TRAMO estimates via regression dummy variables and direct ARIMA
modeling (regARIMA) the deterministic components, trading days, moving hol-
idays, and outliers, which are later removed from the input data. In a second
round, SEATS estimates the stochastic components, seasonality, and trend-cycle,
from an ARIMA model fitted to the data where the deterministic components are
removed. SEATS uses the filters derived from the ARIMA model that describes
the behavior of the time series. By imposing certain conditions, a unique canonical
decomposition is performed to obtain the ARIMA models for each component. This
chapter discusses with details the estimation methods used by TRAMO and SEATS
as well as the basic assumptions on the derivation of ARIMA models for each
component. An illustrative example of the seasonal adjustment with the TRAMO-
SEATS software default option is shown with the US New Orders for Durable Goods
series. The illustrative example concentrates on the most important tables of this
software that enable to assess the quality of the seasonal adjustment.

Peter Burman [4] was the first to develop a seasonal adjustment method based on
ARIMA model decomposition that he named SIGEX (Signal Extraction). Later,
working on the same topic, Hillmer and Tiao [16] developed what these authors
called ARIMA model-based seasonal adjustment, largely discussed in Bell and
Hillmer [2]. An ARIMA model is identified from the observed data and, by
imposing certain restrictions, models for each component are derived. Since the
components are unknown, to obtain a unique solution Hillmer and Tiao proposed
a canonical decomposition which has the property of maximizing the variance of
the irregulars and minimizing the variance of the estimated components. Because
ARIMA model identification and estimation are not robust to outliers or extreme
values and cannot deal with deterministic components such as trading days and
moving holidays, further changes were made by combining dummy variables
regression models with ARIMA models. In this regard, Gomez and Maravall [12]
produced at the Bank of Spain a seasonal adjustment software called TRAMO-
SEATS which is currently applied mainly by European statistical agencies. TRAMO

© Springer International Publishing Switzerland 2016 115

E. Bee Dagum, S. Bianconcini, Seasonal Adjustment Methods and Real Time
Trend-Cycle Estimation, Statistics for Social and Behavioral Sciences,
DOI 10.1007/978-3-319-31822-6_5
116 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

stands for “Time series Regression with ARIMA noise, Missing observations
and Outliers,” and SEATS for “Signal Extraction in ARIMA Time Series.” First,
TRAMO estimates via regression the deterministic components, which are later
removed from the input data. In a second round, SEATS estimates the seasonal
and trend-cycle components from the ARIMA model fitted to the data where the
deterministic components are removed. SEATS uses the filters derived from the
ARIMA model that describes the stochastic behavior of the linearized time series.
It should be mentioned that Eurostat, in collaboration with the National Bank
of Belgium, the US Bureau of the Census, the Bank of Spain, and the European
Central Bank, has developed an interface of TRAMO-SEATS and X12ARIMA
called DemetraC. In the Bank of Spain and Eurostat websites it is also possible
to find a considerable number of papers relevant to TRAMO-SEATS as well as in
the European Statistical System (ESS) Guidelines. Next sections briefly describe the
supporting methodology of TRAMO-SEATS.

5.1 TRAMO: Time Series Regression with ARIMA Noise,

Missing Observations, and Outliers

TRAMO (Time Series Regression with ARIMA Noise, Missing Observations, and
Outliers) is a regression method that performs the estimation, forecasting, and
interpolation of missing observations and ARIMA errors, in the presence of possibly
several types of outliers. The ARIMA model can be identified automatically or by
the user. Given the vector of observations y D .y1 ; : : : ; yn /0 , the program fits the
regression model

yt D x0t ˇ C t ; (5.1)

where ˇ D .ˇ1 ; : : : ; ˇp /0 is a vector of regression coefficients, xt D .x1t ; : : : ; xpt /0

are p regression variables that define the deterministic part of the model. On the
other hand, t represents the stochastic part of the model, that is assumed to follow
the general ARIMA model

 .B/t D .B/at ; (5.2)

where B is the backshift operator,  .B/ and .B/ are finite polynomials in B, and
at is assumed to be normally identically distributed, that is, NID(0, a2 ).
The polynomial  .B/ contains the unit roots associated with regular and
seasonal differencing, as well as the polynomial with stationary autoregressive roots
(and complex unit roots, if present). .B/ denotes the invertible moving average
polynomial. In TRAMO, these polynomials assume the following multiplicative
5.1 TRAMO: Time Series Regression with ARIMA Noise, Missing. . . 117

form:

 .B/ D p .B/˚P .Bs /.1  B/d .1  Bs /D

D .1 C 1 B C    C p Bp /.1 C ˚1 Bs C    C ˚P BsP /.1  B/d .1  Bs /D ;
.B/ D q .B/Q .Bs /
D .1 C 1 B C    C q Bq /.1 C 1 Bs C    C Q BsQ /;

where s denotes the number of observations per year, e.g., 12 for monthly and 4
for quarterly data. The model may contain a constant c related to the mean of the
differenced series .1  B/d .1  Bs /D yt . In practice, this parameter is estimated as
one of the regression parameters in (5.1). Initial estimates of the parameters can be
input by the user, set to the default values, or computed by the program.
The regression variables xt can be given by the user or generated by the program.
In the latter case, the variables correspond to trading day variations, Easter effects,
and outliers.
Trading day variations are linked to the calendar of each year. Traditionally,
six variables have been used to model the trading day effects. These are: (no. of
Mondays/  .no. of Sundays), . . . , (no. of Saturdays/  .no. of Sundays). The sum
of the effects of each day of the week cancels out. Mathematically, this can be
expressed by the requirement that the trading day coefficients ˇj ; j D 1; : : : ; 7,
P P
verify 7jD1 ˇj D 0, which implies ˇ7 D  6jD1 ˇj . There is the possibility
to consider a more parsimonious modeling of the trading day effects using one
variable instead of six. In this case, the days of the week are first divided into
two categories: working days and nonworking days. Then the variable is defined
as (no. of Mondays, Tuesdays, Wednesdays, Thursdays, and Fridays) – 5/2(no. of
Saturdays and Sundays). Again, the motivation P is that it is desirable that the trading
day coefficients ˇj ; j D 1; : : : ; 7, verify 7jD1 ˇj D 0. Since ˇ1 D ˇ2 D    D ˇ5
and ˇ6 D ˇ7 , we have 5ˇ1 D 2ˇ6 . The set of trading day variables may also
include the leap-year effect. The leap-year variable assumes value 0 for months that
are not February, 0.25 for Februaries with 28 days, and 0.75 for Februaries with 29
days (leap year). Over a 4-year period, this effect cancels out. Trading day effects
typically characterize series reflecting economic activity, and if significant, most
often adding Easter effect variables improves significantly the results.
Easter effects. This variable models a constant change in the level of daily activity
during the d days before Easter, and its typical value ranges between three and
six. The value of d can also be supplied by the user. The variable has zeros for all
months different from March and April. The value assigned to March is equal to
pM  mM , where pM is the proportion of the d days that fall on that month and mM
is the mean value of the proportions of the d days that fall on March over a long
period of time. The value assigned to April is pA  mA , where pA and mA are defined
analogously. Usually, a value of mM D mA D 1=2 is a good approximation. Since
pA  mA D 1  pM  .1  mM / D .pm  mM /, a desirable feature is that the sum
of the effects of both months, March and April, cancels out. If forecast of the series
118 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

is desired, the Easter effect variable has to be extended over the forecasting horizon
(2 years when TRAMO is used with SEATS). Furthermore, given that the holidays
vary among countries, it is strongly recommended that each country builds its own
(set of) holiday regression variables, extended over the forecast period.
Outliers reflect the effect of some special, nonregularly repetitive events, such
as implementation of a new regulation, major political or economic changes,
modifications in the way the variable is measured, occurrence of a strike or natural
disaster, etc. Consequently, discordant observations and various types of abrupt
changes are often present in times series data. The location and type of outliers are
“a priori” unknown. TRAMO uses an improved Chen and Liu [6] type procedure
for outlier detection and correction. The effect of an outlier is modeled by yt D
! .B/It .t0 / C t , where .B/ is a quotient of polynomials in B that models the type
of outlier (its effect over time), It .t0 / is an indicative variable of the occurrence of
the outlier, that is,

1 if t D t0
It .t0 / D
0 otherwise;

whereas ! represents the impact of the outlier at time t0 and t is the outlier free
series which follows the model specified in Eq. (5.2). In the automatic detection and
correction, by default, three types of outliers can be considered:
(1) Additive Outlier (AO): .B/ D 1;
(2) Transitory Change (TC): .B/ D 1=.1  ıB/, where, by default, ı D 0:7; and
(3) Level Shift (LS): .B/ D 1=.1  B/.
One can also include a fourth type of outlier, that is Innovational Outlier (IO) for
which .B/ D .B/=.B/ı.B/ that resembles a shock in the innovations at .
The procedure followed by TRAMO consists of:
1. Pretest for the log/level specification. A trimmed range mean regression test is
performed to select whether the original data will be transformed into log or
maintain the level. The former is selected if the slope is positive, whereas the
latter is chosen if it is negative. When the slope is close to zero, a selection
is made based on the Bayesian Information Criterion (BIC) when applied to
the default model using both specifications. A second test consists of direct
comparison of the BICs of the default model in levels and in logs (with a proper
correction).
2. Pretest for trading days and Easter effects. These pretests are made with
regressions using the default model for the noise and, if the model is subsequently
changed, the test is redone. Thus, the output file of TRAMO may say at the
beginning “Trading day is not significant,” but the final model estimated may
contain trading day variables (or vice versa).
3. Automatic detection and correction of outliers. The program has a facility for
detecting outliers and for removing their effects. The outliers can be entered by
the user or they can be automatically identified by the program, using an original
5.1 TRAMO: Time Series Regression with ARIMA Noise, Missing. . . 119

approach based on those of Tsay [29] and Chen and Liu [6]. The outliers are
identified one by one, as proposed by Tsay [29], and multiple regressions are
used, as in [6], to detect spurious outliers. The procedure used to incorporate or
reject outliers is similar to the stepwise regression procedure for selecting the
“best” regression equation. This results in a more robust procedure than that of
Chen and Liu [6], which uses “backward elimination” and may therefore detect
too many outliers in the first step of the procedure.
4. Automatic model selection. The program further performs an automatic identi-
fication of the ARIMA model. This is done in two steps. The first one yields
the nonstationary polynomial .1  B/d .1  Bs /D of model (5.2). This is done by
iterating on a sequence of AR and ARMA models (with constant) which have a
multiplicative structure when the data is seasonal. The procedure is based on the
results of Tiao and Tsay [27], and Tsay [28]. Regular and seasonal differences
are obtained, up to a maximum order of .1  B/2 .1  Bs /.
The second step identifies an ARMA model for the stationary series (modified
for outliers and regression-type effects) following the Hannan–Rissanen proce-
dure [15], with an improvement which consists of using the Kalman filter to
calculate the first residuals in the computation of the estimator of the variance of
the innovations of model (5.2). For the general multiplicative model

p .B/˚P .Bs /.1  B/d .1  Bs /D yt D q .B/Q .Bs /at ; at  WN.0; a2 /;

the search is made over the range 0  .p; q/  3, and 0  .P; Q/  2. This is
done sequentially (for fixed regular polynomials, the seasonal ones are obtained,
and vice versa), and the final orders of the polynomials are chosen according to
the BIC criterion, with some possible constraints aimed at increasing parsimony
and favoring “balanced” models (similar AR and MA orders).
Finally, the program combines the facilities for automatic identification and
correction of outliers and automatic ARIMA model identification just described
so that it has an option for automatic model identification of a nonstationary
series in the presence of outliers.
5. Estimation of the regARIMA model. The basic methodology is described by
Gomez and Maravall [11], and additional documentation is contained in [8, 9,
13]. Estimation of the regression parameters, including outliers, and the missing
observations among the initial values of the series, plus the ARIMA model
parameters, can be made by concentrating the former out of the likelihood, or
by joint estimation. Several algorithms are available for computing the nonlinear
sum of squares to be minimized. When the differenced series can be used, the
algorithm of Morf et al. [26], with a simplification similar to that of Mélard [25],
is used.
For the nonstationary series, it is possible to use the ordinary Kalman filter
(default option), or its square root version (see [1]). The latter is adequate when
numerical difficulties arise. By default, the exact maximum likelihood method
is employed, and the unconditional and conditional least squares methods are
available as options. Nonlinear maximization of the likelihood function and
120 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

computation of the parameter estimates standard errors are made using the
Marquardt’s method and first numerical derivatives.
Estimation of regression parameters is made by using first the Cholesky
decomposition of the inverse error covariance matrix to transform the regression
equation. In particular, the Kalman filter provides an efficient algorithm to
compute the variables in this transformed regression. Then, the resulting least
squares problem is solved by applying the QR algorithm, where the Householder
orthogonal transformation is used. This procedure yields an efficient and numer-
ically stable method to compute Generalized Least Square (GLS) estimators
of the regression parameters, which avoids matrix inversion. For forecasting,
the ordinary Kalman filter or the square root filter options are available. These
algorithms are applied to the original series. See Gomez and Maravall [10] for
a more detailed discussion on how to build initial conditions on a nonstationary
case.
Missing observations can be handled in two equivalent ways. The first one
is an extension to nonstationary models of the skipping approach by Jones [17].
In this case, interpolation of missing values is made by a simplified Fixed Point
Smoother, and yields identical results to those of Kohn and Ansley [19]. The
second one consists of assigning a tentative value and specifying an additive
outlier to each missing observation. If this option is used, the interpolator is the
difference between the tentative value and the estimated regression parameter. If
the ARIMA parameters are the same, it coincides with the interpolator obtained
with the skipping approach. Also, the likelihood can be corrected so that it
coincides with that of the skipping approach. When concentrating the regression
parameters out of the likelihood, the mean squared errors of the forecasts and
interpolations are obtained following the approach of Kohn and Ansley [18].
When some of the initial missing values cannot be estimated (free parameters),
the program detects them, and flags the forecasts or interpolations that depend on
these free parameters. The user can then assign arbitrary values (typically, very
large or very small) to the free parameters and rerun the program. Proceeding
in this way, all parameters of the ARIMA model can be estimated because the
function to minimize does not depend on the free parameters. Moreover, it will be
evident which forecasts and interpolations are affected by these arbitrary values
because they will strongly deviate from the rest of the estimates. However, if
all unknown parameters are jointly estimated, the program may not flag all free
parameters. It may happen that there is convergence to a valid arbitrary set
of solutions, i.e., some linear combinations of the initial missing observations,
including the free parameters, can be estimated.
In brief, regression parameters are initialized by OLS and the ARMA model
parameters are first estimated with two regressions, as in Hannan and Rissanen
[15]. Next, the Kalman filter and the QR algorithm provide new regression
parameter estimates and regression residuals. For each observation, t-tests are
computed for four types of outliers, namely additive, innovational, level shift,
and transitory. If there are outliers whose absolute t-values are greater than a
5.2 SEATS: Signal Extraction in ARIMA Time Series 121

preselected critical level C, the one with the greatest absolute t-value is selected.
Otherwise, the series is free from outlier effects and the algorithm stops.
If outliers are identified, the series is corrected and the ARMA model
parameters are reestimated. Then, a multiple regression is performed using the
Kalman filter and the QR algorithm. If some outliers have absolute t-values below
the critical value C, the one with the lowest absolute value is removed and the
multiple regression is reestimated. In the next step, using the residuals from the
last multiple regression model, t-tests are calculated for the four types of outliers
and for each observation. If there are outliers with absolute t-values greater than
the critical level C, the one with the greatest absolute t value is chosen and the
algorithm goes on to the estimation of the ARMA model to iterate, otherwise it
stops.
6. Diagnostic checks. The main diagnostic tests are on the residuals. The estimated
residuals aO t are analyzed to test the hypothesis that at are independent and
identically normally distributed with zero mean and constant variance a2 .
Besides inspection of the residual graph, the corresponding sample autocorre-
lation function is examined. The lack of residual autocorrelation is tested using
the Ljung–Box test statistics, and skewness and kurtosis tests are applied to
test for normality of the residuals. Specific out-of-sample forecast tests are also
performed to evaluate if forecasts behave in agreement with the model.
7. Optimal forecasts. If the diagnostics are satisfied, the model is used to compute
optimal forecasts for the series, together with their Mean Square Error (MSE).
These are obtained using the Kalman filter applied to the original series.
TRAMO is to be used as a preadjustment process to eliminate all the determinis-
tic components such that the residual is a linearized series that can be modeled with
an ARIMA process. The latter is decomposed by SEATS in stochastic trend-cycle,
seasonality, and irregulars. Both programs can handle routine applications to a large
number of series and provide a complete model-based solution to the problems of
forecasting, interpolation, and signal extraction for nonstationary time series.

5.2 SEATS: Signal Extraction in ARIMA Time Series

SEATS stands for “Signal Extraction in ARIMA Time Series.” It belongs to the
class of procedures based on ARIMA models for the decomposition of time series
into unobserved components. The method was mainly developed for economic
series, and the corresponding computer program is based on a seasonal adjustment
method developed by Burman [4] at the Bank of England. In particular, the SEATS
procedure consists of the following steps.
1. ARIMA model estimation. SEATS starts by fitting an ARIMA model to a series
not affected by deterministic components, such as trading day variations, moving
holidays, and outliers. Let yt denote this linearized series, and consider an
additive decomposition model (multiplicative if applied to the log transformation
122 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

of yt ), such that

zt D .1  B/d .1  Bs /D yt ; (5.3)

represent the “differenced” series. The model for the differenced linearized series
zt can be written as

p .B/˚P .Bs /.zt  zN/ D q .B/Q .Bs /at ; (5.4)

where zN is the mean of zt ; t D 1; : : : ; n, at is a series of innovations, normally

distributed with zero mean and variance a2 , p .B/˚P .Bs / and q .B/Q .Bs / are
Autoregressive (AR) and Moving Average (MA) polynomials in B, respectively,
which are expressed in multiplicative form as the product of a regular polynomial
in B and a seasonal polynomial in Bs . The complete model can be written in
detailed form as

p .B/˚P .Bs /.1  B/d .1  Bs /D yt D q .B/Q .Bs /at C c; (5.5)

and, in concise form, as

 .B/yt D .B/at C c; (5.6)

where c is equal to  .B/Ny, being yN the mean of the linearized series yt . In words,
the model that SEATS assumes is that of a linear time series with Gaussian
innovations. When used with TRAMO, estimation of the ARIMA model is made
by the exact maximum likelihood method described in Gomez and Maravall
[11]. When used by itself, SEATS applies the quasi-maximum likelihood method
described by Burman [4]. In both cases, a (faster) least squares algorithm is also
available. SEATS contains an option where the automatic model identification
procedure is replaced by a faster one. This simplified procedure starts directly by
fitting the default model and, if found inappropriate, tries a sequence of models
according to the results.
The program starts with the estimation of the ARIMA model. The (inverse)
roots of the AR and MA polynomials are always constrained to remain in or
inside the unit circle. When the modulus of a root converges within a preset
interval around 1 (by default (0.98,1]), the program automatically fixes the root.
If it is an AR root, the modulus is made 1. If it is an MA root, it is fixed to
the lower limit. This simple feature makes the program very robust to over-
and underdifferencing. In the standard case in which SEATS and TRAMO are
used jointly, SEATS controls the AR and MA roots mentioned above, uses the
ARIMA model to filter the linearized series, obtains in this way new residuals,
and produces a detailed analysis of them.
2. Derivation of the ARIMA models for each component. The program proceeds
by decomposing the series that follows model (5.6) into several components.
The decomposition can be multiplicative or additive. Next, we shall discuss the
5.2 SEATS: Signal Extraction in ARIMA Time Series 123

additive model, since the multiplicative relation can be taken care with the log
transformation of the data. That is,

yt D Tt C Ct C St C It ; (5.7)

where Tt denotes the trend component, Ct the cycle, St represents the seasonal
component, and It the irregulars.
The decomposition is done in the frequency domain. The spectrum (or
pseudospectrum) is partitioned into additive spectra, associated with the different
components which are determined, mostly, from the AR roots of the model. The
trend component represents the long-term evolution of the series and displays
a spectral peak at frequency 0, whereas the seasonal component captures the
spectral peaks at seasonal frequencies (e.g., for monthly data these are 0.524,
1.047, 1.571, 2.094, 2.618, and 3.142). The cyclical component captures periodic
fluctuations with period longer than a year, associated with a spectral peak
for a frequency between 0 and (2=s), and short-term variation associated
with low order MA components and AR roots with small moduli. Finally, the
irregular component captures white noise behavior, and hence has a flat spectrum.
The components are determined and fully derived from the structure of the
(aggregate) ARIMA model (5.6) for the linearized series directly identified from
the data. The program is aimed at monthly or quarterly frequency data and the
maximum number of observations that can be processed is 600.
One important assumption is that of orthogonality among the components,
and each one will be described through an ARIMA model. In order to identify
the components, the canonical decomposition is used, such that the variance of
the irregulars is maximized, whereas the trend, seasonal, and cycle are as stable
as possible.
The canonical condition on the trend, seasonal, and cyclical components
identifies a unique decomposition, from which the ARIMA models for the
components are obtained (including the component innovation variances). This
is achieved as follows.
Let the total AR polynomial  .B/ of the ARIMA model (5.6) be factorized as

 .B/ D p .B/˚P .Bs /.1  B/d .1  Bs /D :

The roots of  .B/ are assigned to the unobserved components as follows:

• the roots of .1  B/d D 0 are assigned to trend component;
• the roots of .1  Bs /D D 0 are factorized such that ..1  B/.1 C B C    C
Bs1 //D D 0. In particular, the root of .1  B/ D 0 goes to the trend, whereas
those of .1 C B C    C Bs1 / D 0 go to the seasonal component;
• the positive real roots of p .B/ D 0 are assigned to the trend if they are in
modulus greater than or equal to k, and assigned to cycle if in modulus smaller
than k;
124 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

• the negative real roots of p .B/ D 0 are assigned to the seasonal component
if s ¤ 1, and to the cycle if s D 1;
• the complex roots of p .B/ D 0 are assigned to the seasonal if ! 2 Œ a seasonal
frequency ˙ , being ! the frequency of the root. Otherwise, they are assigned
to the cycle; and
• the positive real roots of ˚P .Bs / D 0, if they are greater than or equal to k
they are assigned to the trend. On the other hand, when they are smaller than
k they are assigned to the cycle.
The parameters k and  are automatically set equal to 0.5 and 0.4, respectively,
but they can be entered by the user as well.
The factorization of  .B/ can be rewritten as

 .B/ D T .B/S .B/C .B/;

where T .B/; S .B/, and C .B/ are the AR polynomials with the trend, seasonal,
and cyclical roots, respectively. Let P and Q denote the orders of the polynomials
 .B/ and .B/ in (5.6).
Considering the case in which P  Q, a polynomial division of the spectrum
(or pseudospectrum) of model (5.6) yields a first decomposition of the type

.B/ .B/
Q
at D a1t C v1 ;
 .B/  .B/

where the order of .B/

Q is min.Q; P  1/, and v1 is a constant (0 if P > Q).
A partial fraction expansion of the spectrum of Œ.B/=
Q .B/ a1t yields the
decomposition

.B/
Q Q T .B/ Q S .B/ Q C .B/
a1t D aQ T C aQ S C aQ C ;
 .B/ T .B/ t s .B/ t C .B/ t

where, given j D T; S; C, we have order(Q j )  order(j ). Let gQ j .!/ denote the

spectrum of [Q j .B/=j .B/]Qajt , with vj D minfQgj .!/ W 0  !  g: Imposing the
canonical condition

gj .!/ D gQ j .!/  vj ; j D T; S; C;
X
v D v1 C vj ;
j

the spectrum of the final components are obtained, which give the models for the
components

T .B/Tt D T .B/aTt
S .B/St D S .B/aSt
5.2 SEATS: Signal Extraction in ARIMA Time Series 125

C .B/Ct D C .B/aCt
It  WN.0; v/:

All components have balanced models, in the sense that the order of the AR
polynomial equals that of the MA one. On the other hand, when Q > P, the
decomposition proceeds as follows. First a decomposition is performed, whereby

ARIMA.P; Q/ D ARIMA.P; P  1/ C MA.Q  P/:

The first component falls under the previous case of P  Q, and hence can be
decomposed in the previous way. In general,

ARIMA.P; P  1/ D Tt C St C Ct C It ;

where Tt , St , Ct , and It denote the trend, seasonal, cyclical, and irregular

component. The MA.Q  P/ component, which represents stationary short-term
deviations, is added to the cyclical component. The series is decomposed then
into a balanced trend model, a balanced seasonal model, a top heavy cycle model,
and a white noise irregular. The first three components are made canonical (i.e.,
noise free).
As a general rule, it is recommended that balanced models be favored because
the excess of MA structure often produces nonadmissible decompositions.
3. Estimation of the components. To perform the series decomposition, SEATS
selects a different ARIMA model from the one estimated by TRAMO if the
latter does not admit an admissible decomposition. In this regard, the forecasts
of the original series done by SEATS differ from the ones obtained by TRAMO.
This can also happen when the ARIMA models are the same. Indeed, whereas
TRAMO uses white noise residuals, SEATS uses residuals obtained by the
filtering of the linearized series. The latter is obtained using an ARIMA
model plus maximum likelihood estimates of the residuals deleted using the
differentiation process, such that the number of residuals equals the sample size.
Finally, the components estimated by SEATS with their forecasts are modified
to reincorporate the deterministic effects that have been previously removed by
TRAMO.
For a particular time series .y1 ; y2 ; : : : ; yn /0 , the program yields the Minimum
Mean Square Error (MMSE) estimators of the components, computed with a
Wiener–Kolmogorov type filter applied to the finite series by extending the latter
with forecasts and backcasts (see [4]). For i D 1; : : : ; n, the estimate yO tjn , equal
to the conditional expectation E.yt jy1 ; : : : ; yn /, is obtained for all components.
When n ! 1, the estimate yO tjn becomes the “final” or “historical” estimate,
which is denoted by yO t . For t D n, the concurrent estimate, yO njn , is obtained,
i.e., the estimate for the last observation of the series. The final and concurrent
estimates are the ones of most interest. When n  k < t < n, yO tjn yields
a preliminary estimator, and, for t > n, a forecast. Besides their estimates,
126 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

the program produces several years of forecasts of the components, as well as

corresponding Standard Errors (SE).
4. Diagnostic checks. For the last two and the next two forecasted years, the SE of
the revision of the preliminary estimates and of the forecast are provided. The
program further computes MMSE estimates of the innovations in each one of
the components. The joint distributions of the components and of their MMSE
estimators are obtained; they are characterized by the variances and auto- and
cross-correlations. The comparison between the theoretical moments for the
MMSE estimators and the corresponding estimates provides additional elements
for diagnosis (see [21]). The program also presents the Wiener–Kolmogorov
filter for each component, and the filter which expresses the weights with which
the different innovations aj in the observed series contribute to the estimate yO tjn .
These weights directly provide the moving average expressions for the revisions.
Next, an analysis of the estimation errors for the trend and for the seasonally
adjusted series (and for the cycle, if present) is performed. Let denote with

dt D yt  yO t ;
dtjn D yt  yO tjn ;
rtjn D yO t  yO tjn :

for t D 1; : : : ; n, the final estimation error, the preliminary estimation error, and
the revision error in the preliminary estimator. The variances and autocorrelation
functions for dt ; dtjn , rtjn are displayed. The program shows how the variance of
the revision error in the concurrent estimator rnjn decreases as more observations
are added, and hence the time it takes in practice to converge to the final
estimator. Similarly, the program computes the decrease in precision as the
forecast moves away from the concurrent estimate. Finally, the SE of the
estimates of the linearized rates of growth for the concurrent estimate of the
rate and its successive revisions are presented, both for the trend and seasonally
adjusted series. All SE computed are exact given that the ARIMA model for the
observed series is correct. Further details can be found in [22, 23].

5.3 Illustrative Example: TRAMO-SEATS Seasonal

Adjustment of the US NODG Series

TRAMO-SEATS is a DOS program, but its Windows version with some modifi-
cations and additions, called TSW, has been developed by Caporello and Maravall
[5]. TRAMO-SEATS is also available, along with X12ARIMA, in the open source
software Gretl and in the commercial software Eviews. The two seasonal adjustment
procedures are also implemented in the Eurostat software Demetra+, recently
updated in JDemetra+ [14], both developed by the National Bank of Belgium.
5.3 Illustrative Example: TRAMO-SEATS Seasonal Adjustment of the US. . . 127

The main feature of Demetra+ and JDemetra+ is to normalize the two different
methods and provide ease of use for national statistics agencies. This is also the aim
declared in the European Statistical System (ESS) Guidelines for Seasonal Adjust-
ment [7] that want to harmonize and provide guidelines for seasonal adjustment.
From the viewpoint of a typical user there is little difference between JDemetra+
and Demetra+. The actual added value of JDemetra+ to the development of the
seasonal adjustment tools lies in the rewriting of the original FORTRAN codes
of X12ARIMA and TRAMO-SEATS in Java. This solution is a crucial factor that
enables the long-term maintenance of the tool, integration of the libraries in the IT
environments of many institutions, and reuse of the modules and algorithms for
other purposes. It runs on operating systems that support the Java VM (Virtual
Machine), such as Microsoft Windows XP SP3/Vista SP1/Windows 7, Ubuntu
9.10, various other Linux distributions, Solaris OS version 11 Express (SPARC and
x86/x64 Platform Edition), and Macintosh OS X 10.6 Intel.
Eurostat uses the TRAMO-SEATS method in the JDemetra+ interface for
seasonal adjustment, and this is the common practice in several European countries.
Hence, in this section, JDemetra+ is applied to perform the seasonal adjustment of
the US New Orders for Durable Goods (NODG) series, observed from February
1992 to December 2013, using the TRAMO-SEATS method. The results of the
seasonal adjustment provided by JDemetra+ are divided into six parts: input, main
results, pre-processing, decomposition, benchmarking, and diagnostics. A detailed
discussion of all the tables, charts, and results produced by the software is not
provided. We focus on the most important outputs for the interpretation of the
seasonal adjustment of the NODG series performed using the TRAMO-SEATS
method.

5.3.1 Input: Specifications

The first part in the JDemetra+ output, called Input, presents information about
the specification used for seasonal adjustment including most important settings.
Several predefined specifications are available, but also user-defined specifications
can be applied. The predefined specifications refer to automatically test whether a
logarithmic transformation is made of the series or just left as is (log/level), to test for
the presence of working days (working days), or trading days effects (trading days),
as well as Easter (Easter) and leap-year effects. Furthermore, an automatic detection
of several types of outliers (AO, LS, and TC) can be performed by the software. For
some predefined specifications, the airline model ARIMA(0,1,1)(0,1,1)s is used as
default, but also there are cases in which JDemetra+ is required to automatically
identify and estimate the best ARIMA model for the series. In detail, the predefined
TRAMO-SEATS specifications available in JDemetra+ are the following:
• RSA0: level, airline model.
128 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

• RSA1: log/level, outliers detection (AO, TC, and LS), airline model.
• RSA2: log/level, working days, Easter, outliers detection, airline model.
• RSA3: log/level, outliers detection, automatic model identification.
• RSA4: log/level, working days, Easter, outliers detection, automatic model
identification.
• RSA5/RSAfull: log/level, trading days, Easter, outliers detection, automatic
model identification.
The default option in JDemetra+ is RSAfull, in agreement with recommendation
of Maravall and Perez [24], who suggested to start with the automatic option that
pretests for trading day and Easter effects, and for the presence of outliers. Hence,
we select this predefined specification in JDemetra+ for the seasonal adjustment of
the NODG series.

5.3.2 Testing for the Presence of Identifiable Seasonality

Before going through the TRAMO-SEATS decomposition, as suggested in

Sect. 4.6.2. for the X12ARIMA seasonal adjustment, it is necessary to check if
the original series presents identifiable seasonality. The Diagnostics part includes
Combined seasonality tests that are useful to check for the presence of seasonality in
the NODG series. The corresponding output is illustrated in Table 5.1. The tests are
the same discussed for the X12ARIMA decomposition, and we remind the reader
to Sect. 4.6.2 for a detail description. The nonparametric Friedman and Kruskal–
Wallis tests confirm the presence of stable seasonality in the series, whereas the test
for the presence of moving seasonality rejects such a hypothesis. It turns out that
the combined test confirms the presence of identifiable seasonality in the NODG
series.

5.3.3 Pre-processing

The Main Results part includes basic information about pre-processing (TRAMO)
and the quality of the decomposition (SEATS). Information about the TRAMO
part of the process includes the estimation span, the number of observations,
transformation (if any), and deterministic effects. On the other hand, the second
part of the Main results summaries the results from the decomposition of the
variance of the white noise innovation for each component extracted by SEATS
from the linearized time series where the regression effects estimated by TRAMO
are removed. The general interest is on a rather stable seasonal component and
hence, when competing models are compared, the preference is given to the model
that minimizes the innovation variance of the seasonal component. For RSA5 and
RSAfull options the trading day and leap-year effects are pretested. If the trading
5.3 Illustrative Example: TRAMO-SEATS Seasonal Adjustment of the US. . . 129

Table 5.1 Combined seasonality test performed on the NODG series

Non parametric tests for stable seasonality

Friedman test

Friedman statistic = 187.9744

Distribution: Chi2(11)
p-value: 0.0000

Stable seasonality present at the 1 % level

Kruskall–Wallis test

Kruskall–Wallis statistic =216.7785474118

Distribution: Chi2(11)
p-value: 0.0000
Stable seasonality present at the 1 % level

Test for the presence of seasonality assuming stability

Sum of squares Degrees of free- Mean square

dom
Between 0.9073464537171304 11.0 0.08248604124701185
months
Residual 0.16251917583422604 251.0 6.474867563116575E-4
Total 1.0698656295513564 262.0 0.004083456601341055

Value: 127.39417515948156
Distribution: F with 11 degrees of freedom in the numerator and 251 degrees of freedom in the
denominator
p-value: 0.0000
Seasonality present at the 1 % level

Evolutive seasonality test

Sum of squares Degrees of free- Mean square

dom
Between 0.00594528505243273 20.0 2.972642526216365E-4
years
Error 0.11407660206322308 220.0 5.1853Q0093782867E-4

Value: 0.573282639857342
Distribution: F with 20 degrees of freedom in the numerator and 220 degrees of freedom in the
denominator
p-value: 0.9283
No evidence of moving seasonality at the 20 % level

Combined seasonality test

Identifiable seasonality present

130 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

day effects have been detected, the message “Trading days effect (6 regressors)” or
“Trading days effect (7 regressors)” is displayed, depending on whether the leap-
year effect has been detected or not. If the Easter effect is statistically significant,
Easter effect detected is displayed. The additional information, i.e., type, location,
and coefficients of every outlier, can be found in the Pre-processing node.
As shown in Table 5.2, it also includes information about the data, such as the
estimation span, the number of observations, the data transformation, the correction
for leap years, and the number of regression variables (calendar-related variables
and outliers). The notation of the estimation span varies according to the frequency
of the observations, in this case monthly, such that [2-1992:12-2013]. Based on
the predefined (default) specification RSAfull, the message “Series has been log-
transformed” is displayed since the logarithmic transformation has been applied
as result of the test performed by the software. Based on the selected predefined
specification, trading days effects, leap-year and the Easter effects are pretested and
estimated if present. It can be noticed that all these effects, except the leap year,
result significant in the series under analysis. Likelihood statistics related to the

Table 5.2 Pre-processing summary information

Summary

Estimation span: [2-1992–12-2013]

263 observations
Series has been log-transformed
Trading days effects (6 variables)
Easter(6) effect detected
3 detected outliers

Final model

Likelihood statistics
Number of effective observations = 250
Number of estimated parameters = 13

Loglikelihood = 460.5213592108896
Transformation adjustment = 3029.8684210569045
Adjusted loglikelihood = 2549.347061846015

Standard error of the regression (ML estimate) = 0.034824441988354465

AIC = 5124.69412369203
AICC = 5126.236496573386
BIC (corrected for length)= 6.449841441444706

Scores at the solution

0.012825 0,001446
5.3 Illustrative Example: TRAMO-SEATS Seasonal Adjustment of the US. . . 131

final model are also shown. The number of effective observations is the number
of observations used to estimate the model, i.e., the number of observations of
regular and seasonal differenced series. The number of estimated parameters is
the sum of the regular and seasonal parameters for both AR and MA, mean effect,
trading/working days effect, outliers, regressors, and standard error of model. In the
pre-processing part, the model is estimated using the exact Maximum Likelihood
Estimation (MLE) method. The standard error of the regression (ML estimate) is
the standard error of the regression from MLE. JDemetra+ displays the maximized
value of the likelihood function, that is used for the selection of the proper ARIMA
model, that is, AIC, AICC, and BIC [14]. The model with the smallest values of
these three criteria is preferred.

5.3.3.1 Regressors

Based on the RSAfull specification, JDemetra+ provides the estimated coefficients

of the trading day variables to evaluate their significant effect on the NODG
series. Table 5.3 illustrates that only the Wednesday variable is characterized by
a coefficient that is significantly different from zero. This implies that the outcome
of the joint F-test indicates that the trading day regressors are jointly significant (the

Table 5.3 Estimates of calendar effects and outliers detection

Regression model
Trading days
Coefficients t-Stat p-value
Monday 0.0043 1.01 0.3128
Tuesday 0.0072 1.69 0.0926
Wednesday 0.0125 2.94 0.0036
Thursday 0.0032 0.76 0.4464
Friday 0.0065 1.51 0.1319
Saturday 0.0034 0.79 0.4309

Joint F-Test = 4.68(0.0002)

Easter [6]
Coefficients t-Stat p-value
0.0246 2.68 0.0078

Outliers
Coefficients t-Stat p-value
LS (1-2009) 0.2379 8.32 0.0000
LS (10-2008) 0.1788 6.26 0.0000
AO (6-2000) 0.1528 5.22 0.0000
132 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

1.010
1.000
0.990

1995 2000 2005 2010

Fig. 5.1 Trading day estimates of the US New Orders for Durable Goods

p-value associated to the F-test statistic is lower than 5 %). Figure 5.1 displays the
monthly trading days component that ranges from 0.988 to 1.016 over the period
February 1992–December 2013.
The Easter effect is also estimated, and the value of the corresponding coefficient
is reported with its standard error, t-statistics, and the related p-value. As shown in
Table 5.3, there is a significant effect lasting over 6 days around Easter. In particular,
Fig. 5.2 displays that the Easter effect is felt 6 days before Easter and on Easter
Sunday but not after Easter.
JDemetra+ presents also the results of the outliers detection. This table includes
the type of outlier, its date, the value of the coefficient, and corresponding t-statistics
and p-values. For the series under study, the software automatically detects the
presence of three outliers: an additive outlier at June 2000, and two level shift
outliers on October 2008 and January 2009, respectively.

5.3.3.2 ARIMA Model

The Arima section shows that, for the log transformed NODG series under study, the
ARIMA(0,1,1)(0,1,1)12 model is chosen and estimated. It should be noticed that we
have reversed the sign of the parameters given by JDemetra+ following the notation
5.3 Illustrative Example: TRAMO-SEATS Seasonal Adjustment of the US. . . 133

1.010
1.005
1.000
0.995
0.990

1995 2000 2005 2010

Fig. 5.2 Moving holiday component of the US NODG

Table 5.4 Estimated values of the ARIMA(0,1,1)(0,1,1)12 model parameters

Arima model
[(0,1,1)(0,1,1)]

Coefficients t-stat p-value

Theta(1) 0.5235 9.47 0.0000
BTheta(1) 0.6940 14.07 0.0000

adopted by Box and Jenkins [3] (see Table 5.4):

.1  B/.1  B12 /Yt D .1  0:5235B/.1  0:6940B12/at ; (5.8)

where at is assumed to follow a white noise process.

All the coefficients result significantly different from zero. Even if this is not the
case, when autoregressive components are estimated, for each regular AR root (i.e.,
the solution of the corresponding characteristic equation) the argument and modulus
are also reported (if present) to inform to which component the regular roots would
be assigned if the SEATS decomposition is used. Rules used by SEATS to assign the
roots to the components have been described in Sect. 5.3, and they are detailed for
the NODG series in the following subsections. The roots with seasonal frequency
belong to the seasonal component while real and positive roots with a high modulus
value are assigned to the trend.
134 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

5.3.3.3 Residuals

The examination of residuals from the fitted model is a crucial step of model
validation. JDemetra+ produces several residual diagnostics and graphs for model
checking. They are presented in the Residuals section and its subcategories.
Specifically, Fig. 5.3 illustrates the residuals from the ARIMA(0,1,1)(0,1,1)12 fitted
to the NODG series, after deterministic effects have been removed.
Their pattern seems to be in agreement with the assumption that they have
been generated by a purely random process with constant mean and variance.
However, to verify this assumption in more detail, several tests are computed on
these residuals, and the results are presented both in summary and detail. Summary
statistics report for each test the corresponding p-value. Bold p-values (ranging from
0.1 to 1) mean “good,” italics (p-value ranging from 0.05 to 0.1) means “uncertain,”
and red (p-value ranging from 0 to 0.05) means “bad” (the latter are not present
in our illustrative example). The four tests considered evaluate the normality, no
correlation, randomness, and linearity of the residuals. Table 5.5 shows that, for the
NODG series, all the tests do not reject the null hypothesis.
The Details sections, not reported here, correspond to the appropriate tables from
the Summary diagnostics. Specifically, detailed information are provided for the
0.10
0.05
0.00
-0.05
-0.10

1995 2000 2005 2010

Fig. 5.3 Residuals from the ARIMA(0,1,1)(0,1,1)12 fitted to the NODG series
5.3 Illustrative Example: TRAMO-SEATS Seasonal Adjustment of the US. . . 135

Table 5.5 Summary statistics for residuals from the regARIMA(0,1,1)(0,1,1)12 model
Analysis of the residuals

Summary

1. Normality of the residuals

p-value
Mean 0.9231
Skewness 0.2516
Kurtosis 0.0948
Normality 0.1151

2. Independence of the residuals

p-value
Ljung-Box(24) 0.3738
Box-Pierce(24) 0.4324
Ljung-Box on seasonality (2) 0.8426
Box-Pierce on seasonality (2) 0.8507

Durbin-Watson statistic: 2.0673

3. Randomness of the residuals

P-value
Runs around the mean: number 0.0992
Runs around the mean: length 1.0000
Up and Down runs: number 1.0000
Up and Down runs: length 1.0000

4. Linearity of the residuals

P-value
Ljung-Box on squared residuals(24) 0.4171
Box-Pierce on squared residuals(24) 0.4742

analysis of the:
1. Distribution of the residuals. For the NODG series, the estimated mean of the
residuals results not significantly different from zero. Furthermore, the normality
assumption is tested in terms of skewness and kurtosis, separately, as well as
jointly using the Doornik–Hansen test which combines both the skewness and
kurtosis tests [14]. In this specific case, the residuals can be considered as
realization of a normal random process.
2. The independence test provided by JDemetra+ is the Ljung–Box test [20].
Denoting with j the sample autocorrelation at rank j of the n residuals, the
136 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

statistics is given by

X
k
j2
LB.k/ D n.n  2/
jD1
.n  j/

which is distributed as a chi-square variable with k  p degrees of freedom if

the residuals are independent. In particular, k depends on the frequency of the
series (k D 24 for monthly series, equal to 8 for quarterly series, and 4s for
other frequencies, where s is a frequency of the time series), and p is the number
of parameters in the ARIMA model with deterministic effects from which the
residuals are derived.
3. Randomness of the residuals. The Wald–Wolfowitz tests, also known as Run
tests, are used to account for the randomness of the residuals. Generally, it
examines the hypothesis that a series of numbers is random. For data centered
around the mean, the test calculates the number and length of runs. A run is
defined as a set of sequential values that are either all above or below the mean.
An up run is a sequence of numbers each of which is above the mean; a down
run is a sequence of numbers each of which is below the mean. The test checks
if the number of up and down runs is distributed equally in time. Both too many
runs and too few runs are unlikely a real random sequence. The null hypothesis
is that the values of the series have been independently drawn from the same
distribution. The test also verifies the hypothesis that the length of runs is random.
In the case under investigation, the randomness assumption is satisfied.
4. The linearity residuals test provides an evidence of autocorrelation in residuals.
In this regard, the Ljung–Box test is performed on the square of the residuals,
and, although not shown, the linearity condition is satisfied by the residuals of
the estimated ARIMA model.

5.3.4 Decomposition

SEATS processes the linearized series obtained with TRAMO. The decomposition
made by SEATS assumes that all the components in the time series, trend, seasonal,
transitory effects (if present), and irregular, are orthogonal and can be expressed by
ARIMA models. The identification of the components requires that only irregular
components include noise.

5.3.4.1 Wiener–Kolmogorov (WK) Analysis

The spectrum of the ARIMA(0,1,1)(0,1,1)12 model estimated by TRAMO is

partitioned into additive spectrums for each component. We recall that the trend
component captures low frequency variations of the series and displays a spectral
5.3 Illustrative Example: TRAMO-SEATS Seasonal Adjustment of the US. . . 137

peak at frequency 0. By contrast, the seasonal component picks up the spectral peaks
at seasonal frequencies and the irregular component captures a white noise behavior.
In some cases, the ARIMA model chosen by TRAMO is changed by SEATS.
This is done when it is nonadmissible for its decomposition. The decomposition is
admissible when the variance of the seasonal component V.aSt / D S2 ranges in a
given interval ŒNS 2 ; QS 2 (see for details Bell and Hillmer [2]), under the assumption
that the series can be decomposed into the sum of a seasonal and nonseasonal
component, that are independent of each other. Furthermore, the decomposition is
defined to be canonical when this variance S2 is exactly equal to NS 2 .
For the NODG series, the ARIMA(0,1,1)(0,1,1)12 model estimated by TRAMO
is kept by SEATS, and, from that, ARIMA representations of the several com-
ponents are derived as described in Sect. 5.3. In the Decomposition part of the
JDemetra+ output, each model is presented in closed form (i.e., using the backshift
operator B) as illustrated in Table 5.6.
Specifically, the theoretical model for the trend results

.1  B/2 Tt D .1 C 0:0299B  0:9701B2/aTt with aTt  WN.0; 0:041/; (5.9)

whereas the model for the seasonal component is

.1 C B C    C B11 /St D 11 .B/aSt ; (5.10)

Table 5.6 The ARIMA models for original series, seasonally adjusted series and components
Model
D: 1.00000  B  B^ 12 + B^ 13
MA: 1.00000  0.523460 B  0.694005 B^ 12 + 0.363284 B^ 13

Sa
D: 1.00000  2.00000 B + B^ 2
MA: 1.00000  1.49745 B + 0.511615 B^ 2
Innovation variance: 0.73560
Trend
D: 1.00000  2.00000 B + B^ 2
MA: 1.00000 + 0.0299450 B  0.970055 B^ 2
Innovation variance: 0.04117
Seasonal
D: 1.00000 + B + B^ 2 + B^ 3 + B^ 4 + B^ 5 + B^ 6 + B^ 7 + B^ 8 + B ^ 9 + B^ 10 + B^ 11
MA: 1.00000 + 1.23274 B + 1.17869 B^ 2 + 1.02437 B^ 3 + 0.790409 B^ 4 + 0.539837 B^ 5
+ 0.294997 B^ 6 + 0.0730214 B^ 7  0.100683 B^ 8  0.252590 B^ 9  0.346686 B^ 10 
0.571634 B^ 11
Innovation variance: 0.02285
Irregular
Innovation variance: 0.41628
138 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

Fig. 5.4 The theoretical spectra of the components and seasonally adjusted function

where aSt  WN.0; 0:023/, and 11 .B/ is a polynomial of order 11 in the backshift
operator B.
Expressions (5.9) and (5.10) theoretically describe the components and at these
equations JDemetra+ refers with the term Components. The corresponding spectra,
as well as the one of the seasonally adjusted series, are presented in the section
WK analysis in the subsection Components. The sum of the components spectra
should be equal to the spectrum for the linearized time series, presented in the Arima
subsection in the Pre-processing part. Figure 5.4 shows the standard spectra for
trend (green), seasonal (blue), and irregular (grey) components defined in Table 5.6.
The seasonally adjusted series spectrum (yellow) is the sum of trend component
spectrum (green) and the irregular component spectrum (grey).
As discussed in Sect. 5.3, the estimation of the components is performed using
symmetric and bi-infinite Wiener–Kolmogorov (WK) filters that define Minimum
Mean Square Error (MMSE) estimators. As example, in the case of the seasonal
component, the estimator is given by
1
X
SO t D WS .B/yt D wS;j Bj yt t D 1; : : : ; n;
jD1
5.3 Illustrative Example: TRAMO-SEATS Seasonal Adjustment of the US. . . 139

where WS .B/ is the Wiener–Kolmogorov filter obtained as

S .B/ S .F/
WS .B/ D S2 yt ;
.B/ .F/

where F D B1 is the forward operator. S .B/ D SS.B/

.B/
, with S .B/ and S .B/ are
the autoregressive and moving average polynomials, respectively, in the ARIMA
model (5.10) defining the seasonal component, whereas .B/ refers to the ARIMA
model (5.8) specified for the linearized NODG series. Replacing in WS .B/ at yt its
ARIMA specification (5.8), and based on the theoretical components specified in
Eqs. (5.9) and (5.10), we obtain
 
O
11 2 S .F/.1  F/2
.1 C B C    C B /St D S S .B/ at : (5.11)
.F/

This equation defines what JDemetra+ indicates as Final estimator, or simply Esti-
mator, for the seasonal component, from which the variance and autocorrelations
of the (theoretical) MMSE estimator can be obtained. Similarly, we can derive the
theoretical estimator for the trend component defined in Eq. (5.9).
The behavior of the WK filters is generally studied in the frequency domain, such
that the spectrum of the seasonal estimator is derived as
1
X fS ./
WS .ei / D wS;j eij D ;
jD1
fY ./

where fS ./ and fY ./ are the spectral densities of the theoretical seasonal compo-
nent and of the original series, respectively. These spectral densities are generally
unknown, but they are approximated based on the ARIMA specifications given in
Eqs. (5.8) and (5.10), respectively. The transfer functions of the WK estimators of
the other components are similarly derived. They are represented graphically in
the subsection Final estimators in the WK analysis part of the JDemetra+ output.
Beyond the spectra of the theoretical estimators, JDemetra+ also provides the
corresponding squared gain functions and WK filter weights, for a maximum of
71 weights. In particular, the spectra of the final estimators are shown in Fig. 5.5,
and they result similar to those of the theoretical components shown in Fig. 5.4,
although the spectra of the final estimators show spectral zeros at the frequencies
where the corresponding theoretical components have spectral peaks. However, the
estimator adapts to the structure of the analyzed series, that is, the spectral holes in
seasonally adjusted series (yellow line) are wider as more stochastic is the seasonal
component.
140 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

Fig. 5.5 Final WK estimators spectra

5.3.4.2 Estimated Components

The Stochastic series part presents the table containing the estimated components
produced by SEATS, that is, the seasonally adjusted series, the trend, the seasonal
component, and the irregular component. The seasonally adjusted NODG series is
shown in Fig. 5.6 together with the original series.
The estimated components, denoted by JDemetra+ as Estimates, are derived by
applying the WK filter to the observations extended at both ends with forecasts
and backcasts from the ARIMA model specified for the linearized NODG series.
Alternatively, the spectrum of the component is estimated by applying the spectrum
(transfer function) of the filter to the spectral density of the ARIMA model (5.8) [2].

5.3.4.3 Diagnostic and Quality Assessment

The Model based tests part analyzes the theoretical distribution of estimators making
use of the estimated components. Specifically, the tests concern the variance and the
autocorrelation function of each estimator, as well as the cross-correlation among
them. Hence, the output is organized in three main sections.
5.3 Illustrative Example: TRAMO-SEATS Seasonal Adjustment of the US. . . 141

NODG

Seasonally adjusted series

200000
100000

1995 2000 2005 2010

Fig. 5.6 The US NODG original and seasonally adjusted series

Table 5.7 Cross-correlation between the components estimators and the components sample
estimates
Cross-correlation
Estimator Estimate p-value
Trend/seasonal 0.1011 0.1708 0.3719
Trend/irregular 0.2041 0.1173 0.2919
Seasona/irregular 0.0513 0.1131 0.2324

The most important one is on the Cross-correlation function. The decom-

position made by SEATS assumes that the components are orthogonal. To test
this assumption, Table 5.7 contains the cross-correlations between pairs of final
components estimators (denoted by “Estimator”) that are compared with the cross-
correlations computed on the estimated components (indicated as “Estimate”).
The cross-correlations among the theoretical components are not reported since,
by assumption, they are uncorrelated, whereas the MMSE estimators could be
correlated. The p-values in Table 5.7 are related to the test statistics used to analyze
if there is a statistically significant difference among the expected cross-correlations
between the final estimators and the cross-correlations computed between the
estimated components. It is evident that, for the NODG series, there is no significant
difference between them.
142 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

Table 5.8 Variance of the components, of their theoretical WK estimators and of empirical
estimates
Variance

Component Estimator Estimate p-value

Trend 0.0799 0.0096 0.0074 0.0655
Seasonally adjusted 2.5776 2.1820 2.0913 0.7398
Seasonal 0.1482 0.0162 0.0080 0.1238
Irregular 0.4163 0.2686 0.2458 0.3849

SEATS identifies the components assuming that, except for the irregular, they
are clean of noise. This implies that the variance of the irregular is maximized in
contrast to that of the trend-cycle and seasonal components which are as stable
as possible. In this regard, JDemetra+ compares the variance expected for the
theoretical estimator with the variance empirically computed on the corresponding
estimated component. In the NODG example, the variance of the WK estimator of
the seasonal component is derived using its theoretical specification (5.11). For all
the components, Table 5.8 shows that, for the NODG series, there are no statistically
significant differences between the variances of the theoretical WK estimators and
their estimated values.
Finally, for each component, JDemetra+ exhibits autocorrelations corresponding
to the ARIMA model selected for each component, the corresponding WK estima-
tors and the estimated components. They are calculated from the first lag up to the
seasonal lag.
For each table, the user should check whether the empirical estimates agree with
the model, i.e., if their ACFs are close to those of the WK estimator. Special attention
should be given to first and/or seasonal order autocorrelations. The autocorrelation
functions for the trend and seasonal components of the NODG series are provided
in Tables 5.9 and 5.10, respectively. It can be noticed that for both components there
is evidence of difference in the theoretical and estimated autocorrelations in the first
lags, but not at the seasonal lag.
It has to be noticed that the coefficients of the autocorrelation function of the
irregular component are always null (“Component” column) as shown in Table 5.11
since the theoretical model for irregular component is always a white noise. As
outlined by the p-values column, the estimated irregular component can be thought
as realization of an uncorrelated process.
5.3 Illustrative Example: TRAMO-SEATS Seasonal Adjustment of the US. . . 143

Table 5.9 Autocorrelations of the stationary transformation of trend, the estimators and the
sample estimates
Trend
Lag Component Estimator Estimate p-value
1 0.0005 0.4622 0.3619 0.0078
2 0.4995 0.2542 0,3984 0.0261
3 0.0000 0.3292 0.3102 0.7656
4 0.0000 0.1708 0.0062 0.0525
5 0.0000 0.0865 0.1200 0.0216
6 0.0000 0.0399 0.1204 0.0397
7 0.0000 0.0108 0.0038 0.8717
8 0.0000 0.0137 0.0983 0.2171
9 0.0000 0.0440 0.0451 0.3054
10 0.0000 0.0357 0.0515 0.8594
11 0.0000 0.0723 0.0125 0.4977
12 0.0000 0.1536 0.1440 0.9009

Table 5.10 Autocorrelations of the stationary transformation of the seasonal component, the
estimators and the sample estimates
Seasonal
Lag Component Estimator Estimate p-value
1 0.8654 0.7040 0.5605 0.0188
2 0.6674 0.3311 0.1542 0. 0216
3 0.4467 0.0428 0.1614 0.0374
4 0.2340 0.3588 0.3517 0.8991
5 0.0510 0.5750 0.3705 0.0491
6 0.0886 0.6661 0.4018 0.0452
7 0.1781 0.6233 0.3492 0.0073
8 0.2168 0.4537 0.3495 0.1107
9 0.2087 0.1798 0.1658 0.8765
10 0.1621 0.1604 0.1012 0.6234
11 0.0881 0.5139 0.4329 0.4704
12 0.0000 0.8129 0.7992 0.8198
144 5 Seasonal Adjustment Based on ARIMA Model Decomposition: TRAMO-SEATS

Table 5.11 Autocorrelations of the stationary transformation of the irregular component, the
estimators and the sample estimates
Irregular
Lag Component Estimator Estimate p-value
1 0.0000 0.2382 0.2980 0.2870
2 0.0000 0.1246 0.1918 0.3210
3 0.0000 0.0651 0.0111 0.2715
4 0.0000 0.0338 0.0292 0.9476
5 0.0000 0.0171 0.0672 0.2300
6 0.0000 0.0079 0.0110 0.7568
7 0.0000 0.0022 0.0511 0.4486
8 0.0000 0.0027 0.1111 0.1004
9 0.0000 0.0086 0.0321 0.7347
10 0.0000 0.0184 0.0040 0.7465
11 0.0000 0.0361 0.0732 0.5936
12 0.0000 0.1532 0.1196 0.5818

References

1. Anderson, B., & Moore, J. (1979). Optimal filtering. New Jersey: Prentice Hall.
2. Bell, W. R., & Hillmer, S. C. (1984). Issues involved with the seasonal adjustment of economic
time series. Journal of Business and Economic Statistics, 2, 291–320.
3. Box, G. E. P., & Jenkins, G. M. (1970). Time series analysis: Forecasting and control. San
Francisco: Holden-Day (Second Edition 1976).
4. Burman, J. P. (1980). Seasonal adjustment by signal extraction. Journal of the Royal Statistical
Society A, 143, 321–337.
5. Caporello, G., Maravall, A., & Sanchez, F. J. (2001). Program TSW reference manual. Working
Paper 0112, Servicio de Estudios, Banco de España.
6. Chen, C., & Liu, L. M. (1993). Joint estimation of model parameters and outlier effects in time
series. Journal of the American Statistical Association, 88, 284–297.
7. Eurostat. (2009). ESS guidelines on seasonal adjustment. Luxembourg: Office for Official
Publications of the European Communities.
8. Gomez, V. (1997). Automatic model identification in the presence of missing observations
and outliers. Mimeo. Ministerio de Economia y Hacienda, Direccion General de Analisis y
Programacion Presupuestaria, June 1997
9. Gomez, V., & Maravall. A. (1992). Time series regression with ARIMA noise and missing
observations - Program TRAM. EUI Working Paper Eco No. 92/81, Department of Economics,
European University Institute.
10. Gomez, V., & Maravall, A. (1993). Initializing the Kalman filter with incompletely specified
initial conditions. In G. R. Chen (Ed.), Approximate Kalman filtering. Series on approximation
and decomposition. London: World Scientific Publishing Company.
11. Gomez, V., & Maravall, A. (1994). Estimation, prediction and interpolation for nonstationary
series with the Kalman filter. Journal of the American Statistical Association, 89, 611–624.
12. Gomez, V., & Maravall, A. (1996). Program TRAMO and SEATS: Instructions for users.
Working Paper 9628, Service de Estudios, Banco de Espana.
13. Gomez, V., Maravall, A., & Pena, D. (1999). Missing observations in ARIMA models:
Skipping strategy versus additive outlier approach. Journal of Econometrics, 88, 341–364.
14. Grudkowska, S. (2015). JDemetra+ reference manual. Warsaw: National Bank of Poland.
References 145

15. Hannan, E. J., & Rissanen, J. (1982). Recursive estimation of mixed autoregressive moving
average order. Biometrika, 69, 81–94.
16. Hillmer, S. C., & Tiao, G. C. (1982). An ARIMA-model based approach to seasonal
adjustment. Journal of the American Statistical Association, 77, 63–70.
17. Jones, R. (1980). Maximum likelihood fitting of ARMA models to time series with missing
observations. Technometrics, 22, 389–395.
18. Kohn, R., & Ansley, C. F. (1985). Efficient estimation and prediction in time series regression
models. Biometrika, 72, 694–697.
19. Kohn, R., & Ansley, C. F. (1986). Estimation, prediction and interpolation for ARIMA models
with missing data. Journal of the American Statistical Association, 81, 751–761.
20. Ljung, G., & Box, G. (1979). The likelihood function of stationary autoregressive-moving
average models. Biometrika, 66, 265–270.
21. Maravall, A. (1987). On minimum mean squared error estimation of the noise in unobserved
component models. Journal of Business and Economic Statistics, 5, 115–120.
22. Maravall, A. (1988). The use of ARIMA models in unobserved components estimation.
In W. Barnett, E. Berndt, & H. White (Eds.), Dynamic econometric modeling. Cambridge:
Cambridge University Press.
23. Maravall, A. (1995). Unobserved components in economic time series. In H. Pesaran, P.
Schmidt, & M. Wickens (Eds.), The handbook of applied econometrics (Vol. 1). Oxford: Basil
Blackwell.
24. Maravall, A., & Pérez, D. (2012). Applying and interpreting model-based seasonal adjustment
- The Euro-area industrial production series. In W. R. Bell, S. H. Holan, & T. S. McElroy
(Eds.), Economic time series: Modeling and seasonality. London: Chapman and Hall.
25. Melard, G. (1984). A fast algorithm for the exact likelihood of autoregressive moving average
models. Applied Statistics, 35, 104–114.
26. Morf, M., Sidhu, G. S., & Kailath, T. (1974). Some new algorithms for recursive estimation on
constant, linear, discrete-time systems. IEEE Transactions on Automatic Control, 19, 315–323.
27. Tiao, G. C., & Tsay, R. S. (1983). Consistency properties of least squares estimates of
autoregressive parameters in ARMA Models. The Annals of Statistics, 11, 856–871.
28. Tsay, R. S. (1984). Regression models with time series errors. Journal of the American
Statistical Association, 79, 118–124.
29. Tsay, R. S. (1986). Time series model specification in the presence of outliers. Journal of the
American Statistical Association, 81, 132–141.
Chapter 6
Seasonal Adjustment Based on Structural Time
Series Models

Abstract The structural model decomposition method starts directly with an

observation equation (sometimes called measurement equation) that relates the
observed time series to the unobserved components. Simple ARIMA or stochastic
trigonometric models are a priori assumed for each unobserved component. Struc-
tural Time series Analyzer, Modeler, and Predictor (STAMP) is the main software
and includes several types of models for each component. This chapter discusses in
detail the basic structural time series model with explicit specifications for each
component. It deals also with the estimation of the parameters which is carried
out by the method of maximum likelihood where the maximization is done by
means of a numerical optimization method. Based on the parameter estimates, the
components can be estimated using the observed time series. Model adequacy is
generally diagnosed using classical test statistics applied to the standardized one-
step ahead prediction errors. An illustrative example of the seasonal adjustment
performed using the default option of the STAMP software is shown with the US
Unemployment Rate for Males (16 years and over) series.

The structural time series approach involves decomposing a series into components
which have a direct interpretation. A structural model, also called unobserved
component (UC) model, consists of a number of stochastic linear processes that
stand for the trend, cycle, seasonality, and remaining stationary dynamic features in
an observed time series. The trend component typically represents the longer term
movement of the series and is often specified as a smooth function of time. The
recurring but persistently changing patterns within the years are captured by the
seasonal component. In economic time series, the cycle component can represent
the dynamic features associated with the business cycle.
The key to handling structural time series models is the state space form with
the state of the system representing the various unobserved components such
as trend, cycle, and seasonality. The estimate of the unobservable state can be
updated by means of a filtering procedure as new observations become available.
Predictions are made by extrapolating these estimated components into the future,
while smoothing algorithms give the best estimate of the state at any point within
the sample. The statistical treatment can therefore be based on the Kalman filter
and its related methods. A detailed discussion of the methodological and technical

© Springer International Publishing Switzerland 2016 147

E. Bee Dagum, S. Bianconcini, Seasonal Adjustment Methods and Real Time
Trend-Cycle Estimation, Statistics for Social and Behavioral Sciences,
DOI 10.1007/978-3-319-31822-6_6
148 6 Seasonal Adjustment Based on Structural Time Series Models

concepts underlying structural time series models is contained in the monographs

by Harvey [11], Kitagawa and Gersh [17], and Kim and Nelson [16]. Durbin and
Koopman [8] also provide an exhaustive overview on state space methods for time
series, whereas an introduction is given by Commandeur and Koopman [5].
In this chapter, we first introduce the basic structural time series model with
explicit specifications for each component. The estimation of parameters is carried
out by the method of maximum likelihood in which the likelihood is evaluated via
the Kalman filter. The likelihood is maximized by means of a numerical optimiza-
tion method. Based on the parameter estimates, the components can be estimated
using the observed time series. Model adequacy is generally diagnosed using
classical test statistics applied to the standardized one-step ahead prediction errors.

6.1 Structural Time Series Models

The structural time series model for quarterly or monthly observations yt ; t D

1; : : : ; n, is given by

yt D Tt C Ct C St C It ; (6.1)

where Tt stands for the trend, Ct for the cycle, St for seasonality, and It is the irregular
component generally assumed to be NID.0; I2 /. All four components are stochastic
and the disturbances driving them are assumed to be mutually uncorrelated. The
definitions of the components are given below, but a full explanation of the
underlying rationale can be found in Harvey [11]. The effectiveness of structural
time series models compared to ARIMA-type models is discussed in Harvey et al.
[12].

6.1.1 Trend Models

The trend component can be specified in many different ways. The most common
specifications are listed below.
Local Level Model The trend component can be simply modeled as a random walk
process, that is

TtC1 D Tt C aTt ; (6.2)

where the disturbance series aTt is normally independently distributed with mean
zero and variance T2 , that is NID.0; T2 /. It is also mutually independent on all other
disturbance series related to yt in (6.1). The initial trend T1 is generally treated as an
unknown coefficient that has to be estimated together with the variance T2 .
6.1 Structural Time Series Models 149

In the specification (6.2), the trend component is an integrated process of order 1,

I.1/. When this trend is included in the decomposition of yt , the time series yt is at
least I.1/ as well. In particular, when Tt is given by Eq. (6.2), the model yt D Tt C It
is called local level model.
Local Linear Trend An extension of the random walk trend is obtained by
including a stochastic drift component

TtC1 D Tt C ˇt C aTt (6.3)

ˇtC1 D ˇt C
t
t  NID.0; 
2 /; (6.4)

where the disturbance series aTt is as in (6.2). The initial values T1 and ˇ1 are treated
as unknown coefficients. When Tt is given by (6.3), Harvey [10] defines yt D Tt C It
as the local linear trend model. In case 
2 D 0, the trend (6.3) reduces to an I.1/
process given by TtC1 D Tt C ˇ1 C aTt , where the drift ˇ1 is fixed. This specification
is referred as a random walk plus drift process. If in addition T2 D 0, the trend
reduces to the deterministic linear trend TtC1 D T1 Cˇ1 t. When T2 D 0 and 
2 > 0,
the trend Tt in (6.3) remains an I.2/ process.
Trend with Stationary Drift An extension of the previous model is obtained by
including a stationary stochastic drift component:

TtC1 D Tt C ˇt C aTt ; (6.5)

ˇtC1 D 'ˇ ˇt C
t ;

with the autoregressive coefficient 0 < 'ˇ < 1, and where the disturbance series
aTt and
t are as in Eqs. (6.3) and (6.4). The restriction for 'ˇ is necessary to have a
stationary process for the drift ˇt . In this case, the initial variable T1 is treated as an
unknown coefficient, while the initial drift is specified as ˇ1  N.0; 
2 =1'ˇ2 /. The
stationary drift process for ˇt can be generalized to a higher order autoregressive
process and can include moving average terms. However, in practice it may be
difficult to empirically identify such drift processes without very large data samples.
Higher Order Smooth Trend The local linear trend (6.3) with T2 D 0 is a smooth
I.2/ process. It can be alternatively specified as 2 TtC2 D
t , where the initial
variables T1 and T2 are treated as unknown coefficients. To enforce more smoothness
in the trend component, it can be generalized as k TtCk D
t , where the initial
variables T1 ; : : : ; Tk are treated as unknown coefficients for k D 1; 2; : : :. In the
.k/
usual way, we can specify the higher order smooth trend component by Tt D Tt ,
where
. j/ . j/ . j1/ .0/
TtC1 D Tt C Tt ; Tt D
t ; (6.6)

for j D k; k  1; : : : ; 1, and where the disturbance series

t is as in (6.4). In case
.2/ .1/
k D 2, we obtain the trend model (6.3) with T2 D 0, where Tt D Tt and ˇt D Tt .
This trend specification is discussed in more detail by Gomez [9].
150 6 Seasonal Adjustment Based on Structural Time Series Models

Trend with Smooth Stationary Drift Although the smoothness of a trend is a

desirable feature for many economic time series, the fact that the smooth trend is
defined by an I.k/ process can be less convincing. Therefore, Koopman and Ooms
[20] propose a smooth I.1/ trend given by
.m/
TtC1 D Tt C ˇt ; (6.7)
. j/ . j/ . j1/ .0/
ˇtC1 D 'ˇ ˇt C ˇt ; ˇt D
t ;

for j D m; m1; : : : ; 1, and where the disturbance series

t is as in (6.3). In case m D
.1/
1, the trend with stationary drift model (6.5) is derived with T2 D 0 and ˇt D ˇt .
. j/
The autoregressive coefficient 0 < 'ˇ < 1 is the same for each ˇtC1 with j D m; m
1; : : : ; 1. This restriction can be lifted by having different autoregressive coefficients
for each j, but generally the parsimonious specification (6.7) is preferred.

6.1.2 The Cyclical Component

To capture the business cycle features of a time series, various stochastic specifica-
tions of the cycle component can be considered.
Autoregressive Moving Average Process The cycle component Ct can be formu-
lated as a stationary autoregressive moving average (ARMA) process as

'C .B/CtC1 D C .B/aCt aCt  NID.0; C2 /; (6.8)

where 'C .B/ is the autoregressive polynomial of order p in the lag operator B with
coefficients 'C;1 ; : : : ; 'C;p and C .B/ is the moving average polynomial of order
q with coefficients C;1 ; : : : ; C;q . The requirement of stationarity applies to the
autoregressive polynomial 'C .B/, and states that the roots j'C .B/j D 0 lie outside
the unit circle. The theoretical autocorrelation function of an ARMA process has
cyclical properties when the roots of j'C .B/j D 0 are within the complex range. It
requires p > 1. In this case, the autocorrelations converge to zero with increasing
lags, but the convergence pattern is cyclical. It implies that the time series itself
has cyclical dynamic properties. Once the autoregressive coefficients are estimated,
it can be established whether the empirical model with Ct as in (6.8) has detected
cyclical dynamics in the time series.
Time-Varying Trigonometric Cycle Another stochastic formulation of the cycle
component is based on a time-varying trigonometric process as follows:
      
CtC1 cos c sin c Ct a Ct
 D C t D 1; : : : ; n; (6.9)
CtC1  sin c cos c Ct aCt

with frequency c associated with the typical length of a business cycle, say between
1.5 and 8 years according to Burns and Mitchell [4]. The factor 0 <  < 1 is
6.1 Structural Time Series Models 151

introduced to enforce a stationary process for the stochastic cycle component. The
disturbance and the initial conditions for the cycle variables are given by
     
a Ct
C1 2
 NID 0; C2 I2  NID 0; C2 I2 (6.10)
aCt C1 

where the cyclical disturbance series aCt and aCt are serially and mutually inde-
pendent, also with respect to all other disturbance series. The coefficients ; c ,
and C2 are unknown and need to be estimated together with the other parameters.
This stochastic cycle specification is discussed by Harvey [11], that shows that the
process (6.9) to be the same as and ARMA process (6.8) with complex roots where
p D 2 and q D 1.
Smooth Time-Varying Trigonometric Cycle To enforce smoothness on the cycle
.m/
component, Harvey and Trimbur [13] propose the specification Ct D Ct , where
!   ! !
. j/ . j/ . j1/
CtC1 cos c sin c Ct Ct
. j/ D . j/ C . j1/ ; (6.11)
CtC1  sin c cos c Ct Ct

for j D m; m  1; : : : ; 1, and
!  
.0/
CtC1 a Ct
.0/ D  NID.0; C2 I2 /; (6.12)
CtC1 aCt

for t D 1; : : : ; n. The initial conditions for this stationary process are provided by
Trimbur [27].

6.1.3 Seasonality

As discussed in Chap. 2, to account for the seasonal variation in a time series, the
component St is included in model (6.1). Common specifications for the seasonal
component St are provided in the following:
Dummy Seasonal Since the seasonal effects should sum to zero over a year, a basic
model for this component is given by

X
s1
St D  Stj C !t t D s; : : : ; n; (6.13)
jD1

where s denotes the number of “seasons” in a year. In words, the seasonal effects are
allowed to change over time by letting their sum over the previous year be equal to a
random disturbance term !t with mean zero and variance !2 . Writing out Eq. (6.13)
152 6 Seasonal Adjustment Based on Structural Time Series Models

in terms of the lag operator B gives

.1 C B C    C Bs1 /St D .B/St D !t t D s; : : : ; n: (6.14)

However, since .1  Bs / D .1 C B C    C Bs1 /.1  B/ D .B/.1  B/ the

model can also be expressed in terms of the seasonal difference operator as .1 
Bs /St D .1  B/!t . The normally distributed disturbance !t drives the changes in
the seasonal effect over time and is serially and mutually uncorrelated with all other
disturbances and for all time periods. In the limiting case where !2 D 0 for all t,
the seasonal effects are fixed over time and are specified as a set of unknown fixed
dummy coefficients that sum up to zero.
Trigonometric Seasonal Alternatively, a seasonal pattern can also be modeled by a
set of trigonometric terms at the seasonal frequencies, j D 2j=s, j D 1; : : : ; Œs=2 ,
where Œs=2 is equal to s=2 if s is even, and .s  1/=2 if s is odd. The seasonal effect
at time t is then described as
Œs=2
X
St D .Sj cos j t C Sj sin j t/: (6.15)
jD1

When s is even, the sine term disappears for j D s=2, and so the number of
trigonometric parameters, the Sj and Sj , is always .s  1/=2, which is the same as
the number of coefficients in the seasonal dummy formulation. A seasonal pattern
based on (6.15) is the sum of Œs=2 cyclical components, each with  D 1, and it may
be allowed to evolve over time in exactly the same way as the cycle was allowed to
move in Eq. (6.10). The model is

Œs=2
X
St D Sj;t ; (6.16)
jD1

where, following Eq. (6.10),

!   ! !
Sj;t cos j sin j Sj;t1 !j;t
 D  C  ; (6.17)
Sj;t  sin j cos j Sj;t1 !j;t


with !j;t and !j;t ; j D 1; : : : ; Œs=2 , being white noise processes which are uncorre-
lated with each other and have a common variance !2 . Sj;t 
appears as a matter of
construction, and its interpretation is not particularly important. Note that, when s
is even, the component at j D s=2 collapses to

Sj;t D Sj;t1 cos j C !j;t :

6.2 Linear State Space Models 153

Other seasonal models that incorporate time-varying seasonal effects in the

unobserved components time series model can be considered. We refer the reader
to Proietti [25] for a review of other different seasonal specifications and their
properties.

6.1.4 Regression Component

The model (6.1) may provide a good description of the time series, although it
may sometimes be necessary to include additional components. As discussed in the
previous chapters, seasonal economic time series is often affected by trading day
effects and holiday effects which can influence the dynamic behavior of the series.
Hence, a set of explanatory variables need to be included in the model to capture
specific (dynamic) variations in the time series, as well as outliers and breaks.
Therefore, Koopman and Ooms [20] suggest to extend model (6.1) as follows:

yt D Tt C Ct C St C x0t ı C It ; It  N.0; I2 /;

for t D 1; : : : ; n, and where xt is a K-dimensional vector of predetermined covariates

and ı is a K  1 vector of regression coefficients, that can be allowed to change over
time.

6.2 Linear State Space Models

The statistical treatment of structural time series models is based on the correspond-
ing state space representation, according to which the observations are assumed to
depend linearly on a state vector that is unobserved and generated by a stochastic
time-varying process. The observations are further assumed to be subject to a
measurement error that is independent on the state vector. The state vector can be
estimated or identified once a sufficient set of observations becomes available.
The state space form provides a unified representation of a wide range of linear
time series models, as discussed by Harvey [11], Kitagawa and Gersch [17], and
Durbin and Koopman [8].
The general linear state space model for a sequence of n observations, y1 ; : : : ; yn ,
is specified as follows:

yt D zTt ˛t C It ; t D 1; : : : ; n (6.18)
˛tC1 D  t ˛t C Rt "t : (6.19)

Equation (6.18) is called the observation or measurement equation which relates

the observations yt ; t D 1; : : : ; n, to the state vector ˛t through zt , that is, an m  1
154 6 Seasonal Adjustment Based on Structural Time Series Models

vector of fixed coefficients. In particular, ˛t is the m  1 state vector that contains

the unobserved trend Tt , cycle Ct , and the seasonal component St . The irregular
component It is generally assumed to follow a white noise process with zero mean
and variance I2t .
On the other hand, Eq. (6.19) is called the state or transition equation, where the
dynamic evolution of the state vector ˛t is described through the fixed matrix  t
of order m  m. "t is a r  1 vector of disturbances which are assumed to follow a
multivariate white noise process with zero mean vector and covariance matrix ˙ "t .
It is assumed to be distributed independently on It at all time points. The matrix Rt
is an m  r selection matrix with r < m, that in many standard cases is the identity
matrix Im , being generally r D m. Indeed, although matrix Rt can be specified
freely, it is often composed of a selection from the first r columns of the identity
matrix Im .
Initial conditions have to be specified for the state vector at the first time
point, ˛1 . It is generally assumed to be generated as ˛1  N.a1 ; P1 / , and to be
independent on the observation and state disturbances, It and "t . The mean vector a1
and covariance matrix P1 can be treated as given and known in almost all stationary
processes. For nonstationary processes and in presence of regression effects in the
state vector, the associated elements in the initial mean vector a1 can be treated as
unknown and estimated. For an extensive discussion of initialization in state space
analysis, we refer the reader to Durbin and Koopman [8].
By appropriate choices of ˛t ; It , and "t , of the matrices zt ;  t , Rt , and of the
scalar I2t , a wide range of different structural time series models discussed in
Sect. 6.2 can be specified from Eqs. (6.18) and (6.19).
Basic Decomposition Model Consider the model yt D Tt C St C It with the trend
component specified through Eqs. (6.3) and (6.4), and St as in (6.13) in presence of
quarterly data (s D 4). A state vector of five elements and a disturbance vector of
four elements are required, and they are given by
0 1
Tt
Bˇ C 0 1
B t C aTt
B C
˛t D B S1;t C ; "t D @
t A :
B C
@ S2;t A !t
S3;t

The state space formulation of the basic decomposition model is given by

Eqs. (6.18) and (6.19) with the system matrices
0 1 0 1 0 1
10 0 0 0 1 0 1 100
B0 1 0 0 0 C B0C 2 0 0 B0 1 0C
B C B C B C
B C B C B T 2 C B C
 t D B0 0 1 1 1 C ; zt D B 1 C ; ˙ "t D @ 0 
0 A ; Rt D B 0 0 1 C :
B C B C B C
@0 0 1 0 0 A @0A 0 0 !2 @0 0 0A
00 0 1 0 0 000
6.2 Linear State Space Models 155

Here the system matrices  t and zt do not depend on time. The variances of the
disturbances are unknown and need to be estimated. They are I2 ; T2 ; 
2 , and !2 .
For the trend component, the initial variables T1 and ˇ1 are treated as unknown,
whereas for the seasonal component, with s D 4, the unknown coefficients are the
initial variables S1 ; S0 ; S1 . All these coefficients form part of the initial state vector
˛1 . This is the basic structural time series model. A typical application of this basic
model is for the seasonal adjustment of time series, being the seasonally adjusted
series given by yt  St D Tt C It for t D 1; : : : ; n.
Local Trend Plus Cycle Model Another model is given by yt D Tt C Ct C It with
Tt specified as in Eq. (6.2) and Ct as in Eq. (6.9). Defining
0 1 0 1 0 1
Tt aTt 1 0 0
˛t D @ Ct A ; "t D @ aCt A ;  t D @ 0  cos.c /  sin.c / A ; zTt D .1; 1; 0/;
Ct aCt 0  sin.c /  cos.c /
0 2 1 0 1
 0 0 100
˙ "t D @ 0 C2 .1  2 / 0 A ; Rt D @ 0 1 0 A
2 2 001
0 0 C .1   /

in (6.18) and (6.19), we obtain the following local level plus cycle model:

yt D Tt C Ct C It ; It  N.0; I2 /
TtC1 D Tt C t ; t  N.0; 2 /
CtC1 D Œcos.c /Ct C sin.c /Ct C aCt ; aCt  NID.0; C2 .1  2 //

CtC1 D Πsin.c /Ct C cos.c /Ct C aCt ; aCt  NID.0; C2 .1  2 //

for t D 1; : : : ; n, where 0 <  < 1 is the damping factor and c is the frequency of
the cycle in radians so that 2=c is the period of the cycle. In case  D 1, the cycle
reduces to a fixed sine–cosine wave, but the component is still stochastic since the
initial values C1 and C1 are stochastic variables with mean zero and variance C2 . A
typical application of this model is for the signal extraction of business cycles from
macroeconomic time series.

6.2.1 The Kalman Filter

Given the values of all system matrices and initial conditions a1 and P1 , the
predictive estimator of the state vector ˛tC1 is based on observations y1 ; : : : ; yn . The
Kalman filter [15] computes the minimum mean square linear estimator (MMSLE)
of the state vector ˛tC1 conditional on the observations y1 ; : : : ; yt , denoted by atC1jt ,
together with its mean square error (MSE) matrix, denoted by PtC1jt . atC1jt is
156 6 Seasonal Adjustment Based on Structural Time Series Models

generally called the state prediction estimate, whereas PtC1jt is its state prediction
error variance matrix. Specifically, the one-step ahead prediction error vt D yt 
E.yt jy1 ; : : : ; yt1 / and its variance Var.vt / D ft are derived as

vt D yt  zTt atjt1 ; ft D zTt Pt zt C I2 ;

such that the innovations have mean zero and are serially independent by construc-
tion, that is, E.vt vs0 / D 0 for t ¤ s, and t; s D 1; : : : ; n. For particular initial values
a1j0 and P1j0 , the MMSLE and corresponding MSE matrix are then determined as
follows:

atC1jt D  t atjt1 C kt vt
PtC1 D  t Pt  T
t C Rt ˙ "t RTt  kt MTt

where kt D Mt ft1 is the Kalman gain matrix, being Mt D  t Pt zt .

Before the MMSLE atC1jt and the MSE PtC1jt are computed in the Kalman filter,
the MMSLE of the state vector ˛t conditional on y1 ; : : : ; yt , denoted by atjt , and its
corresponding MSE matrix, denoted by Ptjt , can be computed as

atjt D atjt1 C Ptjt1 zt ft1 vt ; Ptjt D Ptjt1  Ptjt1 zt ft1 zTt Ptjt1 :

It follows that

atC1jt D  atjt ; PtC1jt D  t Ptjt  0t C Rt ˙ "t RTt :

Formal proofs of the Kalman filter can be found in Anderson and Moore [1],
Harvey [11], and Durbin and Koopman [8].

6.2.2 Likelihood Estimation

The Kalman filter can be used to evaluate the likelihood function via the prediction
error decomposition [11, 14, 26], according to which the joint density of y1 ; : : : ; yn
is specified as follows:

Y
n
p.y1 ; : : : ; yn / D p.y1 / p.yt jy1 ; : : : ; yt1 /:
tD2

The predictive density p.yt jy1 ; : : : ; yt1 / is Gaussian with mean E.yt jy1 ; : : : ; yt1 / D
zTt atjt1 and variance V.yt jy1 ; : : : ; yt1 / D zTt Ptjt1 zt CI2 D ft . Hence, for a realized
6.2 Linear State Space Models 157

time series y1 ; : : : ; yn , and denoting with  the vector of model parameters, the log-
likelihood function is given by

X
n
`./ D log p.y1 ; : : : ; yn / D log p.yt jy1 ; : : : ; yt1 / (6.20)
tD1

1X 1 X 2 1
n n
n
D  log.2/  log j ft j  v f :
2 2 tD1 2 tD1 t t

The one-step ahead prediction errors vt and their variances ft are computed by the
Kalman filter for a given value of the parameter vector .
The log likelihood (6.20) has not closed form and has to be numerically
maximized with respect to . This can be done using a numerical quasi-Newton
method, such as the Broyden–Fletcher–Goldfarb–Shanno (BFGS) method. It is
generally regarded as computationally efficient in terms of convergence speed and
numerical stability [24]. This iterative optimization method is based on information
from the gradient (or score). Analytical and computationally fast methods for
computing the score for a current value of  in a state space analysis are developed
by Koopman and Shepard [21].
An alternative method for maximum likelihood estimation is the EM-algorithm.
It is based on the joint density p.y1 ; : : : ; yn ; ˛1 ; : : : ; ˛n / and consists of two step.
The expectation (E) step takes the expectation of the components of the joint
density conditional on y1 ; : : : ; yn . Hence, in the E-step, the estimated state space
vector is evaluated using a smoothing algorithm related to the Kalman filter. In the
maximization (M) step the resulting expression of the joint density is maximized
with respect to . It is usually done analytically, and is simpler than maximizing the
full likelihood function directly. Given the updated estimate of  from the M-step,
the algorithm returns to the E-step and evaluate the smoothed estimates based on
the new parameter estimates. This iterative procedure converges to the maximum
likelihood estimate of . Details on the properties of the EM algorithm for state
space models can be found in Koopman [18].
The nonstationary trend and seasonal components, as discussed in Sect. 6.2, rely
on initial variables that are treated as fixed unknown coefficients. A straightforward
approach to the estimation of ˛1 is to estimate it jointly with  by the method of
maximum likelihood as previously discussed. However, numerical problems may
arise when the likelihood function is maximized with respect to a parameter vector
of high dimension. Several methods have been used in the literature, and most of
them can be embedded in a unified treatment for the initialization of the Kalman
filter with respect to the initial variables, and we refer the reader to Ansley and Kohn
[2], de Jong [6], and Koopman [18] for a detailed discussion. The latter references
also provide a detailed treatment in presence of stationary variables, where in this
case the initial conditions for ˛1 can be obtained from the theoretical autocovariance
function.
158 6 Seasonal Adjustment Based on Structural Time Series Models

6.2.3 Diagnostic Checking

The assumptions underlying model (6.1), and all the component specifications given
in Sect. 6.2, are that all disturbances related to the unobserved components are
normally distributed and serially and mutually independent with constant variances.
Under these assumptions the standardized one-step ahead prediction errors (or
prediction residuals) are given by
vt
et D p ; t D 1; : : : ; n;
ft

being also normally distributed and serially independent with unit variance.
The Normality assumption is analyzed looking at the first four moments of the
standardized forecast errors, that are given by

1X 1X
n n
m1 D et ; mq D .et  m1 /q ; q D 2; 3; 4:
n tD1 n tD1

When the model assumptions are valid, skewness M3 and kurtosis M4 are normally
distributed as [3]
   
m3 6 m4 24
M3 D q  N 0; ; M4 D 2  N 3; :
m32 n m2 n

Standard statistical tests can be used to check whether the observed values of M3
and M4 are consistent with their asymptotic densities.
A simple test for heteroscedasticity is obtained by comparing the sum of squares
of two exclusive subsets of the sample. For example, the statistic
Pn 2
tDnhC1 et
Ph 2
tD1 et

is Fh;h -distributed under the null hypothesis of homoscedasticity.

Finally, the correlogram of the prediction residuals should not reveal significant
serial correlation. A standard Portmanteau test statistic for serial correlation is based
on the Ljung–Box statistic [23]. This is given by

X
q
c2j
Q.q/ D n.n C 2/ ;
jD1
nj

for some positive integer q, that depends on the frequency of the seriesP (generally
q D 24 for monthly series), where cj is the j-th correlation cj D nm1 2 ntDjC1 .et 
m1 /.etj  m1 /. The statistic Q is distributed as a chi-square variable with q  p
6.3 Illustrative Example: Analysis of the US Unemployment Rate for Males 159

degrees of freedom if the residuals are independent, being p the number of

parameters in the model.

6.3 Illustrative Example: Analysis of the US Unemployment

Rate for Males Using STAMP

The implementation and estimation of the structural time series models discussed in
the previous sections are made simpler by the availability of a specific software
called STAMP (Structural Time series Analyzer, Modeller, and Predictor). It is
a graphical user interface (GUI)-based package designed to model and forecast
time series using the structural time series approach developed by Koopman et al.
[19]. STAMP runs on Windows, Macintosh, and Linux operating systems as part
of OxMetrics, a software system for (econometric) data analysis and forecasting
developed by Doornik [7]. It provides a powerful, flexible, easy to use, and up-to-
date environment to perform state space analysis of time series data.
STAMP enables a large variety of models to be analyzed, including all the
structural time series models discussed in the previous sections. The software also
provides an extensive array of test statistics, residual, and auxiliary residual graphics
after estimation for assessing the goodness of the model fit. Furthermore, OxMetrics
integrates STAMP into a larger class of related models, and provides a broad suite
of graphics, data transformation, and arithmetic options.
STAMP default model is the basic structural model discussed in Sect. 6.3 which
includes a stochastic level and slope specified through Eqs. (6.3) and (6.4), as well
as a stochastic seasonal component given in (6.13). By default, STAMP does not
allow for the treatment of trading day and holiday effects. Hence, it is appropriate
for the analysis of stock series. In this regard, we analyze the monthly series of US
Unemployment Rate for Males (16 years and over) observed from January 1992
to December 2013. In general, we recommend to select the automatic detection of
outliers in the Select components window among the several possible options. Once
the estimation is complete, some basic information appears in the Results window
as illustrated in Table 6.1.
The “Estimation report” tells us that the convergence was very strong. Maximum
likelihood estimation has been carried out by numerical optimization.
The “Diagnostic summary report” provides some of the basic diagnostic and
goodness of fit statistics discussed in Sect. 6.2.3. In Table 6.1, T denotes the sample
size (indicated by n in our notation), p the number of model parameters to estimate,
and N is the number of analyzed time series. The Ljung–Box test for residual serial
correlation, Q.24; 21/, is based on q D 24 residual autocorrelations and should
be tested against a chi-square distribution with q  p D 21 degrees of freedom.
Based on the observed value of Q equal to 24.628, we can conclude that there is no
serial correlations in the residuals, a strong indication that the model is adequately
capturing the dynamic structure of the series. The classical Durbin–Watson (DW)
160 6 Seasonal Adjustment Based on Structural Time Series Models

Table 6.1 STAMP output: diagnostic summary report

Estimating...
Very strong convergence relative to 1e-07
- likelihood cvg 7.35241e-15
- gradient cvg 1.03322e-08
- parameter cvg 5.02437e-08
- number of bad iterations 0
Estimation process completed.

UC( 1) Estimation done by Maximum Likelihood (exact score)

The selection sample is: 1992(1) - 2013(12) (T = 264, N = 1)
The dependent variable Y is: UNEMP_RATE
The model is: Y = Trend + Seasonal + Irregular + Interventions
Steady state ......found without full convergence

Log-Likelihood is 355.342 (-2 LogL = -710.684).

Prediction error variance is 0.041706

Summary statistics
UNEMP_RATE
T 264.00
p 3.0000
std.error 0.2042
Normality 0.7291
H(82) 1.2800
DW 2.0335
r(1) -0.0248
q 24.000
r(q) -0.0316
Q(q,q-p) 24.628
Rsˆ2 0.2933

test is also given. Since the DW statistic is approximately equal to 2.1  /,O where O
is the sample first order autocorrelation of the residuals, a value of the DW statistic
equal to 2 indicates no autocorrelation. The value of the DW statistics always lies
between 0 and 4. If the Durbin–Watson statistic is substantially less than 2, there is
evidence of positive serial correlation. As a rough rule of thumb, if Durbin/Watson
is less than 1, there may be cause for alarm. Small values of the DW statistic
indicate that successive error terms are, on average, close in value to one another, or
positively correlated. If the statistic is greater than 2, successive error terms are, on
average, much different in value from one another.
The heteroscedasticity test statistic H.82/, illustrated in Sect. 6.2.3, presents a
small value indicating that the residuals are characterized by constant variance
over time. Finally, the Normality test statistic of Bowman–Shenton discussed in
6.3 Illustrative Example: Analysis of the US Unemployment Rate for Males 161

1.0
UNEMP_RATE-Standardised Residuals +/- 2SE ACF-Standardised Residuals

2
0.5

0.0
0

-0.5

-2

1995 2000 2005 2010 0 20 40 60

Density QQ plot
0.4 N(s=1.06) resid normal

2
0.3

0.2 0

0.1
-2

-3 -2 -1 0 1 2 3 4 -2 -1 0 1 2 3

Fig. 6.1 Standardized residual graphics

Sect. 6.2.3, and based on third and fourth moments of the residuals, has a 2
distribution with 2 degrees of freedom when the model is correctly specified. The
5 % critical value is 5.99. The small value of the Normality test statistic together
with the results shown for the other statistics in Table 6.1 clearly indicates that the
basic structural model is adequate for the US Unemployment Rate series. This is
also confirmed by the standardized residuals graphs illustrated in Fig. 6.1.
STAMP also provides further information for the estimated model. As shown in
Table 6.2, in the Print parameter section it reports on the parameters, in this example
estimated by maximum likelihood, where the variances and all the parameters are
organized by each component. Concerning the variances of the disturbances, a zero
parameter estimate would suggest that the corresponding component is fixed. In our
specific case, this is true for the seasonal component, whereas the trend appears to
be stochastic as assumed. The q-ratios in parenthesis are the ratio of each standard
deviation to the standard deviation associated with the largest variance, being the
162 6 Seasonal Adjustment Based on Structural Time Series Models

12.5
UNEMP_RATE
UNEMP_RATE Level+Intv
Level+Intv

10.0

7.5

5.0

1995 2000 2005 2010

1.0
UNEMP_RATE-Seasonal
UNEMP_RATE-Seasonal

0.5

0.0

1995 2000 2005 2010

0.2
UNEMP_RATE-Irregular
UNEMP_RATE-Irregular

0.1

0.0

-0.1

1995 2000 2005 2010

Fig. 6.2 Estimated components in the basic structural model

latter related to the level of the trend component. The original series together with
the estimated components are shown in Fig. 6.2.
The final state vector contains information on the values taken by the various
components at the end of the sample. Following Koopman et al. [22], the final state
is the filtered estimate at time n, that is anjn as defined in Sect. 6.3. The square
roots of the diagonal elements of the corresponding MSE matrix, PO njn , are the root
mean square errors (RMSEs) of the corresponding state elements. The t-value is the
estimate of the state divided by the RMSE, and the corresponding p-value is also
reported. The null hypothesis of a zero value is rejected (at 5 % level) for both the
level and slope in the trend component, and for the seasonal dummies, except for
those related to June and December (at 1 % level).
STAMP automatically detected three outliers. In particular, level breaks are
identified in January 1994, December 2008, and April 2010.
References 163

Table 6.2 STAMP output: print parameter section

Variances of disturbances:
Value (q-ratio)
Level 0.0161857 (1.000)
Slope 0.000674890 (0.042)
Seasonal 0.00000 (0.000)
Irregular 0.00937361 (0.579)

State vector analysis at period 2013(12)

Value Prob
Level 6.13693 [0.00000]
Slope -0.13053 [0.03842]
Seasonal chi2 test 1209.31038 [0.00000]
Seasonal effects:
Period Value Prob
1 0.86587 [0.00000]
2 0.79596 [0.00000]
3 0.53084 [0.00000]
4 -0.11948 [0.00075]
5 -0.24775 [0.00000]
6 0.03332 [0.33913]
7 -0.11718 [0.00087]
8 -0.39016 [0.00000]
9 -0.47199 [0.00000]
10 -0.49448 [0.00000]
11 -0.31218 [0.00000]
12 -0.07277 [0.04151]

Regression effects in final state at time 2013(12)

Coefficient RMSE t-value Prob
Level break 1994(1) 0.69029 0.19110 3.61224 [0.00037]
Level break 2008(12) 0.84900 0.19090 4.44744 [0.00001]
Level break 2010(4) -0.72008 0.19089 -3.77215 [0.00020]

References

1. Anderson, B. D. O., & Moore, J. B. (1979). Optimal filtering. Englewood Cliffs: Prentice-Hall.
2. Ansley, C. F., & Kohn, R. (1985). Estimation, filtering and smoothing in state space models
with incompletely specified initial conditions. Annals of Statistics, 13, 1286–1316.
3. Bowman,p K. O., & Shenton, L. R. (1975). Omnibus test contours for departures from normality
based on b1 and b2 . Biometrika, 62, 243–250.
4. Burns, A., & Mitchell, W. (1946). Measuring business cycles. Working paper. New York:
NBER.
164 6 Seasonal Adjustment Based on Structural Time Series Models

5. Commandeur, J. J. F., & Koopman, S. J. (2007). An introduction to state space time series
analysis. Oxford: Oxford University Press.
6. de Jong, P. (1991). The diffuse Kalman filter. Annals of Statistics, 19, 1073–1083.
7. Doornik, J. A. (2009). An introduction to OxMetrics 6: A software system for data analysis and
forecasting. London: Timberlake Consultants Ltd.
8. Durbin, J., & Koopman, S. J. (2001). Time series analysis by state space methods. Oxford:
Oxford University Press.
9. Gomez, V. (2001). The use of Butterworth filters for trend and cycle estimation in economic
time series. Journal of Business and Economic Statistics, 19, 365–373.
10. Harvey, A. C. (1985). Trends and cycles in macroeconomic time series. Journal of Business
and Economic Statistics, 3, 216–227.
11. Harvey, A. C. (1989). Forecasting, structural time series models and the Kalman filter.
Cambridge: Cambridge University Press.
12. Harvey, A. C., Koopman, S. J., & Penzer, J. (1998). Messy time series: A unified approach. In
T. B. Fomby & R. Carter Hill (Eds.), Advances in econometrics (Vol. 13, pp. 103–143). New
York, NY, USA: JAI Press.
13. Harvey, A. C., & Trimbur, T. M. (2003). General model-based filters for extracting trends and
cycles in economic time series. Review of Economics and Statistics, 85, 244–255.
14. Jones, R. H. (1980). Maximum likelihood fitting of ARIMA models to time series with missing
observations. Technometrics, 22, 389–395.
15. Kalman, R. E. (1960). A new approach to linear filtering and prediction problems. Journal of
Basic Engineering, Transactions ASMA, Series D, 82, 35–45.
16. Kim, C. J., & Nelson, C. R. (1999). State space models with regime switching. Cambridge,
MA: MIT Press.
17. Kitagawa, G., & Gersch, W. (1996). Smoothness priors analysis of time series. New York:
Springer.
18. Koopman, S. J. (1997). Exact initial Kalman filtering and smoothing for non-stationary time
series models. Journal of the American Statistical Association, 92, 1630–1638.
19. Koopman, S. J, Harvey, A. C., Doornik, J. A., & Shepard, N. (2000). STAMP 8.2: Structural
time series analyser, modeller and predictor. London: Timberlake Consultants Ltd.
20. Koopman, S. J., & Ooms, M. (2011). Forecasting economic time series using unobserved
components time series models. In M. Clements & D. F. Hendry (Eds.), Oxford handbook
of economic forecasting. New York: Oxford University Press.
21. Koopman, S. J., & Shephard, N. (1992). Exact score for time series models in state space form.
Biometrika, 79, 823–826.
22. Koopman, S. J., Shephard, N., & Doornik, J. A. (1999). Statistical algorithms for models in
state space form using SsfPack 2.2. Econometrics Journal, 2, 113–166.
23. Ljung, G. M., & Box, G. E. P. (1978). On a measure of lack of fit in time series models.
Biometrika, 66, 67–72.
24. Nocedal, J., & Wright, S. J. (1999). Numerical optimization. New York: Springer.
25. Proietti, T. (2000). Comparing seasonal components for structural time series models. Interna-
tional Journal of Forecasting, 16, 247–260.
26. Schweppe, F. (1965). Evaluation of likelihood functions for Gaussian signals. IEEE Transac-
tions on Information Theory, 11, 61–70.
27. Trimbur, T. M. (2006). Properties of higher-order stochastic cycles. Journal of Time Series
Analysis, 27, 1–17.
 Part II
Trend-Cycle Estimation
Chapter 7
Trend-Cycle Estimation

Abstract Economists and statisticians are often interested in the short-term trend
of socioeconomic time series. The short-term trend generally includes cyclical
fluctuations and is referred to as trend-cycle. In recent years, there has been an
increased interest to use trend-cycle estimates. Among other reasons, this interest
originated from major economic and financial changes of global nature which have
introduced more variability in the data, and, consequently, in the seasonally adjusted
numbers. This makes very difficult to determine the direction of the short-term trend,
particularly to assess the presence or the upcoming of a turning point. This chapter
discusses in detail stochastic and deterministic trend-cycle models formulated for
global and local estimation with special reference to the TRAMO-SEATS and
STAMP software as well as the Henderson filters used in the Census II-X11 and its
variants X11/X12ARIMA. It includes also trend-cycle models obtained with locally
weighted averages as the Gaussian kernel, locally weighted regression smoother
(LOESS), and cubic smoothing spline. Many illustrative examples are exhibited.

The trend corresponds to sustained and systematic variations over a long period
of time. It is associated with the structural causes of the phenomenon in question,
for example, population growth, technological progress, new ways of organization,
or capital accumulation. For the majority of socioeconomic time series, the trend
is very important because it dominates the total variation of the series. The
identification of the trend component has always posed a serious statistical problem.
The problem is not of mathematical or analytical complexity, but of conceptual
complexity. This problem exists because the trend and the remaining components
of a time series are latent (not directly observable) variables and, therefore,
assumptions must be made on their behavioral pattern. The trend is generally
thought of as a smooth and slow movement over a long term. The concept of “long”
in this connection is relative and what is identified as trend for a given series span
might well be part of a long cycle once the series is considerably augmented, such as
the Kondratieff economic cycle. Kondratieff [35] estimated the length of this cycle
to be between 47 and 60 years. Often, a long cycle is treated as a trend because the
length of the observed time series is shorter than one complete cycle.
To avoid the complexity of the problem posed by a statistically vague definition,
statisticians have resorted to two simple solutions. One consists of estimating trend

© Springer International Publishing Switzerland 2016 167

E. Bee Dagum, S. Bianconcini, Seasonal Adjustment Methods and Real Time
Trend-Cycle Estimation, Statistics for Social and Behavioral Sciences,
DOI 10.1007/978-3-319-31822-6_7
168 7 Trend-Cycle Estimation

and cyclical fluctuations together calling this combined movement trend-cycle. The
other consists of defining the trend in terms of the series length, denoting it as the
longest no periodic movement. The estimation of the trend-cycle can be done via a
specified model applied to the whole data, called the global trend-cycle, or by fitting
a local polynomial function in such a way that, at any time point, its estimates
only depend on the observations at that point and some specified neighboring
observations.
Local polynomial fitting has a long history in the smoothing of noisy data.
Henderson [30], Whittaker and Robinson[49], and Macauley [37] are some of the
earliest classical references. These authors were very much concerned with the
smoothing properties of linear estimators, being Henderson [30] the first to show
that the smoothing power of a linear filter depends on the shape and values of its
weighting system.
On the other hand, more recent contributions (among others, [8, 18–21, 25, 26,
44, 48]) concentrated on the asymptotic statistical properties of optimally estimated
smoothing parameters. Optimality is being defined in terms of minimizing a given
loss function, usually the mean square error of the estimator or the prediction risk.
In this chapter, we will discuss stochastic and deterministic trend-cycle models
formulated for global and local estimation with special reference to those available
in TRAMO-SEATS and STAMP, as well as the Henderson filters used in the Census
II-X11 and its variants X11/X12ARIMA.

7.1 Deterministic Global Trend Models

Deterministic global trend models are based on the assumption that the trend or
nonstationary mean of a time series can be approximated closely by simple functions
of time over the entire span of the series.
The most common representation of a deterministic trend is by means of
polynomial functions. The time series from which the trend has to be identified
is assumed to be generated by a nonstationary process where the nonstationarity
results from a deterministic trend. A classical model is the regression or error model
[3] where the observed series is treated as the sum of a systematic part or trend and
a random part. This model can be written as

yt D Tt C et (7.1)

where fet g is a purely random process, that is, fet g  i:i:d:.0; e2 / (independent and
identically distributed with expected value 0 and variance, e2 ).
In the case of a polynomial trend-cycle,

Tt D a0 C a1 t C a2 t2 C    C ap tp (7.2)
7.1 Deterministic Global Trend Models 169

where generally n  3. The trend is said to be of a deterministic character because

it is not affected by random shocks which are assumed to be uncorrelated with the
systematic part.
Model (7.1) can be generalized by assuming that fet g is a second order linear
stationary stochastic process, that is, its mean and variance are constant and its
autocovariance is finite and depends only on the time lag.
Besides polynomials in time, other suitable mathematical functions are used to
represent deterministic trends. Three of the most widely applied functions, known
as growth curves, are the modified exponential, the Gompertz, and the logistic.
The modified exponential trend can be written as

Tt D a C bct ; a 2 R; b ¤ 0; c > 0; c ¤ 1: (7.3)

For a D 0, model (7.3) reduces to the unmodified exponential trend

Tt D bct D T0 e˛t ; b D T0 ; ˛ D log c: (7.4)

When b > 0 and c > 1, and so ˛ > 0, model (7.4) represents a trend that increases
at a constant relative rate ˛. For 0 < c < 1, the trend decreases at the rate ˛.
Models (7.3) and (7.4) are solutions of the differential equation

dTt
D ˛.Tt  a/; ˛ D log c (7.5)
dt
which specifies the simple assumption of no inhibited growth.
Several economic variables during periods of sustained growth or of rapid
inflation, as well as population growths measured in relative short periods of time,
can be well approximated by trend models (7.3) and (7.4). But in the long run,
socioeconomic and demographic time series are often subject to obstacles that
slow their time path, and if there are no structural changes, their growth tends to
a stationary state. Quetelet [40] made this observation with respect to population
growth and Verhulst [46] seems to have been the first to formalize it by deducing
the logistic model. Adding to Eq. (7.5) an inhibit factor proportional to Tt2 , the
result is
dTt
D ˛Tt  ˇTt2 D ˛Tt .1  Tt =k/ (7.6)
dt
k D ˛=ˇ; ˛; ˇ > 0

which is a simple null form of the Riccati differential equation.

Solving Eq. (7.6), we obtain the logistic model,

Tt D k.1 C ae˛t /1 (7.7)

where a > 0 is a constant of integration.

170 7 Trend-Cycle Estimation

Model (7.7) belongs to a family of S-shaped curves generated from the differen-
tial equation (see Dagum [10]):

dTt
D Tt  .t/˚.Tt =k/; ˚.1/ D 0: (7.8)
dt

Solving Eq. (7.8) for  D log c and ˚ D log.Tt =k/, we obtain the Gompertz curve
used to fit mortality table data. That is,
t
Tt D kbc ; b > 0; b ¤ 1; 0 < c < 1; (7.9)

where b is a constant of integration. It should be noticed that differencing will

remove polynomial trends and suitable mathematical transformations plus differ-
encing will remove trends from nonlinear processes. For example, for (7.7) using
 
Tt
Zt D log
k  Tt

and then taking differences give .1  B/Zt D ˛.

7.2 Stochastic Global Trend Models

The second major class of global trend models is the one that assumes the trend
to be a stochastic process, most commonly that the series from which the trend
will be identified has been generated by a homogeneous nonstationary stochastic
process [51]. Processes of this kind are nonstationary, but applying a homogeneous
filter, usually the difference filter, we obtain a stationary process in the differences
of a finite order. In empirical applications, the nonstationarity is often present in the
level and/or slope of the series. Hence, the order of the difference is low.
An important class of homogeneous linear nonstationary processes are the
Autoregressive Integrated Moving Average (ARIMA) processes which can be
written as [4]

p .B/.1  B/d Tt D q .B/aTt ; aTt  i:i:d:.0; T2 /; (7.10)

where B is the backshift operator, such that Bn Tt D Ttn , p .B/ and q .B/
are polynomials in B of order p and q, respectively, satisfying the conditions of
stationarity and invertibility. .1  B/d is the difference operator of order d, and faTt g
is a purely random process. Model (7.10) is also known as an ARIMA process of
order (p; d; q). If p D 0, the process follows an IMA.0; q/ model.
7.2 Stochastic Global Trend Models 171

Three common stochastic trend models are the IMA.1; 1/, IMA.2; 2/, and
ARIMA.2; 1; 2/ which take the following forms
• IMA.1; 1/

.1  B/Tt D .1  B/aTt ; (7.11)

jj < 1; aTt  i:i:d:.0; T2 /

or, equivalently,

Tt D Tt1 C aTt  aTt1 : (7.12)

• IMA.2; 2/

.1  B/2 Tt D .1  1 B  2 B2 /aTt ; (7.13)

2 C 1 < 1 ; 2  1 < 1; j2 j < 1;
aTt  i:i:d:.0; T2 /;

or, equivalently,

Tt D 2Tt1  Tt2 C aTt  1 aTt1  2 aTt2 :

• ARIMA.2; 1; 2/

.1  1 B  2 B2 /.1  B/Tt D .1  1 B  2 B2 /aTt ; (7.14)

2 C 1 < 1; 2  1 < 1; 1 < 2 < 1
2 C 1 < 1; 2  1 < 1; 1 < 2 < 1
aTt  i:i:d:.0; T2 /

or equivalently

Tt D .1 C 1 /Tt1 C .2  1 /Tt2  2 Tt3 C aTt  1 aTt1  2 aTt2 : (7.15)

The aTt ’s may be regarded as a series of random shocks that drive the trend and  can
be interpreted as measuring the extent to which the random shocks or innovations
incorporate themselves into the subsequent history of the trend. For example, in
model (7.11), the smaller the value of , the more flexible the trend; the higher the
value of , the more rigid the trend (less sensitive to new innovations). For  D 1,
model (7.11) reduces to one type of random walk model which has been used mainly
for economic time series, such as stock market price data [23]. In such models, as
time increases the random variables tend to oscillate about their mean value with an
ever increasing amplitude.
172 7 Trend-Cycle Estimation

Stochastic models for the trend have become widespread in economics (see Stock
and Watson [45]), and several modeling approaches have been suggested. Within the
context of linear stochastic processes, there are two general ones. First, the so-called
ARIMA Model-Based (AMB) approach, adopted by TRAMO-SEATS, in which the
model for the trend is derived from the ARIMA model identified for the observed
series (see Burman [5]). The other approach starts by directly specifying a model for
the trend, as implemented in STAMP. Basic references are Harvey [27], and Harvey
and Todd [29].

7.2.1 TRAMO-SEATS Trend Models

Let yt be a series which is the sum of a trend Tt and a nontrend component et ,

yt D Tt C et ; (7.16)

where the two components are uncorrelated and each generated by an ARIMA
model as follows

'T .B/Tt D T .B/aTt ; (7.17)

'e .B/et D e .B/aet ; (7.18)

where 'T .B/; 'e .B/; T .B/ and e .B/ are finite polynomials in B. In particular,
'T .B/ D T .B/ıT .B/ and 'e .B/ D e .B/ıe .B/ where both ıT .B/ and ıe .B/ contain
unit autoregressive roots. aTt and aet are orthogonal white noise variables, with
variances T2 and e2 , respectively. It is generally assumed that the roots of the AR
polynomials T .B/ and e .B/ are different, since AR roots for the same frequency
should be associated with the same component. The two polynomials have also to
be prime.
Combining Eqs. (7.16)–(7.18), yt is assumed to follow an ARIMA model of the
type

'.B/yt D .B/at ; (7.19)

where '.B/ D .B/ı.B/, being .B/ D T .B/e .B/, and ı.B/ D .1  B/d .1  Bs /D
where s is the number of observations for year. It can be easily derived that the
polynomial .1  Bs / can be factorized as .1  Bs / D .1  B/.1 C B C    C Bs1 /,
where .1 C B C    C Bs1 / denotes the annual aggregation operator. .B/at is a
Moving Average (MA) process such that

.B/at D 'e .B/T .B/aTt C 'T .B/e .B/aet ; (7.20)

7.2 Stochastic Global Trend Models 173

with at  WN.0; a2 /. It has to be noticed that in order to allow model (7.19) to be
invertible it is assumed that T .B/ and e .B/ do not have common unit root.
In the AMB approach '.B/, .B/ and a2 are assumed known, and the parameters
of 'T .B/, 'e .B/, T .B/ and e .B/ as well as the variances T2 and e2 have to be
derived from them. This is done in the frequency domain. Based on Eq. (7.16),
the pseudospectrum of yt is partitioned into additive spectra associated with the
terms in the RHS of Eq. (7.16). Identification of a unique decomposition is achieved
by imposing the canonical condition that the minimum of the trend component
pseudospectrum be zero (see Burman [5]).
The Wiener–Kolmogorov (WK) filter for estimating Tt is given by the ratio
of the Tt and yt pseudospectra, and yields the Minimum Mean Square Error
(MMSE) estimator (also conditional mean) of Tt given yt . The filter is centered,
symmetric, and convergent. Its derivation requires an infinite realization of yt in
the direction of the past and of the future. Hence, to apply the filter to a finite
realization, model (7.19) is used to extend yt with forecasts and backcasts. As new
observations become available, forecasts will be updated and eventually replaced
by observations. As a consequence, the estimator of Tt near the end of the series is
preliminary and will be revised. The duration of the revision process of a preliminary
estimator depends on the ARIMA model identified for the series. The spectral
factorization provides the time domain expression of the components given in
Eqs. (7.17) and (7.18).
Letting F denote the forward shift operator (F D B1 ) and replacing the ratio
of pseudospectra by the ratio of Autocovariance Generating Functions (ACGF), the
time domain expression of the WK filter is

T .B/T .F/'e .B/'e .F/

T .F; B/ D kT ;
.B/.F/

where kT D T2 =e2 , so that the final trend estimator is given by

TO t D T .B; F/yt : (7.21)

As an example, we consider a general class of models widely applied to observed

series, that is,

.1  B/.1  Bs /yt D .B/at at  WN.0; a2 /: (7.22)

174 7 Trend-Cycle Estimation

In this case, the decomposition (7.20) results

.B/at D .1 C B C    C Bs1 /T .B/aTt C .1  B/2 e .B/aet

where the order q of the polynomial .B/ is equal to max.qT C s  1; qe C 2/, being
qT and qe the orders of the polynomials T .B/ and e .B/, respectively.
Following Maravall [38], necessary and sufficient conditions for the decomposi-
tion (7.20) to be underidentified are that qT > 1 and qe > s  2.
The AMB decomposition restricts the order of the AR and MA polynomial in
model (7.17) for the trend to be of the same order. Hence, the trend model becomes
an IMA.2; 2/ defined in Eq. (7.13), that is,

.1  B/2 Tt D .1  1 B  2 B2 /aTt : (7.23)

Based on this trend specification, the decomposition (7.19) is still underidentified

and a further condition is imposed. Due to the fact that the randomness of Tt
is determined by aTt , all the models of type (7.23) that are compatible with the
stochastic structure of the observed series, that is, with model (7.19), the one with
smallest T2 is chosen. This yields the most stable trend, given the observed ARIMA
model. As shown by Hillmer and Tiao [31], minimizing T2 is equivalent to the
requirement that it should not be possible to further decompose Tt as the sum of
a trend plus a noise component, with the latter orthogonal to the former. When
a component satisfies this “noise-free” requirement, it is called “canonical.” The
canonical trend is uncontaminated by noise, such that its spectrum should contain
a zero, that means there is a unit root in the MA polynomial of the trend. Hence,
model (7.23) can be rewritten as

.1  B/2 Tt D .1  1 B/.1 C B/aTt (7.24)

which contains two parameters 1 and T2 . The number of the unknowns in the
system of covariance equations is .qe C 3/ and the number of equations is greater
than .qe C 3/. The decomposition becomes identifiable and there will be a unique
model (7.24) which will represent the trend component contained in model (7.22).
As an illustrative example, Fig. 7.1 shows the trend estimates, derived using
the TRAMO-SEATS software, of the series of the US New Orders for Durable
Goods (NODG) observed for the period February 1992–December 2013. The
trend estimates have been obtained through canonical decomposition of the
ARIMA.0; 1; 1/.0; 1; 1/12 identified by TRAMO-SEATS to be the stochastic
generating process of the original series.
7.2 Stochastic Global Trend Models 175

NODG series

200000
100000 TRAMO-SEATS trend estimate

Jan 1995 Jan 2000 Jan 2005 Jan 2010

Fig. 7.1 The US New Orders for Durable Goods trend-cycle estimates obtained using the
TRAMO-SEATS software

7.2.2 STAMP Trend Models

Structural models can be seen as the sum of unobservable ARIMA models specified
to describe the behavior of each time series component separately, and random walk
processes are basic elements in many of these models.
The simplest structural model consists of a trend component Tt plus a random
disturbance term It , where the level Tt , representing the long-term movements in
the series, is allowed to fluctuate up and down according to a random walk process.
In particular, if the irregular component is assumed to be normally distributed, this
local level model is specified as

yt D Tt C It ; It  NID.0; I2 /; (7.25)

Tt D Tt1 C aTt ; aTt  NID.0; T2 /;

for t D 1; 2; : : : ; n, where the It ’s and aTt ’s are mutually independent. The model is
incomplete without a specification of the trend at T1 , which generally is assumed
to be normally distributed T1  N.1 ; P1 /, where 1 and P1 are known, and
independent of It and aTt .
When I2 is equal to 0, the observations follow a random walk and the forecast
of future observations is the last observation yT . If T2 is equal to 0, then the level is
constant and the best forecast of future observation is the sample mean. The effect
of aTt is to allow the trend to shift up and down. The larger the variances, the greater
the stochastic movements in the trend. Although it is simple, the local level model
176 7 Trend-Cycle Estimation

provides the basis for the treatment of important series in practice. The properties
of series generated by this model are studied in detail by Harvey [28].
An extension of the local level model is represented by the local linear trend
model which is characterized by a stochastic linear trend whose level and slope vary
over time according to random walk processes as follows

yt D Tt C It ; (7.26)
Tt D Tt1 C ˇt1 C aTt ;
ˇt D ˇt1 C
t ;
 NID.0; 
2 /;

where aTt and

t are uncorrelated white noise disturbances. The trend is equivalent
to an ARIMA(0,2,2) process [41]. However, if both T2 and 
2 are equal to zero,
then Tt reduces to a deterministic trend, that is, Tt D T1 C ˇt. On the other hand,
if T2 D 0, but 
2 > 0, the trend is an Integrated process of order 2, I.2/, that
is stationary in the second order differences. A trend component with this feature
tends to be relatively smooth. An important issue is therefore whether or not the
constraint T2 D 0 should be imposed at the outset. There are series where it is
unreasonable to assume a smooth trend a priori, and therefore the question whether
or not T2 is set to zero is an empirical one.
In this model, ˇt shows the contribution of the slope, and
t allows the slope to
change over time. If 
2 is equal to zero, the slope is constant and the local linear
trend model reduces to a random walk plus drift model. On the other hand, when
the estimated value of the slope and its variance are very small, the trend follows a
local level model without drift described in Eq. (7.25).
For the US NODG series introduced in the previous section, Fig. 7.2 illustrates
the trend component estimated by fitting the basic structural model which includes
a local linear trend model as specified in Eq. (7.26) as well as a stochastic
seasonal component. The estimated component has been obtained using the STAMP
(Structural Time series Analyser, Modeller and Predictor) software.
Both the local level model and the local linear trend model can be expressed in
matrix form as special cases of the linear Gaussian state space model defined as
follows

yt D zTt ˛t C It ; It  N.0; I2 /; (7.27)

˛tC1 D  t ˛t C Rt "t ; "t  N.0; ˙ " /; (7.28)
˛1  N.a1 ; P1 /:

Equation (7.27) is called the observation equation and Eq. (7.28) is called the state
equation, being the unobserved vector ˛t defined as the state. The vector zt and the
matrices  t ; Rt , and  " are assumed known as well as I2 , whereas the disturbances
It and "t are independent sequences of normal random variables. The matrix Rt ,
when it is not the identity, is usually a selection matrix whose columns are a subset
of the columns of the identity matrix. It is needed when the dimensionality of ˛t is
7.2 Stochastic Global Trend Models 177

250000
NODGseries
200000
150000
100000 Local linear trend estimate

Jan 1995 Jan 2000 Jan 2005 Jan 2010

Fig. 7.2 The US New Orders for Durable Goods trend estimates obtained using the STAMP
software

greater than that of the disturbance vector "t . The structure of model defined through
Eqs. (7.27) and (7.28) is a natural one for representing the behavior of many time
series as a first approximation. The first equation is a standard multivariate linear
regression model whose coefficient vector ˛t varies over time. The development
over time of ˛t is determined by the first order autoregressive vector given in the
second equation. The Markovian nature of the model has a remarkably wide range
of applications to problems in practical time series analysis where the local level
and local linear trend models are special cases. The local level model admits a state
space representation by setting in Eq. (7.27) ˛t D Tt , zt D 1 and in Eq. (7.28)
 t D Rt D 1, and ˙ " D T2 .
On the other hand, the local linear trend model defined in Eq. (7.26) can be
rewritten in matrix form as follows
 

Tt
yt D 1 0 C It ;
ˇt
      
TtC1 11 Tt aTt
D C ;
ˇtC1 01 ˇt
t
T1  N.1 ; T21 /:
178 7 Trend-Cycle Estimation

For the measurement of the short-term trend, the business cycle component has to
be included in the model. Hence, the trend plus cycle model is specified as follows

yt D Tt C Ct C It ; (7.29)
Tt D Tt1 C ˇt1 C aTt ;
ˇt D ˇt1 C
t ;

where Ct is a cycle with frequency (in radians) c , being 0  c  . It is generated

by the stochastic process

Ct D  cos c Ct1 C  sin c Ct1 C a Ct ; (7.30)
Ct D  sin c Ct1 C 
 cos c Ct1 C aCt ;

being  a damping factor on the amplitude such that 0    1, and both aCt and
aCt are NID.0; C2 /. The irregular component It is assumed to follow an NID.0; I2 /
process, and the disturbances in all the three components are taken to be independent
of each other. It can be seen that cycle and trend are simply added to each other.
When  is zero, then this model is not identifiable and, since Ct is assumed to be a
stationary process,  is strictly less than one. In particular, Ct follows an ARMA(2,1)
process in which both the MA and AR parts are subject to restrictions [27]. The
period of the cycle, which is the time taken to go through its complete sequence of
values, is 2=c . When T2 is equal to zero and C2 is small, the trend plus cycle
model allocates most of the short run variation in the series to the nonstationary
trend component. When both 
2 and T2 are small, then the trend becomes much
smoother and most of the variation in the series is allocated to the stationary cyclical
component. Finally, the trend plus cycle model reduces to an AR(1) process when
c is equal to zero or .
Estimation of the parameters I2 , T2 , 
2 , ; c , and C2 can be carried out by
maximum likelihood either in the time domain or in the frequency domain. Once
this has been done, estimates of the trend, cyclical, and irregular components are
obtained from a smoothing algorithm. These calculations may be carried out using
the STAMP software.

7.3 Stochastic Local Trend-Cycle Models

Economists and statisticians are often interested in the short-term trend of socioeco-
nomic time series. The short-term trend generally includes cyclical fluctuations and
is referred to as trend-cycle. In recent years, there has been an increased interest
to use trend-cycle estimates or smoothed seasonally adjusted data to facilitate
recession and recovery analysis. Among other reasons, this interest originated from
major economic and financial changes of global nature which have introduced more
7.3 Stochastic Local Trend-Cycle Models 179

variability in the data, and, consequently, in the seasonally adjusted numbers. This
makes very difficult to determine the direction of the short-term trend, particularly
to assess the presence or the upcoming of a turning point. The local polynomial
regression predictors developed by Henderson [30] and LOESS due to Cleveland
[7] are widely applied to estimate the short-term trend of seasonally adjusted
economic indicators. Particularly, the former is available in nonparametric seasonal
adjustment software such as the US Bureau of the Census II-X11 method [43] and
its variants the X11ARIMA ([11] and [12]) and X12ARIMA [22], the latter, in the
STL (Seasonal-Trend decomposition procedure based on LOESS) [6].
The basic assumption is that the input series fyt ; t D 1; : : : ; ng, that is seasonally
adjusted or without seasonality, can be decomposed into the sum of a systematic
component called the signal (or nonstationary mean) TCt , that represents the trend
and cyclical components, usually referred to as trend-cycle for they are estimated
jointly, plus an erratic component, called the noise It , such that

yt D TCt C It : (7.31)

The irregular component It is assumed to be either a white noise, WN.0; I2 /, or,
more generally, to follow a stationary and invertible ARMA process. The trend-
cycle can be represented locally by a polynomial of degree of the time distance j
between yt and the neighboring observations ytCj . Hence, it is possible to find a
local polynomial trend-cycle estimator

TCt . j/ D b0 C b1 j C    C bp jp C aTCt . j/; (7.32)

where b0 ; b1 ; : : : ; bp are real and aTCt is assumed to be purely random and mutually
uncorrelated with It . The coefficients, b0 ; b1 ; : : : ; bp , can be estimated by ordinary or
b
weighted least squares or by summation formulae. The solution for TCt .0/ provides
the trend-cycle estimate, which equivalently is a weighted average applied in a
moving manner [32], such that

b b X
m
TCt .0/ D TCt D wj ytj ; (7.33)
jDm

where wj ; j D m; : : : ; m, denotes the weights to be applied to the observations ytj

b
to get the estimate TCt for each point in time t D 1; 2; : : : ; n.
The weights depend on:
(1) the degree p of the fitted polynomial,
(2) the amplitude (2m C 1) of the neighborhood, and
(3) the shape of the function used to average the observations in each neighborhood.
Once a (symmetric) span 2m C 1 of the neighborhood has been selected,
the weights associated with the observations corresponding to points falling out
of the neighborhood of any target point are null or approximately null, such
180 7 Trend-Cycle Estimation

that the estimates of the n  2m central observations are obtained by applying

2m C 1 symmetric weights to the observations neighboring the target point. The
missing estimates for the first and last m observations can be obtained by applying
asymmetric moving averages of variable length to the first and last m observations,
respectively. The length of the moving average or time invariant symmetric linear
filter is 2m C 1, whereas the asymmetric linear filters length is time-varying.
Using the backshift operator B, such that Bn yt D ytn , Eq. (7.33) can be written
as

b X
m
TCt D wj Bj yt D W.B/yt ; t D 1; 2; : : : ; n; (7.34)
jDm

where W.B/ is a linear nonparametric estimator. The nonparametric estimator W.B/

is said to be a second order estimator if it satisfies the conditions

X
m X
m
wj D 1; jwj D 0; (7.35)
jDm jDm

such that it preserves a constant and a linear trend. On the other hand, W.B/ is a
higher order estimator if

X
m X
m
wj D 1; ji wj D 0 i D 1; 2; : : : ; p: (7.36)
jDm jDm

In other words, it will reproduce a polynomial trend of degree (p  1) without

distortion.
Nonparametric estimators are based on different assumptions of smoothing. We
shall next discuss: (1) the local polynomial regression predictor LOESS due to
Cleveland [7], (2) the Henderson linear filters, (3) the Gaussian kernel, and (4) the
cubic smoothing spline. The first three estimators are often applied to estimate the
short-term trend of seasonally adjusted economic indicators.

7.3.1 Locally Weighted Regression Smoother (LOESS)

The locally weighted regression smoother (LOESS) developed by Cleveland [7]

fits local polynomials of a degree p where the parameters are estimated either by
ordinary or weighted least squares. Given a series of equally spaced observations
and corresponding target points f.yti ; ti /; i D 1; : : : ; ng; t1 <    < tn , where ti
denotes the time the observation yti is taken, LOESS produces a smoothed estimate
as follows

c ti D t0i bi
TC (7.37)
7.3 Stochastic Local Trend-Cycle Models 181

where ti is a (p C 1)-dimensional vector of generic component tid ; d D 0; : : : ; p,

p D 0; 1; 2; : : : denotes the degree of the fitted polynomial, and bi is the (p C 1)-
dimensional least squares estimate of a weighted regression computed over a
neighborhood of ti constituting a subset of the full span of the series.
The weights of the regression depend on the distance between the target point
ti and any other point belonging to its neighborhood, through a weighting function
W.t/. The weighting function most often used is the tricube proposed by Cleveland
et al. [6], that is,

3
W.t/ D 1  jtj3 IŒ1;1 .t/: (7.38)

In particular, at each point tk in the neighborhood of the target point ti , denoted by

N.ti /, has assigned a weight
 
jti  tk j
w.tk / D W 8tk 2 N.ti / (7.39)
5.ti /

with 5.ti / representing the distance of the furthest near-neighbor from ti .

Each neighborhood is made of the same number of points chosen to be the nearest
to ti , and the ratio between the amplitude of the neighborhood and the full span
of the series defines the bandwidth or smoothing parameter. For the first and last
observations, Cleveland [7] derived the filters by weighting the data belonging to
an asymmetric neighborhood which contains the same number of data points of the
symmetric one.

7.3.2 Henderson Smoothing Filter

The Henderson smoothing filters are derived from the graduation theory and used in
Census II-X11 method and its variants X11/ X12ARIMA. The basic principle of the
graduation theory is the combination of operations of differencing and summation
in such a manner that, when differencing above a certain order is ignored, they
will reproduce the functions operated on. The merit of these procedures is that the
smoothed values thus obtained are functions of a large number of observations
whose errors, to a considerable extent, cancel out. These smoothers have the
properties that when fitted to second or third degree parabolas, their output will
fall exactly on those parabolas and when fitted to stochastic data, they will give
smoother results than can be obtained from weights that give the middle point to a
second degree parabola fitted by ordinary least squares.
Recognition of the fact that the smoothness of the resulting graduation depends
directly on the smoothness of the weight diagram led Henderson [30] to develop a
formula which makes the sum of squares of the third differences of the smoothed
series a minimum for any number of terms.
182 7 Trend-Cycle Estimation

Henderson’s starting point was the requirement that the filter should reproduce a
cubic polynomial trend without distortion. Henderson showed that three alternative
smoothing criteria lead to the same formula:
1. minimization of the variance of the third differences of the series defined by the
application of the moving average;
2. minimization of the sum of squares of the third differences of the coefficients of
the moving average formula; and
3. fitting a cubic polynomial by weighted least squares, where the weights are
chosen as to minimize the sum of squares of their third differences.
Kenny and Durbin [33] and Gray and Thomson [24] showed the equivalence of
these three criteria. The problem is one of locally fitting a cubic trend by weighted
least squares to the observations where the weights are chosen to minimize the sum
of squares of their third differences (smoothing criterion). The objective function to
be minimized is

X
m
 2
Wj yatCj  a0  a1 j  a2 j2  a3 j3 ; (7.40)
jDm

c t , Wj D
where the solution for the constant term aO 0 is the smoothed observation TC
Wj , and the filter length is 2m C 1. The solution is a local cubic smoother with
weights
   
Wj / .m C 1/2  j2 .m C 2/2  j2 .m C 3/2  j2 ; (7.41)

and the weight diagram known as Henderson’s ideal formula is obtained, for a filter
length equal to 2m  3,
   
315  .m  1/2  j2 .m2  j2 / .m C 1/2  j2 .3m2  16  11j2 /
wj D :
8m.m2  1/.4m2  1/.4m2  9/.4m2  25/
(7.42)

By making m D 8 the wj values are obtained for each j of the 13-term filter (m D 6
for the 9-term Henderson filter, and m D 13 for the 23-term Henderson).
On the contrary, the weights of the usually known as asymmetric Henderson
filters developed by Musgrave [39] are based on the minimization of the mean
squared revision between the final estimates and the preliminary estimates subject
to the constraint that the sum of the weights is equal to one. The assumption made
is that the most recent values of the series (where seasonality has been removed if
present in the original observations) follow a constant linear trend plus an erratic
7.3 Stochastic Local Trend-Cycle Models 183

component (see Laniel [36] and Doherty [17]). The equation used is
0 12
h i2 X
m X
m
.i;m/
E rt D c21 @t  wa;j .t  j/A C I2 .wj  wa;j /2 ; (7.43)
jDa jDm

where w and wa are the weights of the symmetric (central) filter and the asymmetric
filters, respectively; wa;j D 0 for j D m; : : : ; .a C 1/, c1 is the slope of the line
and I2 denotes the noise variance. There is a relation between c1 and I such that
the noise to signal ratio, I=C, is given by

.4I2 =/1=2 c21 4

I=C D or 2
D 2
: (7.44)
jc1 j I .I=C/

The noise to signal ratio I=C (7.44) determines the length of the Henderson trend-
cycle filter to be applied. Thus, setting t D 0 and m D 6 for the end weights of the
13-term Henderson filter, we have
0 12
.i;6/ 6
X X6
EŒr0 2 4
D @ wa;j A C .wj  wa;j /2 : (7.45)
I2 .I=C/2 jDa jD6

Making I=C D 3:5 (the most noisy situation where the 13-term Henderson filter is
applied), Eq. (7.45) gives the same set of end weights of Census II-X11 variant [43].
The end weights for the remaining monthly Henderson filters are calculated using
I=C D 0:99 for the 9-term filter and I=C D 4:5 for the 23-term filter.
The estimated final trend-cycle is obtained by cascade filtering that results from
the convolution of various linear trend and seasonal filters. In fact, if the output from
the filtering operation H is the input to the filtering operation Q, the coefficients of
the cascade filter C result from the convolution of H  G. For symmetric filters,
H  G D G  H, but this is not valid for asymmetric filters. Assuming an input
series yt ; t D 1; 2; : : : ; n, we can define a matrix W D Œwa;j ; a D 1; 2; : : : ; mI j D
1; 2; : : : ; 2m C 1, where each row is a filter and m is the half length of the symmetric
filter. w0; denotes an asymmetric filter where the first m coefficients are zeroes,
whereas wm; denotes the symmetric filter.
Given data up to time n, the m C 1 most recent values of the output (filtered
series) are given by

2mC1
X
c nC1a D
TC wa;j yn2m1Cj a D 1; 2; : : : ; m C 1: (7.46)
jDmaC2
184 7 Trend-Cycle Estimation

The 13-term Henderson filter, which is the most often applied, can then be put in
matrix form as follows:
2 3
0:019 0:028 0:000 0:066 0:147 0:214 0:240 0:214 0:147 0:066 0:000 0:028 0:019
6 7
6 0 0:017 0:025 0:001 0:066 0:147 0:213 0:238 0:212 0:144 0:061 0:006 0:034 7
6 7
6 7
6 0 0 0:011 0:022 0:003 0:067 0:145 0:210 0:235 0:205 0:136 0:050 0:018 7
6 7
6 7
6 0:145 0:208 0:230 0:201 0:131 0:046 7
6 0 0 0 0:009 0:022 0:004 0:066 7:
6 7
6 0 0 0 0 0:016 0:025 0:003 0:068 0:149 0:216 0:241 0:216 0:148 7
6 7
6 7
6 0 0 0 0 0 0:043 0:038 0:002 0:080 0:174 0:254 0:292 0:279 7
4 5
0 0 0 0 0 0 0:092 0:058 0:012 0:120 0:244 0:353 0:421
(7.47)

7.3.3 Gaussian Kernel Smoother

Kernel type smoothers are locally weighted averages. Given a series of seasonally
adjusted observations and corresponding target points .ytj ; tj /; j D 1; : : : ; n, a kernel
smoothing gives, at time th ; 1  h  n, the smoothed estimate

X
n
c th D
TC whj ytj ; (7.48)
jD1

where
th tj

K
whj D Pn th tj

b
(7.49)
jD1 K b

are the weights from a parametric kernel which is a nonnegative symmetric function
that integrated over its domain gives unity, and b ¤ 0 is the smoothing parameter.
Kernel smoothers are local functions since their weighting systems are local. In
fact, for any observed value yth , a weighted average is computed. Each weight is
obtained as a function of the distance between the target point th and the tj ’s close
to th that belong to an interval whose amplitude is established by the smoothing
parameter b, otherwise said bandwidth parameter. Increasing the distance between
th and tj ; j D 1; : : : ; n, the weights assigned to the corresponding observations
decrease till a certain point when they become zero. Such a point depends on the
bandwidth parameter that, in practice, determines the lengths of the smoother, i.e.,
the number of observations near yth that have non-null weight.
In matrix form the action of kernel smoother can be represented by the relation

c D Wb ya
TC (7.50)
7.3 Stochastic Local Trend-Cycle Models 185

where Wb is the smoothing matrix of generic element whj as defined in (7.49) for
h; j D 1; : : : ; n.
Once the smoothing parameter has been selected, the whj ’s corresponding to the
target points falling out of the bandwidth of any th turn out to be null, where the
number of no null weights depends on both the value of the bandwidth parameter
and on the number of decimals chosen for each weight. Notice that as long as the
kernel is a symmetric function, the number of no null weights turns out to be odd.
For instance, for a standard Gaussian kernel function
t  t  
h j 1 1  th  tj 2
K D p exp  (7.51)
b 2 2 b

with b = 5 and weights whj taken with three decimals, the smoothing matrix W5 of
no null weights is of size 31  31. For further details, we refer the reader to Dagum
and Luati [15, 16].
In kernel smoothing, the bandwidth parameter is of great importance. Similar to
LOESS, increasing the smoothing parameter from the extreme case of interpolation
(b ! 0) to that of oversmoothing (b ! 1) produces an increase in bias and a
decrease in variance.

7.3.4 Cubic Smoothing Spline

The current literature on spline functions, particularly on smoothing splines, is very

large and we refer the reader to Wahba [47] for an excellent summary of the most
important contributions on this topic. The name is due to the resemblance with the
curves obtained by draftsmen using a mechanical spline, that is, a thin and flexible
rod with weights or “ducks” used to position the rod at points through which it was
desired to draw a smooth interpolating curve.
The problem of smoothing via spline functions is closely related to that of
smoothing priors and signal extraction in time series, where these latter are
approached from a parametric point of view (see, among others, Akaike [1, 2], and
Kitagawa and Gersch [34]).
Similar to the Henderson filters, the original work on smoothing spline functions
was based on the theory of graduation. The first two seminal works are due to
Whittaker [50] and Whittaker and Robinson [49] who proposed a new graduation
method that basically consisted of a trade-off between fidelity and smoothing. The
problem was that of estimating an unknown “smooth” function, in our case the
trend-cycle, TC, observed with errors assumed to be white noise. That is, given
a set of seasonally adjusted observations yt ; t D 1; 2; : : : ; n such that,

yt D TCt C It ; It  i:i:d:.0; I2 /; (7.52)

186 7 Trend-Cycle Estimation

we want to minimize

X
n X
n

2
.yt  TCt /2 C 2 .1  B/k TCt ; (7.53)
tD1 tDkC1

where .1  B/k denotes the k-th order difference of TCt , e.g., .1  B/TCt D TCt 
TCt1 , .1  B/2 TCt D .1  B/..1  B/TCt /, and so on. The smoothing trade-off
parameter  must be appropriately chosen.
Following this direction, Schoenberg [42] extended Whittaker smoothing method
to the fitting of a continuous function to observed data, not necessarily evenly
spaced. In this case, the model is

yt D TC.t/ C It ; It  i:i:d:.0; I2 /; (7.54)

where the unobserved function TC.t/; t 2 Œa; b ; is assumed to be smooth on the

interval Œa; b , and the observations are at the n points, t1 ; t2 ; : : : ; tn . The problem is
to find
Z
1X  2
n b
minm .yt  TC.ti //2 C  TC.m/ .t/ ; (7.55)
TC2C n iD1 i a

where Cm is the class of functions with m continuous derivatives and  > 0.

The solution to (7.55) known as a smoothing spline is unique and given by a
univariate natural polynomial or piecewise polynomial spline of degree 2m  1 with
knots at the data points t1 ; t2 ; : : : ; tT . The smoothing trade-off parameter  controls
the balance between the fit to the data as measured by the residual sum of squares,
and the smoothness as measured by the integrated squared m-th derivative of the
function. When m D 2, which is the case of a cubic smoothing spline, the integral
of the squared second order derivative TC.2/ is the curvature, and a small value for
the integral corresponds visually to a smooth curve. As  ! 0, the solution TC c
tends to the univariate natural polynomial spline which interpolates the data, and as
 ! 1, the solution tends to the polynomial of degree m that represents the best
fitting of the data in the least squares sense. The smoothing trade-off parameter  is
known as hyperparameter in the Bayesian terminology, and it has the interpretation
of a noise to signal ratio: the larger the , the smoother the trend-cycle.
The estimation of  was first done using Ordinary Cross Validation (OCV).
OCV consisted of deleting one observation and solving the optimization problem
with a trial value of , computing the difference between the predicted value and
the deleted observation, accumulating the sums of squares of these differences as
one runs through each of the data points in turn, and finally choosing the  for
which the accumulated sum is the smallest. This procedure was improved by Craven
7.3 Stochastic Local Trend-Cycle Models 187

and Wahba [9] who developed the Generalized Cross Validation (GCV) method
currently available in most computer packages. GCV can be obtained from OCV by
rotating the system to a standard coordinate system, doing OCV, and rotating back.
The GCV estimate of  is obtained by minimizing

.1=T/ Œ.In  W.//y 2

V./ D ; (7.56)
Œ.1=n/tr.In  W.// 2

c that is,
where W./ is the influential matrix associated with TC,
2c 3 2 3
TCt1 yt1
c t2 7
6 TC 6 yt 7
6 7 6 27
6 : 7 D W./ 6 : 7 (7.57)
4 :: 5 4 :: 5
c tn
TC ytn

or in compact form

c D W./y:
TC

The trace of W./ can be viewed as the degrees of freedom for the signal and
so, (7.56) can be interpreted as minimizing the standardized sum of squares of the
residuals. Since  is usually estimated by cross validation, cubic smoothing splines
are considered nonlinear smoothers. However, Dagum and Capitanio [13] showed
that if  is fixed, i.e.,  D 0 , then W.0 / D W0 has the form
"  1 #
1
W0 D In  DT B C DDT D ; (7.58)
0

where B 2 R.n2/.n2/ and D 2 R.n2/T are as follows:

21 3
3 .t3  t1 / 16 .t3  t2 / 0  0
6 1
.t  t2 / 13 .t4  t2 / 16 .t4  t3 /  0 7
6 6 3 7
6 :: :: 7
BD6
6 0 : :   7
7 (7.59)
6 :: :: :: 7
4 : : :   5
1 1
0   .t  tn1 / 3 .tn  tn2 /
6 n
188 7 Trend-Cycle Estimation

and
2 1 1 1 1
3
t2 t1
 t2 t1
 t3 t2 t3 t2
 0
6 :: 7
6 0 1 1
 t3 t  1
: 0 7
6 t3 t2 t4 t3 7
6 :: ::
2
:: 7
DD6
6 : : : 
7
 7 (7.60)
6 :: :: :: 7
6 7
4 : : :   5
1 1 1 1
0  tn1 tn2  tn1 tn2
 tn tn1 tn tn2

Making tj D j; j D 1; : : : ; n, (7.59) and (7.60) become, respectively,

22 3 2 3
1
0  0 1 2 1    0
3 6
6 6 : 7
6
1
6
2
3
1
6
 0 7 7 6 0 1 2 :: 0 7
6 :: :: 7 6 7
6 7
BD6
60 : :      77 and D D 6 ::: ::: :::       7 : (7.61)
6: :: :: 7 6 7
4 :: 6: : : 7
: :      5 4 :: :: ::       5
0       16 23 0  0  1

7.3.5 Theoretical Properties of Symmetric and Asymmetric

Linear Trend-Cycle Filters

The theoretical properties of the symmetric and asymmetric linear filters can be
studied by analyzing their frequency response functions. It is defined by

X
m
H.!/ D wj ei!j ; 0  !  1=2; (7.62)
jDm

where wj are the weights of the filter and ! is the frequency in cycles per unit of
time. In general, the frequency response functions can be expressed in polar form as
follows:

H.!/ D A.!/ C iB.!/ D G.!/ei.!/ ; (7.63)

 1=2
where G.!/ D A2 .!/ C B2 .!/ is called the gain of the filter and .!/ D
tan1 .B.!/=A.!// is called the phase shift of the filter and is usually expressed
in radians. The expression (7.63) shows that if the input function is a sinusoidal
variation of unit amplitude and constant phase shift .!/, the output function will
also be sinusoidal but of amplitude G.!/ and phase shift .!/ C .!/. The gain
and phase shift vary with !. For symmetric filters the phase shift is 0 or ˙, and
for asymmetric filters take values between ˙ at those frequencies where the gain
7.3 Stochastic Local Trend-Cycle Models 189

function is zero. For a better interpretation the phase shifts will be here given in
months instead of radians, that is, given by .!/=2! for ! ¤ 0.
The gain function should be interpreted as relating the spectrum of the original
series to the spectrum of the output obtained with a linear time invariant filter. For
example, let yO t be the estimated seasonally adjusted observation for the current
period based on data y1 ; y2 ; : : : ; yn by application of the concurrent linear time-
invariant filter w0;j . Thus, the gain function shown in the below figures relates the
spectrum of fyt ; t D 1; : : : ; ng to the spectrum of yO t , and not to the spectrum of
the complete seasonally adjusted series produced at time n (which includes yt , a
first revision of time t  1, a second revision at time t  2, and so on). Dagum
et al. [14] derived the gain functions of standard, short, and long convolutions
corresponding to the 13-term (H13), 9-term (H9), and 23-term (H23) symmetric
Henderson filters, respectively. These authors showed how cycles of 9 and 10
months periodicity (in the 0.08–0.18 frequency band) are not suppressed by any
of the cascade filters, particularly, those using H13 and H9 filters. In fact, about
90 %, 72 %, and 21 % of the power of these short cycles are left in the output by
the 9-, 13-, and 23-term Henderson filters, respectively. Figure 7.3 shows the gain
functions of four symmetric trend-cycle filters of 13-term each. In the context of
trend-cycle estimation, it is useful to divide the total range of ! 2 Œ0; 0:50 into
two major intervals, one, for the signal, and another for the noise. There are no-
fixed rules on defining the cut-off frequency, but for monthly data the intervals are
1.2

Cubic spline

Gaussian kernel

Henderson filter

LOESS
0.8
G( )
0.4
0.0

0.0 0.1 0.2 0.3 0.4 0.5

Fig. 7.3 Gain functions G.!/; 0  !  0:5, of several symmetric 13-term filters
190 7 Trend-Cycle Estimation

usually given by:

1. 0  !  0:06 associated with cycles of 16 months and longer attributed to the
trend-cycle of the series, and
2. 0:06  !  0:50 corresponding to seasonality, if present, and noise.
It is apparent that the 13-term Henderson and LOESS are very close to each other
and pass well the power in the frequency band 0  !  0:06 reproducing well the
cycles associated with the trend-cycle and suppressing a large amount of noise. But
they have the limitation of passing too much power at ! D 0:10, that will produce
a large number of 10-month cycles in the output, also known as unwanted ripples.
On the other hand, the Gaussian kernel will not pass too much power at ! D
0:10, but will suppress also the power attributed to the trend-cycle. In other words,
it will produce a very smooth trend-cycle whereas the 13-term cubic spline will
do the opposite that is why in empirical applications the length of this filter is
selected according to other criteria. Figures 7.4 and 7.5 exhibit the gain and phase
shifts of the last point of the corresponding asymmetric filters. The gains of the
asymmetric filters suppress less noise and pass more power related to the trend-
cycle (particularly, H13 and LOESS amplify the power) relative to the symmetric
filters. All the filters introduce a time lag to detect a true turning point as shown by
their phase shift functions.
1.2
0.8
G( )

Cubic spline
0.4

Gaussian kernel

Henderson filter

LOESS
0.0

0.0 0.1 0.2 0.3 0.4 0.5

Fig. 7.4 Gain functions G.!/; 0  !  0:5, of last point asymmetric filters
7.4 Illustrative Results 191

1.5
Cubic spline

Gaussian kernel

Henderson filter

LOESS
0.5
( )
−0.5
−1.5

0.0 0.1 0.2 0.3 0.4 0.5

Fig. 7.5 Phase shift functions .!/; 0 < !  0:5, of last point asymmetric filters

7.4 Illustrative Results

We now illustrate with a real series the various trend-cycle values obtained with the
different filters discussed before. To select the appropriate lengths of the Henderson,
LOESS, and Gaussian kernel filters we use the irregular/trend-cycle (I=C) ratio,
whereas for the cubic spline we look at the smoothing parameter estimated by means
of the generalized cross validation criterion. In particular, we consider the US New
Orders for Durable Goods (NODG) observed from February 1992 till December
2013.
Figures 7.6, 7.7, and 7.8 exhibit the various trend-cycle estimates for the US
New Orders for Durable Goods series, where 13-term filters have been chosen for
the Henderson, the LOESS, and the Gaussian kernel and the smoothing parameter
for the cubic spline is 16.438 and the equivalent degrees of freedom is 47.289.
Figure 7.6 shows that the LOESS and Henderson filter give values close to each
other whereas the Gaussian kernel (Fig. 7.7) produces the smoothest trend, and the
cubic spline (Fig. 7.8) gives the most volatile. The difference is more noticeable at
cyclical turning points where the Gaussian kernel cuts them more whereas the cubic
spline cuts them the least.
192 7 Trend-Cycle Estimation

250000
NODG series
Henderson
LOESS
150000

Jan 1995 Jan 2000 Jan 2005 Jan 2010

Fig. 7.6 US New Orders for Durable Goods trend-cycle estimates based on LOESS and Hender-
son 13-term filters
250000

NODG series
Gaussian kernel
150000

Jan 1995 Jan 2000 Jan 2005 Jan 2010

Fig. 7.7 The US New Orders for Durable Goods trend-cycle estimates based on the 13-term
Gaussian kernel
References 193

250000
NODG series
200000
150000
100000 Cubic smoothing spline

Jan 1995 Jan 2000 Jan 2005 Jan 2010

Fig. 7.8 The US New Orders for Durable Goods trend-cycle estimates based on cubic smoothing
spline

References

1. Akaike, H. (1980). Likelihood and the Bayes procedure (with discussion). In J. M. Bernardo,
M. H. De Groot, D. V. Lindley, & A. F. M. Smith (Eds.), Bayesian statistics (pp. 143–203).
Valencia, Spain: University Press.
2. Akaike, H. (1981). Likelihood of a model and information criteria. Journal of Econometrics,
16, 3–14.
3. Anderson, T. W. (1971). The statistical analysis of time series. New York: Wiley.
4. Box, G. E. P., & Jenkins, G. M. (1970). Time series analysis: Forecasting and control. San
Francisco, CA, USA: Holden-Day (Second Edition 1976).
5. Burman, J. P. (1980). Seasonal adjustment by signal extraction. Journal of the Royal Statistical
Society, Series A, 143, 321–337.
6. Cleveland, R. B., Cleveland, W. S., McRae, J. E., & Terpenning, I. (1990). STL: A seasonal
trend decomposition procedure based on LOESS. Journal of Official Statistics, 6(1), 3–33.
7. Cleveland, W. S. (1979). Robust locally regression and smoothing scatterplots. Journal of the
American Statistical Association, 74, 829–836.
8. Cleveland, W. S., & Devlin, S. J. (1988). Locally weighted regression: An approach to
regression analysis by local fitting. Journal of the American Statistical Association, 83(403),
596–610.
9. Craven, P., & Wahba, G. (1979). Smoothing noisy data with spline functions. Numerical
Mathematics, 31, 377–403.
10. Dagum, C. (1985). Analyses of income distribution and inequality by education and sex in
Canada. In R. L. Basmann & G. F. Rhodes Jr. (Eds.), Advances in econometrics (Vol. IV, pp.
167–227). Greenwich, CN: JAI Press.
11. Dagum, E. B. (1980). The X11ARIMA seasonal adjustment method. Ottawa, Canada: Statistics
Canada. Catalogue No. 12-564.
194 7 Trend-Cycle Estimation

12. Dagum, E. B. (1988). The X11ARIMA/88 seasonal adjustment method-foundations and user’s
manual. Ottawa, Canada: Time Series Research and Analysis Centre, Statistics Canada.
13. Dagum, E. B., & Capitanio, A. (1998). Smoothing methods for short-term trend analysis: Cubic
splines and Henderson filters. Statistica, LVIII(1), 5–24.
14. Dagum, E. B., Chhab, N., & Chiu, K. (1996). Derivation and analysis of the X11ARIMA and
Census X11 linear filters. Journal of Official Statistics, 12(4), 329–347.
15. Dagum, E. B., & Luati, A. (2000). Predictive performance of some nonparametric linear and
nonlinear smoothers for noisy data. Statistica, LX(4), 635–654.
16. Dagum, E. B., & Luati, A. (2001). A study of asymmetric and symmetric weights of kernel
smoothers and their spectral properties. Estadistica, Journal of the Inter-American Statistical
Institute, 53(160–161), 215–258.
17. Doherty, M. (2001). The surrogate Henderson filters in X-11. Australian and New Zealand
Journal of Statistics, 43(4), 385–392.
18. Eubank, R. L. (1999). Spline smoothing and nonparametric regression. New York: Marcel
Dekker.
19. Fan, J. (1992). Design-adaptive nonparametric regression. Journal of the American Statistical
Association, 87, 998–1004.
20. Fan, J. (1993). Local linear regression smoothers and their minimax efficiencies. Annals of
Statistics, 21, 196–216.
21. Fan, J., & Gijbels, I. (1996). Local polynomial modelling and its applications. New York:
Chapman and Hall.
22. Findley, D. F., Monsell, B. C., Bell, W. R., Otto, M. C., & Chen, B. C. (1998). New capabilities
and methods of the X12ARIMA seasonal adjustment program. Journal of Business and
Economic Statistics, 16(2), 127–152.
23. Granger, C. W., & Morgenstern, O. (1970). Predictability of stock market prices. Lexington,
MA: D.C. Heath.
24. Gray, A., & Thomson, P. (1996). Design of moving-average trend filters using fidelity and
smoothness criteria. In P. M. Robinson & M. Rosenblatt (Eds.), Time series analysis (in
memory of E.J. Hannan), Vol. II. Springer Lecture notes in statistics (Vol. 115, pp. 205–219).
New York: Springer.
25. Green, P. J., & Silverman, B. W. (1994). Nonparametric regression and generalized linear
models. London: Chapman and Hall.
26. Hardle, W. (1990). Applied nonparametric regression. Cambridge: Cambridge University
Press.
27. Harvey, A. C. (1985). Trends and cycles in macroeconomic time series. Journal of Business
and Economic Statistics, 3, 216–227.
28. Harvey, A. C. (1989). Forecasting, structural time series models and the Kalman filter.
Cambridge: Cambridge University Press.
29. Harvey, A. C., & Todd, P. H. J. (1983). Forecasting economic time series with structural and
Box-Jenkins models: A case study. Journal of Business and Economic Statistics, 1, 299–306.
30. Henderson, R. (1916). Note on graduation by adjusted average. Transaction of Actuarial
Society of America, 17, 43–48.
31. Hillmer, S. C., & Tiao, G. C. (1982). An ARIMA-model-based approach to seasonal adjust-
ment. Journal of the American Statistical Association, 77, 63–70.
32. Kendall, M. G., Stuart, A., & Ord, J. (1983). The advanced theory of statistics (Vol. 3).
New York: Wiley.
33. Kenny, P., & Durbin, J. (1982). Local trend estimation and seasonal adjustment of economic
and social time series. Journal of the Royal Statistical Society A, 145, 1–41.
34. Kitagawa, G., & Gersch, W. (1996). Smoothness priors analysis of time series. Lecture notes
in statistics (Vol. 116). New York: Springer.
35. Kondratieff, N. (1925). Long economic cycles. Voprosy Konyuktury (Vol. 1, No. (1)) (English
translation: The long wave cycle. Richardson and Snyder: New York, 1984).
36. Laniel, N. (1985). Design criteria for the 13-term Henderson end weights. Working paper.
Methodology Branch. Statistics Canada, Ottawa.
References 195

37. Macauley, F. (1931). The smoothing of time series. New York: National Bureau of Economic
Research.
38. Maravall, A. (1993). Stochastic linear trends: Models and estimators. Journal of Econometrics,
56, 5–37.
39. Musgrave, J. (1964). A set of end weights to end all end weights. Working paper. U.S. Bureau
of Census, Washington, DC.
40. Quetelet, A. (1835). Sur l’homme et le développement de ses facultés. Paris: Bachelier.
41. Sbrana, G. (2011). Structural time series models and aggregation: Some analytical results.
Journal of Time Series Analysis, 32, 315–316.
42. Schoenberg, I. (1964). Monosplines and quadrature formulae. In T. Greville (Ed.), Theory and
applications of spline functions. Madison, WI: University of Wisconsin Press.
43. Shiskin, J., Young, A. H., & Musgrave, J. C. (1967). The X-11 variant of the Census Method
II seasonal adjustment program. Technical Paper 15 (revised). US Department of Commerce,
Bureau of the Census, Washington, DC.
44. Simonoff, J. S. (1995). Smoothing methods in statistics. New York: Springer.
45. Stock, J. H., & Watson, M. W. (1988). Variable trends in economic time series. Journal of
Economic Perspectives, 2, 147–174.
46. Verhulst, P. F. (1838). Notice sur la loi que la population suit dans son accroissement. In A.
Quetelet (Ed.), Correspondance Mathematique et Physique (Tome X, pp. 113–121).
47. Wahba, G. (1990). Spline models for observational data. Philadelphia: SIAM.
48. Wand, M., & Jones, M. (1995). Kernel smoothing. Monographs on statistics and applied
probability (Vol. 60). London: Chapman and Hall.
49. Whittaker, E., & Robinson, G. (1924). Calculus of observations: A treasure on numerical
calculations. London: Blackie and Son.
50. Whittaker, E. T. (1923). On a new method of graduation. In Proceedings of the Edinburg
Mathematical Association (Vol. 78, pp. 81–89).
51. Yaglom, A. M. (1962). An introduction to the theory of stationary random functions.
Englewood Cliffs, NJ: Prentice-Hall.
Chapter 8
Further Developments on the Henderson
Trend-Cycle Filter

Abstract The linear filter developed by Henderson is the most widely applied to
estimate the trend-cycle component in seasonal adjustment software such as the US
Bureau of Census II-X11 and its variants, the X11/X12ARIMA. Major studies have
been done on trend-cycle estimation during the last 20 years by making changes
to the Henderson filters. The emphasis has been on determining the direction of
the short-term trend for an early detection of a true turning point. This chapter
introduces in detail three major contributions: (1) a nonlinear trend-cycle estimator
also known as Nonlinear Dagum Filter (NLDF), (2) a Cascade Linear Filter (CLF)
that closely approximates the NLDF, and (3) an approximation to the Henderson
filter via the Reproducing Kernel Hilbert Space (RKHS) methodology.

The linear filter developed by Henderson [22] is the most widely applied to estimate
the trend-cycle component in seasonal adjustment software such as the US Bureau
of Census II-X11 method [30] and its variants, the X11ARIMA [8] and X12ARIMA
[20]. Major studies have been done on trend-cycle estimation during the last 20
years by making changes to the Henderson filters. The emphasis has been made
on determining the direction of the short-term trend for an early detection of a
true turning point. Major changes of global character in the financial and economic
sectors have introduced high levels of variability in time series making difficult to
detect the direction of the short-term trend by simply looking at seasonally adjusted
data and the use of trend-cycle data has been supported.
In 1996, Dagum [9] developed a nonlinear trend-cycle estimator to improve on
the classical 13-term Henderson filter [22]. The Nonlinear Dagum Filter (NLDF)
results from applying the 13-term symmetric Henderson filter (H13) to seasonally
adjusted series where outliers and extreme observations have been replaced and
which have been extended with extrapolations from an ARIMA model [3]. Later
on, in 2009, Dagum and Luati [16] developed a linear approximation to the NLDF
to facilitate a broader application since the new linear approximation does not need
an ARIMA model identification.
From another perspective, in 2008 Dagum and Bianconcini [11] developed an
approximation to the Henderson filter via reproducing kernel methodology that was
shown to produce better results for real time trend-cycle estimation. All three papers
make explicit reference to the Henderson filters applied to monthly time series

© Springer International Publishing Switzerland 2016 197

E. Bee Dagum, S. Bianconcini, Seasonal Adjustment Methods and Real Time
Trend-Cycle Estimation, Statistics for Social and Behavioral Sciences,
DOI 10.1007/978-3-319-31822-6_8
198 8 Further Developments on the Henderson Trend-Cycle Filter

but the results can be easily extended to quarterly series. It should be noticed that
following the tradition of statistical agencies, in these papers it is called asymmetric
Henderson filter the one that corresponds to the asymmetric weights developed
by Musgrave [27]. In the next sections, we will discuss the major assumptions
and modifications introduced in each of these studies and refer the reader to the
published papers for more details.

8.1 The Nonlinear Dagum Filter (NLDF)

The modified Henderson filter developed by Dagum [9] is nonlinear and basically
consists of:
(a) extending the seasonally adjusted series with ARIMA extrapolated values, and
(b) applying the 13-term Henderson filter to the extended series where extreme
values have been modified using very strict sigma limits.
To facilitate the identification and fitting of simple ARIMA models, Dagum [9]
recommends, at step (a), to modify the input series for the presence of extreme
values using the standard ˙2:5 limits of X11/X12ARIMA and X13. These
computer programs are those that can be used to implement the Henderson filter.
In this way, a simple and very parsimonious ARIMA model, the ARIMA(0,1,1), is
often found to fit a large number of series.
Concerning step (b), it is recommended to use very strict sigma limits, such as
˙0:7 and ˙1:0. The main purpose of the ARIMA extrapolations is to reduce the
size of the revisions of the most recent estimates, whereas that of extreme values
replacement is to reduce the number of unwanted ripples produced by H13. An
unwanted ripple is a 10-month cycle (identified by the presence of high power at
! D 0:10 in the frequency domain) which, due to its periodicity, often leads to
the wrong identification of a true turning point. In fact, it falls in the neighborhood
between the fundamental seasonal frequency and its first harmonic. On the other
hand, a high frequency cycle is generally assumed to be part of the noise pertaining
to the frequency band 0:10  ! < 0:50. The problem of the unwanted ripples is
specific of H13 when applied to seasonally adjusted series.
The NLDF can be formally described in matrix notation as follows. Let y 2 Rn
be the n-dimensional seasonally adjusted time series to be smoothed, which consists
of a nonstationary trend-cycle TC plus an erratic component e, that is,

y D TC C e: (8.1)
8.1 The Nonlinear Dagum Filter (NLDF) 199

It is assumed that the trend-cycle is smooth and can be well estimated by means of
the 13-term Henderson filter applied to y. Hence,

c D Hy;
TC (8.2)

where H is the n  n matrix (canonically) associated with the 13-term Henderson

c in Eq. (8.1) by Eq. (8.2), we have
filter. Replacing TC

y D Hy C e; (8.3)

or,

.In  H/y D e; (8.4)

where In is the n  n identity operator on Rn . Assign now a weight to the residuals

in such a way that if the observation yt ; t D 1; : : : ; n, is recognized to be an
extreme value (with respect to ˙2:5 limits, where  is a 5-year moving standard
deviation), then the corresponding residual et is zero weighted (i.e., the extreme
c t which is a preliminary estimate of the trend). If yt is not
value is replaced by TC
an extreme value, then the weight for et is one (i.e., the value yt is not modified). In
symbols,

W0 e D W0 .In  H/y; (8.5)

where W0 is a zero-one diagonal matrix, being the diagonal element wtt equal to
zero when the corresponding element yt of the vector y is identified as an outlier.
For instance, if in the series y the only extreme value is y2 , then the weight matrix
for the residuals will be
2 3
1 0 0  0
6 0 0 0  0 7
6 7
6 7
W0 D 6 0 0 1    0 7 : (8.6)
6 7
4              5
0 0 0  1

Denoting by

e0 D W0 e (8.7)

the vector of the modified residuals, then the series modified by extreme values with
zero weights becomes

y0 D TC C e0 ; (8.8)
200 8 Further Developments on the Henderson Trend-Cycle Filter

which can be written as

y0 D Hy C W0 .In  H/y D ŒH C W0 .In  H/ y: (8.9)

Using (8.9), one year of ARIMA extrapolations is obtained in order to extend the
series modified by extreme values. Denoting with yE0 the extended series, that is,
the n C 12 vector whose first n elements are given by y0 while the last 12 are the
extrapolated ones, in block-matrix notation we have
 
ŒH C W0 .In  H/ y
yE0 D ; (8.10)
y12

where y12 is the 12  1 block of extrapolated values. Setting

 
ŒH C W0 .In  H/ On12
ŒH C W0 .In  H/ 12 D ; (8.11)
O12n I12

and
 
C;12 y
y D 12 ; (8.12)
y

yE0 becomes

yE0 D ŒH C W0 .In  H/ 12 yC;12 : (8.13)

This concludes the operations involved in step (a) of the NLDF.

Step (b) follows. The procedure for obtaining y0 on the series yE0 is repeated, but
with stricter sigma limits (such as ˙0:7 and ˙1:0) and with different weights
assigned to the residuals for the replacement of the extreme values. The estimates
yE computed over the series yE0 are

yE D ŒH C W.In  H/ E yE0 : (8.14)

The .nC12/.nC12/ matrix ŒH C W.In  H/ E is analogue to ŒH C W0 .In  H/ 12

except for the matrix W that is also diagonal, but with generic diagonal element
wtt , such that wtt D 0 if the corresponding value yt falls out of the upper bound
selected limits, say, ˙1:0, and wtt D 1 if the corresponding yt falls within the lower
bound selected limits, say, ˙0:7 and wtt decreases linearly (angular coefficient
equal to 1) from 1 to 0 in the range from ˙0:7 to ˙1:0. Under the assumption
of normality, these sigma limits imply that 48 % of the values will be modified
(replaced by the preliminary smoothed trend), 32 % will be zero weighted while the
remaining 16 % will get increasing weights from zero to one.
8.2 The Cascade Linear Filter 201

Notice that yE can also be written as

yE D ŒH C W.In  H/ E ŒH C W0 .In  H/ 12 yC;12 : (8.15)

Finally, the NLDF estimates are given by applying a 13-term Henderson filter to
Eq. (8.15), that is,

HyE D H ŒH C W.In  H/ E ŒH C W0 .In  H/ 12 yC;12 (8.16)

" #
c
TC
D ;
c 12
TC

where TCc is the n-dimensional vector of smooth estimates of y.

It is apparent that the NLDF method reduces drastically the effects of extreme
values by repeatedly smoothing the input data via downweighting points with large
residuals. Furthermore, the ARIMA extension enables the use of the symmetric
weights of the 13-term Henderson filter for the last six observations and, thus,
reduces the size of the revisions of the last estimates.
Since the values of the matrices W0 and W corresponding to extreme values
replacement and matrix ŒH C W0 .In  H/ 12 pertaining to ARIMA extrapolations
are data dependent, this filter is nonlinear. Studies by Dagum, Chhab, and Morry
[5, 12], and Darnè [18] showed the superior performance of the NLDF with respect
to both structural and ARIMA standard parametric trend-cycle models applied to
series with different degrees of signal-to-noise ratios. The criteria evaluated were:
1. the number of unwanted ripples,
2. size of revisions, and
3. time delay to detect a turning point.
In another study, the good performance of the NLDF is shown relative to
nonparametric smoothers, namely: locally weighted regression (LOESS), Gaussian
kernel, cubic smoothing spline, and supersmoother [14]. The NLDF is currently
used by many statistical agencies to produce trend-cycle estimates or trend-cycle
figures to show the direction of the short-term trend.

8.2 The Cascade Linear Filter

Given the excellent performance of the NLDF according to the three criteria
mentioned above, Dagum and Luati [16] developed a cascade linear filter that
closely approximates it. The cascading is done via the convolution of several filters
chosen for noise suppression, trend estimation, and extrapolation.
A linear filter offers many advantages over a nonlinear one. First, its application
is direct and hence, does not require knowledge of ARIMA model identification.
202 8 Further Developments on the Henderson Trend-Cycle Filter

Furthermore, linear filtering preserves the crucial additive constraint by which the
trend of an aggregated variable should be equal to the algebraic addition of its
component trends, thus avoiding the selection problem of direct versus indirect
adjustments. Finally, the properties of a linear filter concerning signal passing and
noise suppression can always be compared to those of other linear filters by means
of spectral analysis.
The symmetric filter is the one applied to all central observations. In this case,
the purpose is to offer a linear solution to the unwanted ripples problem. To avoid
the latter, the NLDF largely suppresses the noise in the frequency band between
the fundamental seasonal and first harmonic. In this regard, a cascade linear filter
is derived by double smoothing the residuals obtained from a sequential application
of H13 to the input data. The residuals smoothing is done by the convolution of two
short smoothers, a weighted 5-term and a simple 7-term linear filters. The linear
approximation for the symmetric part of the NLDF is truncated to 13 terms with
weights normalized to add to one.
On the other hand, the asymmetric filter is applied to the last six data points
which are crucial for current analysis. It is obtained by means of the convolution
between the symmetric filter and the linear extrapolation filters for the last six data
points. The extrapolation filters are linearized by fixing both the ARIMA model and
its parameter values. The latter are chosen so as to minimize the size of revisions and
phase shifts. The model was selected among some parsimonious processes found to
fit and extrapolate well a large number of seasonally adjusted series. Such model
is the ARIMA(0,1,1) with  D 0:40. A simple linear transformation [15] allows to
apply the asymmetric filter to the first six observations.
The new filter is called the Cascade Linear Filter (CLF) and there is a distinction
between the Symmetric (SLF) and the Asymmetric Linear Filter (ALF).

8.2.1 The Symmetric Linear Filter

The smoothing matrix associated with the symmetric linear filter results
  
H H C M7.0:14/ .In  H/ H C M5.0:25/ .In  H/ ; (8.17)

where M5.0:25/ is the matrix representative of a 5-term moving average with weights
.0:250; 0:250; 0:000; 0:250; 0:250/, and M7.0:14/ is the matrix representative of a 7-
term filter with all weights equal to 0:143.
5- and 7-term filters are chosen following the standard filters length selected
in Census II-X11 and X11/X12ARIMA software for the replacement of extreme
Seasonal-Irregular (SI) values. In these computer packages, a first iteration is made
by means of a short smoother, a 5-term (weighted) moving average, and a second
one by a 7-term (weighted) average. In this case, 5- and 7-term filters are applied
to the residuals from a first pass of the H13 filter. These two filters have the
good property of suppressing large amounts of power at the frequency ! D 0:10,
8.2 The Cascade Linear Filter 203

Fig. 8.1 Gain function of the symmetric weights of the modified irregulars filter

corresponding to cycle of 10 months periodicity.

 Figure 8.1 shows  the gain function
of the filter convolution M7.0:14/ .In H/ H C M5.0:25/ .In H/ .
It is apparent that the filter convolution applied to a series consisting of trend plus
irregulars suppresses all the trend power and a great deal of the irregular variations.
Hence given the input series, the results from the convolution are the modified
irregulars needed to produce a new series which will be extended with ARIMA
extrapolations, and then smoothed with H13.
Equation (8.17) produces a symmetric filter of 31 terms with very small weights
at both ends. This long filter is truncated to 13 terms, and normalized such that its
weights add up to unity. Normalization is needed to avoid a biased mean output.
To normalize the filter, the total weight discrepancy (in our case the 13 truncated
weights add up to 1.065) is distributed over the 13 weights, wj , j D 6; : : : ; 6,
according to a well-defined pattern. This is a very critical point since adjustment for
different distributions produces linear filters with very distinctive properties.
The weights of the symmetric linear filter with mixed distribution normalization,

denoted by SLF, are given by 0:027; 0:007; 0:031; 0:067; 0:136; 0:188; 0:224 .

8.2.1.1 Theoretical Properties of SLF

From a theoretical viewpoint, the properties of SLF were studied by means of

classical spectral analysis techniques. It was found that in comparison to H13,
SLF suppressed more signal only in the frequency band pertaining to very short
cycles, ranging from 15 to 24 months periodicity (of corresponding frequency
0:03 < !  0:06), whereas it passed without modification cycles of 3 year
and longer periodicity. Furthermore, it reduced by 14 % the power of the gain
204 8 Further Developments on the Henderson Trend-Cycle Filter

corresponding to the unwanted ripples frequency ! D 0:10. We refer the reader

to Dagum and Luati [16] for further details.

8.2.1.2 Empirical Properties of SLF

To perform an empirical evaluation of the fitting and smoothing performances of

SLF relative to H13, applications were made to a large sample of 100 seasonally
adjusted real time series pertaining to various socioeconomic areas, characterized
by different degrees of variability. As a measure of fitting the Root Mean Square
Error (RMSE) is used, and defined as follows:
v !2
u
u 1 X TC
n5 c t  yt
RMSE D t ;
n  12 tD7 yt

c t the estimated trend-

where yt denotes the original seasonally adjusted series, and TC
cycle values. The main goal was to assess the extent to which the fitting properties
of the SLF were equivalent to those of H13, which is known to be a good signal
estimator but at the expense of producing many unwanted ripples. There was also
the interest to reduce the number of unwanted ripples that may lead to the detection
of false turning points. In this regard, the accepted definition of a turning point for
smoothed data (see, among others, Zellner et al. [32]) is used. A turning point occurs
at time t if (downturn) ytk      yt1 > yt  ytC1      ytCm or (upturn)
ytk      yt1 < yt  ytC1      ytCm for k D 3 and m D 1. An unwanted
ripple arises whenever two turning points occur within a 10-month period.
Table 8.1 shows the mean values of the RMSE calculated over the 100 series and
standardized with respect to H13 to facilitate the comparison. In the same way, the
number of false turning points produced in the final estimates from SLF is given
relative to that of H13. The empirical results were consistent with those inferred
from the theoretical analysis. Furthermore, SLF reduced by 20 % the number of
false turning points produced by H13.

Table 8.1 Empirical RMSE and the number of false turning points (ftp) for SLF and H13 filters
applied to real time series (mean values standardized by those of H13)
Empirical measures of fitting and smoothing SLF H13
RMSE=RMSEH13 1.05 1
ftp=ftpH13 0.80 1
8.2 The Cascade Linear Filter 205

8.2.2 The Asymmetric Linear Filter

The smoothing matrix associated with the asymmetric linear filter for the last six
data points is obtained in two steps:
1. A linear extrapolation filter for six data points is applied to the input series. This
filter is represented by a .n C 6/  n matrix A
" #
 In
A D ;
O6n12 ˘ 612

where ˘ 612 is the submatrix containing the weights for the n  5; n  4; : : : ; n

data points. ˘ 612 results from the convolution
 E  
H H C M7.0:14/ .InC12 H/ A H C M5.0:25/ .In H/ ; (8.18)
 
where H C M5.0:25/ .In H/ is the n  n matrix representative of trend filter
 E
plus a first suppression of extreme values, H C M7.0:14/ .InC12 H/ is the
n  .n C 12/ matrix for the second suppression of the irregulars applied to the
input series plus 12 extrapolated values, generated by
 
In
AD :
˘ 12n

This .n C 12/  n matrix A is associated with an ARIMA(0,1,1) linear

extrapolations filter with parameter value  D 0:40.
It is well-known that for all
ARIMA models that
admit a convergent AR(1)
representation of the form 1  1 B  2 B2     yt D at , the coefficients
j ; j D 1; 2; : : : can be explicitly calculated. For any lead time , the
extrapolated values yt ./ may be expressed as a linear function of current and
. /
past observations yt with weights j that decrease rapidly as they depart from
the current observations. That is,
1
X . /
yt ./ D j ytjC1 ;
jD1

where
 1
X
. / . h/
j D jC 1 C h j ;
hD1
206 8 Further Developments on the Henderson Trend-Cycle Filter

.1/
for j D 1; 2; : : : and j D j . Strictly speaking, the yt ’s go back to infinite
past, but since the power series is convergent, their dependence on ytj can be
. /
ignored after some time has elapsed. For this reason, j D 0 for j > 12.
Furthermore, to generate one year of ARIMA extrapolation  D 12 is fixed. In
terms of signal passing and noise suppression the model chosen is  D 0:40.
For an ARIMA(0,1,1) model, the coefficients are j D .1  /  j1 , and it can
. /
be shown that j D j , 8. Hence, ˘ 12n is the matrix whose generic row is
(0, . . . , 0, 0.001, 0.002, 0.006, 0.015, 0.038, 0.096, 0.24, 0.6).
Because only six observations are needed to be extrapolated, the centered
12-term filter is truncated and the weights are uniformly normalized to obtain
the 6  12 matrix ˘ 612 given by

–0,023 –0,004 0,004 0,086 0,143 0,187 0,200 0,180 0,130 0,071 0,021 –0,024

0,000 –0,023 –0,004 0,033 0,088 0,148 0,186 0,195 0,160 0,116 0,085 0,017
0,000 0,000 –0,023 –0,005 0,035 0,093 0,146 0,180 0,166 0,155 0,165 0,085

0,000 0,000 0,000 –0,024 –0,003 0,037 0,090 0,141 0,148 0,182 0,255 0,173

0,000 0,000 0,000 0,000 –0,021 –0,006 0,034 0,089 0,114 0,196 0,331 0,264

0,000 0,000 0,000 0,000 0,000 –0,032 –0,009 0,039 0,075 0,200 0,386 0,342

2. The symmetric filter is applied to the series extrapolated by A , that is,

y D SA y
b

where S is the n  .n C 6/ matrix given by

  
H H C M7.0:14/ .In  H/ H C M5.0:25/ .In H/ :

The convolution SA produces 12-term asymmetric filters for the last six
observations, that are truncated and uniformly normalized in order to obtain the
following final Asymmetric Linear Filters (ALF) for the last observations:

–0,026 –0,007 0,030 0,065 0,132 0,183 0,219 0,183 0,132 0,066 0,030 –0,006

0,000 –0,026 –0,007 0,030 0,064 0,131 0,182 0,218 0,183 0,132 0,065 0,031
0,000 0,000 –0,025 –0,004 0,034 0,069 0,137 0,187 0,222 0,185 0,131 0,064

0,000 0,000 0,000 –0,020 0,006 0,046 0,083 0,149 0,196 0,226 0,184 0,130

0,000 0,000 0,000 0,000 0,001 0,033 0,075 0,108 0,167 0,205 0,229 0,182

0,000 0,000 0,000 0,000 0,000 0,045 0,076 0,114 0,134 0,182 0,218 0,230
8.2 The Cascade Linear Filter 207

Hence, the asymmetric filters for the last six data points result from the convolution
of:
1. the asymmetric weights of an ARIMA(0,1,1) model with  D 0:40,
2. the weights of M7.0:14/ and M5.0:25/ filters repeatedly used for noise suppression,
and
3. the weights of the final linear symmetric filter SLF.

8.2.2.1 Theoretical Properties of ALF

The convergence pattern of the asymmetric filters corresponding to H13 and ALF
is shown in Figs. 8.2 and 8.3, respectively. It is evident that the ALF asymmetric
filters are very close one another, and converge faster to the SLF relative to H13.
The distance of each asymmetric filter with respect to the symmetric one gives an
indication of the size of the revisions due to filter changes, when new observations
are added to the series. Figure 8.4 shows that the gain function of the last point
ALF does not amplify the signal as H13 and suppresses significantly the power
at the frequency ! < 0:10. From the viewpoint of the gain function, the ALF
is superior to H13 concerning the unwanted ripples problem as well as in terms
of faster convergence to the corresponding symmetric filter which implies smaller
revisions.
On the other hand, the phase shift of the last point ALF is much greater (near
2 months at very low frequencies) relative to H13 as exhibited in Fig. 8.5. For the

Fig. 8.2 Gain functions showing the convergence pattern of H13 asymmetric weights to the
symmetric H13
208 8 Further Developments on the Henderson Trend-Cycle Filter

Fig. 8.3 Gain function showing the convergence pattern of the ALF asymmetric weights to the
symmetric SLF

Fig. 8.4 Gain functions of the last point filter of H13 and the last point ALF

remaining asymmetric filters, the differences are much smaller. Nevertheless, the
impact of the phase shift in a given input cannot be studied in isolation of the
corresponding gain function. It is well-known that a small phase shift associated
8.2 The Cascade Linear Filter 209

Fig. 8.5 Phase shifts of the last point filter of H13 and last point ALF

Table 8.2 Mean values of Absolute size of revisions ALF H13

revisions of last and previous ˇ ˇ
ˇDs;l ˇ 0.87 1
to the last point asymmetric
filters of ALF and H13 D2s;l 0.42 1
applied to a large sample of Mean squared size of revisions 1
ˇ ˇ
real time series ˇDs;p ˇ 0.60 1
D2s;p 0.11 1

with frequency gain amplifications may produce as poor results as a much larger
phase shift without frequency gain amplifications.

8.2.2.2 Empirical Properties of ALF

The performance of last and previous to the last point ALF was evaluated empir-
ically using a large sample characterized by different degrees of variability. The
purpose is to reduce the size of the revision of these ALF with respect to the
asymmetric H13 filter. We denote with jDs;l j the mean absolute revision between
the final estimates obtained with the symmetric filter, yO sk , and the last point estimate
obtained with the asymmetric ˇ filter
ˇ for the last point, yO lk , calculated over the whole
ˇ ˇ
sample. We denote with Ds;p the mean absolute error between the previous yO k
p

and the last point yO k . We also calculate the corresponding mean square errors. The
l

results are shown in Table 8.2 standardized by last point H13.It is evident that the
size of revisions for the most recent estimates is smaller for ALF relative to last
point asymmetric H13. This indicates a faster convergence to the symmetric filter,
which was also shown theoretically by means of spectral analysis. Similarly, the
210 8 Further Developments on the Henderson Trend-Cycle Filter

distance between the last and previous to the last filters is smaller for ALF relative
to H13.
In summary, the symmetric and asymmetric weights of the cascade filter are:
• Symmetric weights

0:027 0:007 0:031 0:067 0:136 0:188 0:224 0:188 0:136 0:067 0:031 0:007 0:027

• Asymmetric weights

0:026 0:007 0:030 0:065 0:132 0:183 0:219 0:183 0:132 0:065 0:030 0:006
0:000 0:026 0:007 0:030 0:064 0:131 0:182 0:218 0:183 0:132 0:065 0:031
0:000 0:000 0:025 0:004 0:034 0:069 0:137 0:187 0:222 0:185 0:131 0:064
0:000 0:000 0:000 0:020 0:005 0:046 0:083 0:149 0:196 0:226 0:184 0:130
0:000 0:000 0:000 0:000 0:001 0:033 0:075 0:108 0:167 0:205 0:229 0:182
0:000 0:000 0:000 0:000 0:000 0:045 0:076 0:114 0:134 0:182 0:218 0:230

8.3 The Henderson Filter in the Reproducing Hilbert Space

(RKHS)

The study of the properties and limitations of the Henderson filter has been done
in different contexts and attracted the attention of a large number of authors,
among them, Cholette [6], Kenny and Durbin [23], Castles [4], Dagum and Laniel
[13], Cleveland et al. [7], Dagum [9], Gray and Thomson [21], Loader [25],
Dalton and Keogh [17], Ladiray and Quenneville [24], Quenneville et al. [29],
Findley and Martin [19], and Dagum and Luati [15, 16]. However, none of these
studies have approached the Henderson smoother from a RKHS perspective. In
this regard, in 2008 Dagum and Bianconcini [11] introduced a Reproducing Kernel
Hilbert Space (RKHS) representation of the Henderson filter. The Henderson kernel
representation enabled the construction of a hierarchy of kernels with varying
smoothing properties. Furthermore, for each kernel order, the asymmetric filters
could be derived coherently with the corresponding symmetric weights or from
a lower or higher order kernel within the hierarchy, if more appropriate. In the
particular case of the currently applied asymmetric Henderson filters, those obtained
by means of the RKHS, coherent to the symmetric smoother, were shown to have
superior properties from the viewpoint of signal passing, noise suppression, and
revisions. A comparison was made with real life series.
8.3 The Henderson Filter in the Reproducing Hilbert Space (RKHS) 211

8.3.1 Linear Filters in Reproducing Kernel Hilbert Spaces

Let fyt ; t D 1; 2; : : : ; ng denote the input series. In this study, we work under the
following (basic) specification for the input time series.
Assumption 1 The input time series fyt ; t D 1; 2; : : : ; ng, that is seasonally
adjusted or without seasonality, can be decomposed into the sum of a systematic
component, called the signal (or nonstationary mean or trend-cycle) TCt , plus an
irregular component It , called the noise, such that

yt D TC t C It : (8.19)

The noise It is assumed to be either a white noise, WN.0; I2 /, or, more generally, to
follow a stationary and invertible AutoRegressive Moving Average (ARMA) process.
The trend-cycle can be deterministic or stochastic, and has a global or a local
representation. It can be represented locally by a polynomial of degree p of a
variable j which measures the distance between yt and the neighboring observations
ytCj .
Assumption 2 Given "t for some time point t, it is possible to find a local
polynomial trend estimator

TC tCj D a0 C a1 j C    C ap jp C "tCj ; j D m; : : : ; m; (8.20)

where a0 ; a1 ; : : : ; ap 2 R and "t is assumed to be purely random and mutually

uncorrelated with It .
The coefficients a0 ; a1 ; : : : ; ap can be estimated by ordinary or weighted least
squares or by summation formulae. The solution for aO 0 provides the trend-cycle
c t .0/, which equivalently consists in a weighted average applied in a
estimate TC
moving manner, such that

X
m
c t .0/ D TC
TC ct D wj ytCj ; (8.21)
jDm

where wj , j < n, denotes the weights to be applied to the observations ytCj to get the
estimate TCc t for each point in time t D 1; 2; : : : ; n.
The weights depend on: (1) the degree of the fitted polynomial, (2) the amplitude
of the neighborhood, and (3) the shape of the function used to average the
observations in each neighborhood.
Once a (symmetric) span 2m C 1 of the neighborhood has been selected, the
wj ’s for the observations corresponding to points falling out of the neighborhood
of any target point are null or approximately null, such that the estimates of the
n  2m central observations are obtained by applying 2m C 1 symmetric weights
to the observations neighboring the target point. The missing estimates for the first
212 8 Further Developments on the Henderson Trend-Cycle Filter

and last m observations can be obtained by applying asymmetric moving averages

of variable length to the first and last m observations, respectively. The length of
the moving average or time invariant symmetric linear filter is 2m C 1, whereas the
asymmetric linear filters length is time-varying.
Using the backshift operator B, such that Bd yt D ytd , Eq. (8.21) can be written
as

X
m
ct D
TC wj Bj yt D W.B/yt ; t D 1; 2; : : : ; n; (8.22)
jDm

where W.B/ is a linear nonparametric estimator.

Definition 8.1 Given p  2, W.B/ is a p-th order kernel if

X
m
wj D 1; (8.23)
jDm

and

X
m
ji wj D 0; (8.24)
jDm

for some i D 1; 2; : : : ; p. In other words, it will reproduce a polynomial trend of

degree (p  1) without distortion.
A different characterization of a p-th order nonparametric estimator can be provided
by means of the RKHS methodology.
A Hilbert space is a complete linear space with a norm given by an inner product.
The space of square integrable functions, denoted by L2 , and the finite p-dimensional
space Rp are those used in the study.
Assumption 3 The time series fyt ; t D 1; 2; : : : ; ng is Ra finite realization of a family
of square Lebesgue integrable random variables, i.e., T j Yt j2 dt < 1. Hence, the
random process fYt gt2T belongs to the space L2 .T/.
The space L2 .T/ is a Hilbert space endowed with the inner product defined by
Z
< U.t/; V.t/ >D E.U.t/V.t// D U.t/V.t/f0 .t/dt; (8.25)
T

where U.t/V.t/ 2 L2 .T/, and f0 is a probability density function, weighting each

observation to take into account its position in time. In the following, L2 .T/ will be
indicated as L2 . f0 /.
Under the Assumption 2, the local trend TCt ./ belongs to the Pp space of
polynomials of degree at most p, being p a nonnegative integer.
8.3 The Henderson Filter in the Reproducing Hilbert Space (RKHS) 213

Pp is a Hilbert subspace of L2 . f0 /, hence it inherits its inner product property

given by
Z
< P.t/; Q.t/ >D P.t/Q.t/f0 .t/dt; (8.26)
T

where P.t/; Q.t/ 2 Pp .

Suppose that fyt ; t D 1; 2; : : : ; ng 2 L2 . f0 / can be decomposed as in Eq. (8.19),
c t of Eq. (8.20) can be obtained by minimizing the distance between
the estimate TC
ytCj and TCt .j/, that is,
Z
min ky  TCk2 D min .y.t C s/  TCt .s//2 f0 .s/ds; (8.27)
TC2Pp TC2Pp T

where the weighting function f0 depends on the distance between the target point t
and each observation within a symmetric neighborhood of points around t.
Theorem 8.1 Under the Assumptions 1–3, the minimization problem (8.27) has a
unique and explicit solution.
Proof By the projection theorem (see, e.g., Priestley [28]), each element ytCj of the
Hilbert space L2 . f0 / can be decomposed into the sum of its projection in a Hilbert
subspace of L2 . f0 /, such as the space Pp , plus its orthogonal complement as follows

ytCj D ˘Pp ŒytCj C fytCj  ˘Pp ŒytCj g (8.28)

where ˘Pp ŒytCj denotes the projection of the observations ytCj ; j D m; : : : ; m; on
Pp . By orthogonality, for every j 2 T

c t .0/ D TC
< ytCj ; Rp .j; 0/ >D< ˘Pp ŒytCj ; Rp .j; 0/ >D ˘Pp Œyt D TC c t: (8.29)

c t .0/ is given by
Thus, TC
Z
c t .0/ D
TC y.t C s/Rp .s; 0/f0 .s/ds (8.30)
T
Z
D ˘Pp Œy.t C s/ Rp .s; 0/f0 .s/ds; (8.31)
T

where Rp is the reproducing kernel of the space Pp .

Hence, the estimate TCc t can be equivalently seen as the projection of yt on Pp
and as a local weighted average of the observations for the discrete version of the
filter given in Eq. (8.22), where the weights wj are derived by a kernel function K of
214 8 Further Developments on the Henderson Trend-Cycle Filter

order (p C 1),

KpC1 .t/ D Rp .t; 0/f0 .t/; (8.32)

where p is the degree of the fitted polynomial.

The following result, proved by Berlinet [1], is fundamental.
Corollary 8.1 Kernels of order .p C 1/, p  1, can be written as products of the
reproducing kernel Rp .t; :/ of the space Pp L2 . f0 / and a density function f0 with
finite moments up to order 2p. That is,

KpC1 .t/ D Rp .t; 0/f0 .t/:

Remark 8.1 (Christoffel–Darboux Formula) For any sequence .Pi /0ip of .p C 1/

orthonormal polynomials in L2 . f0 /;

X
p
Rp .t; 0/ D Pi .t/Pi .0/: (8.33)
iD0

Therefore, Eq. (8.32) becomes

X
p
KpC1 .t/ D Pi .t/Pi .0/f0 .t/: (8.34)
iD0

An important outcome of the RKHS theory is that linear filters can be grouped
into hierarchies fKp ; p D 2; 3; 4; : : :g with the following property: each hierarchy is
identified by a density f0 and contains kernels of order 2, 3, 4, . . . which are products
of orthonormal polynomials by f0 .
The weight system of a hierarchy is completely determined by specifying: (a) the
bandwidth or smoothing parameter, (b) the maximum order of the estimator in the
family, and (c) the density f0 .
There are several procedures to determine the bandwidth or smoothing parameter
(for a detailed discussion see, e.g., Berlinet and Devroye [2]). In this study, however,
the smoothing parameter is not derived by data dependent optimization criteria,
but it is fixed with the aim to obtain a kernel representation of the most often
applied Henderson smoothers. Kernels of any length, including infinite ones, can be
obtained with the above approach. Consequently, the results discussed can be easily
extended to any filter length as long as the density function and its orthonormal
polynomials are specified. The identification and specification of the density is one
of the most crucial task for smoothers based on local polynomial fitting by weighted
least squares, as LOESS and the Henderson smoothers. The density is related to the
weighting penalizing function of the minimization problem.
It should be noticed that the RKHS approach can be applied to any linear filter
characterized by varying degrees of fidelity and smoothness as described by Gray
8.3 The Henderson Filter in the Reproducing Hilbert Space (RKHS) 215

and Thomson [21]. In particular, applications to local polynomial smoothers are

treated in Dagum and Bianconcini [10], and to smoothing splines in Wahba [31].

8.3.2 The Symmetric Henderson Smoother and Its Kernel

Representation

Recognition of the fact that the smoothness of the estimated trend-cycle curve
depends directly on the smoothness of the weight diagram led Henderson to develop
a formula which makes the sum of squares of the third differences of the smoothed
series a minimum for any number of terms.
Henderson’s starting point was the requirement that the filter should reproduce a
cubic polynomial trend without distortion. Henderson proved that three alternative
smoothing criteria give the same formula, as shown explicitly by Kenny and Durbin
[23] and Gray and Thomson [21]: (1) minimization of the variance of the third
differences of the series defined by the application of the moving average; (2)
minimization of the sum of squares of the third differences of the coefficients of
the moving average formula; and (3) fitting a cubic polynomial by weighted least
squares, where the weights are chosen as to minimize the sum of squares of their
third differences.
The problem is one of fitting a cubic trend by weighted least squares to the
observations ytCj ; j D m; : : : ; m, the value of the fitted function at j D 0 being
taken as the smoothed observation TC c t . Representing the weight assigned to the
residuals from the local polynomial regression by Wj ; j D m; : : : ; m , where
Wj D Wj , the problem is the minimization of

X
m
Wj ŒytCj  a0  a1 j  a2 j2  a3 j3 2 ; (8.35)
jDm

c t.
where the solution for the constant term aO 0 is the smoothed observation TC
Henderson showed that TC c t is given by

X
m X
m
ct D
TC .j/Wj ytCj D wj ytCj ; (8.36)
jDm jDm

where .j/ is a cubic polynomial whose coefficients have the property that the filter
reproduces the data if they follow a cubic. Henderson also proved the converse: if
the coefficients of a cubic-reproducing summation formula fwj ; j D m; : : : ; mg do
not change their sign more than three times within the filter span, then the formula
can be represented as a local cubic smoother with weights Wj > 0 and a cubic
polynomial .j/, such that .j/Wj D wj .
216 8 Further Developments on the Henderson Trend-Cycle Filter

P
Henderson measured the amount of smoothing of the input series by ntD3 ..1 
B/3 yt /2 or equivalently
P by the sum of squares of the third differences of the
3 2
weight diagram, m jDmC3 ..1  B/ wj / . The solution is that resulting from the
minimization of a cubic polynomial function by weighted least squares with

Wj / f.m C 1/2  j2 gf.m C 2/2  j2 gf.m C 3/2  j2 g (8.37)

as the weighting penalty function of criterion (3) above. The weight diagram
fwj ; j D m; : : : ; mg corresponding to Eq. (8.37), known as Henderson’s ideal
formula, is obtained, for a filter length equal to 2m0  3, by
315Œ.m0  1/2  j2 .m02  j2 /Œ.m0 C 1/2  j2 .3m02  16  11j2 /
wj D : (8.38)
8m0 .m0 2  1/.4m02  1/.4m0 2  9/.4m02  25/

This optimality result has been rediscovered several times in the modern literature,
usually for asymptotic variants. Loader [25] showed that the Henderson’s ideal
formula (8.38) is a finite sample variant of a kernel with second order vanishing
moments which minimizes the third derivative of the function given by Muller [26].
In particular, Loader showed that for large m, the weights of Henderson’s ideal
penalty function Wj are approximately m6 W.j=m/, where W.j=m/ is the triweight
density function. He concluded that, for very large m, the weight diagram is
approximately .315=512/  W.j=m/.3  11.j=m/2 / equivalent to the kernel given
by Muller [26].
To derive the Henderson kernel hierarchy by means of the RKHS methodology,
the density corresponding to Wj and its orthonormal polynomials have to be deter-
mined. The triweight density function gives very poor results when the Henderson
smoother spans are of short or medium lengths, as in most application cases, ranging
from 5 to 23 terms. Hence, the exact density function corresponding to Wj was
derived.
Theorem 8.2 The exact probability density corresponding to the Henderson’s ideal
weighting penalty function (8.37) is given by

.m C 1/
f0H .t/ D W..m C 1/t/; t 2 Œ1; 1 (8.39)
k
R mC1
where k D m1 W.j/dj, and j D .m C 1/t.
Proof The weighting function W.j/ in Eq. (8.37), is nonnegative in the intervals
.m  1; m C 1/, .m  3; m  2/, .m C 2; m C 3/ and negative otherwise. W.j/ is
also equal to zero if j D ˙.m C 1/; ˙.m C 2/; ˙.m C 3/, and on Œm  1; m C 1 ,
W.j/ is increasing in Œm  1; 0/, decreasing in .0; m C 1 , and reaches its maximum
at j D 0.
Therefore, the support chosen is Œm  1; m C 1 to satisfy the positive definite
R mC1
condition. The integral k D m1 W.j/dj is different from 1 and represents the
integration constant on this support. It follows that the density corresponding to
8.3 The Henderson Filter in the Reproducing Hilbert Space (RKHS) 217

W.j/ on the interval Œm  1; m C 1 is given by

f0 .j/ D W.j/=k:

To eliminate the dependence of the support on the bandwidth parameter m, a new

variable ranging on Œ1; 1 , t D j=.m C 1/, is considered. Applying the change of
variables method,
ˇ ˇ
ˇ 1 ˇ
f0H .t/ D f0 .t1 .j// ˇ @t @t .j/ ˇ ;

where t.j/ D j
mC1
, and t1 .j/ D .m C 1/t. The result follows by substitution.
The density f0H .t/ is symmetric, i.e., f0H .t/ D f0H .t/, nonnegative on Œ1; 1 , and
is equal to zero when t D 1 or t D 1. Furthermore, f0H .t/ is increasing on Œ1; 0/,
decreasing on .0; 1 , and reaches its maximum at t D 0.
For m D 6, the filter is the classical 13-term Henderson and the corresponding
probability function results

15
f0H .t/ D .5184  12289t2 C 9506t4  2401t6/; t 2 Œ1; 1 : (8.40)
79376
To obtain higher order kernels the corresponding orthonormal polynomials have to
be computed for the density (8.39). The polynomials can be derived by the Gram–
Schmidt orthonormalization procedure or by solving the Hankel system based on
the moments of the density f0H . This latter is the approach followed in this study.
The hierarchy corresponding to the 13-term Henderson kernel is shown in Table 8.3,
where for p D 3 it provides a representation of the classical Henderson filter.
Since the triweight density function gives a poor approximation of the Henderson
weights for small m (5–23 terms), another density function with well-known
theoretical properties was searched. The main reason is that the exact density (8.39)
is function of the bandwidth and need to be calculated any time that m changes
together with its corresponding orthonormal polynomials. The biweight density was
found to give almost equivalent results to those obtained with the exact density
function without the need to be calculated any time that the Henderson smoother
length changes.

Table 8.3 13-Term Henderson kernel hierarchy

Henderson kernels Kernel orders
15
79376
.5184  12289t2 C 9506t4  2401t6 / pD2
15
79376
.5184  12289t2 C 9506t4  2401t6 /  . 2175
1274
 1372 2
265
t / pD3
218 8 Further Developments on the Henderson Trend-Cycle Filter

Another important advantage is that the biweight density function belongs to the
well-known Beta distribution family, that is,
 
r
f .t/ D 2B.sC1; 1r / .1  jtjr /s IŒ1;1 .t/; (8.41)

R1
where B.a; b/ D 0 ta1 .1  t/b1 dt with a; b > 0 is the Beta function. The
orthonormal polynomials needed for the reproducing kernel associated with the
biweight function are the Jacobi polynomials, for which explicit expressions for
computation are available and their properties have been widely studied in the
literature.
Therefore, another Henderson kernel hierarchy is obtained using the biweight
density

15
f0B .t/ D .1  t2 /2 IŒ1;1 .t/; (8.42)
16
combined with the Jacobi orthonormal polynomials. These latter are characterized
by the following explicit expression:
n   
1 X nC˛ nCˇ
P˛;ˇ
n .t/ D n .t  1/nm .t C 1/m ; (8.43)
2 mD0 m nm

where ˛ D 2 and ˇ D 2. The Henderson second order kernel is given by the density
function f0B , since the reproducing kernel R1 .j; 0/ of the space of polynomials of
degree at most one is always equal to one.
On the other hand, the third order kernel is given by

15

.1  jtj2 /2  74  21 2
4
t : (8.44)
16
Table 8.4 shows the classical and the two kernels (exact and biweight) Henderson
symmetric weights for spans of 9, 13, and 23 terms, where the central weight values
are given in bold.
The small discrepancy of the two kernel functions relative to the classical
Henderson smoother is due to the fact that the exact density is obtained by
interpolation from a finite small number of points of the weighting penalty function
Wj .
On the other hand, the biweight is already a density function which is made
discrete
Pm by choosing selected points to produce the weights. The smoothing measure
3 2
jDmC3 ..1  B/ wj / for each filter span is calculated as shown in Table 8.5.
The smoothing power of the filters is very close except for the exact 9-term
Henderson kernel which gives the smoothest curve. Given the equivalence for
symmetric weights, the RKHS methodology is used to generate the correspondent
asymmetric filters for the m first and last points.
Table 8.4 Weight systems for 9-, 13-, and 23-term Henderson smoothers
Length Henderson filter
9 Classical 0:041 0:010 0:118 0:267 0:331
Exact Kernel 0:038 0:005 0:119 0:262 0:325
Biweight Kernel 0:039 0:011 0:120 0:266 0:328
13 Classical 0:019 0:028 0:000 0:065 0:147 0:214 0:240
Exact Kernel 0:019 0:027 0:001 0:066 0:147 0:213 0:238
Biweight Kernel 0:020 0:030 0:002 0:070 0:149 0:211 0:234
23 Classical 0:004 0:011 0:016 0:015 0:005 0:013 0:039 0:068 0:097 0:122 0:138 0:144
Exact Kernel 0:004 0:011 0:016 0:014 0:005 0:014 0:039 0:068 0:097 0:122 0:138 0:144
Biweight Kernel 0:005 0:014 0:018 0:014 0:001 0:019 0:045 0:072 0:098 0:118 0:132 0:137
8.3 The Henderson Filter in the Reproducing Hilbert Space (RKHS)
219
220 8 Further Developments on the Henderson Trend-Cycle Filter

Table 8.5 Sum of squares of the third differences of the weights for the classical and kernel
Henderson smoother
Filters 9-term 13-term 23-term
Classical Henderson smoother 0:053 0:01 0:00
Exact Henderson kernel 0:048 0:01 0:00
Biweight Henderson kernel 0:052 0:01 0:00

8.3.3 Asymmetric Henderson Smoothers and Their Kernel

Representations

The asymmetric Henderson smoothers currently in use were developed by Musgrave

[27]. The asymmetric weights of the Henderson kernels are derived by adapting the
third order kernel functions to the length of the last m asymmetric filters such that,

K.j=b/
wq;j D Pq ; j D m; : : : ; qI q D 0; : : : ; m  1; (8.45)
jDm K.j=b/

where K./ denotes the third order Henderson kernel, j the distance to the target point
t, b the bandwidth parameter equal to (m C 1), and (m C q C 1) is the asymmetric
filter length. For example, the asymmetric weights of the 13-term Henderson kernel
for the last point (q D 0) are given by

K.j=7/
w0;j D P0 ; j D 6; : : : ; 0: (8.46)
jD6 K.i=7/

Figure 8.6 shows the gain functions of the symmetric Henderson filter together
with the classical and kernel representation for the last point superposed on the
X11/X12ARIMA Seasonal Adjustment (SA) symmetric filter. This latter results
from the convolution of: (1) 12-term centered moving average, (2) 3  3 m.a., (3)
35 m.a., and (4) 13-term Henderson filter. Its property has been widely discussed in
Dagum et al. [12]. The gain function of the SA filter should be interpreted as relating
the spectrum of the original series to the spectrum of the estimated seasonally
adjusted series.
It is apparent that the asymmetric Henderson kernel filter does not amplify
the signal as the classical asymmetric one and converges faster to the final one.
Furthermore, the asymmetric kernel suppresses more noise relative to the Musgrave
filter. Since the weights of the biweight and exact Henderson kernels are equal up to
the third digit, no visible differences are seen in the corresponding gain and phase
shift functions. Hence, we only show those corresponding to the exact Henderson
kernel in Figs. 8.6 and 8.7. There is an increase of the phase shift for the low
frequencies relative to that of the classical H13 but both are less than a month, as
exhibited in Fig. 8.7.
8.3 The Henderson Filter in the Reproducing Hilbert Space (RKHS) 221

Fig. 8.6 Gain functions of the (last point) asymmetric 13-term filter (continuous line) and kernel
(gray continuous line) with the corresponding symmetric (dotted line) and SA cascade filter (gray
dotted line)

Fig. 8.7 Phase shifts of the asymmetric (end point) weights of the Henderson kernel (dotted line)
and of the classical H13 filter (continuous line)
222 8 Further Developments on the Henderson Trend-Cycle Filter

References

1. Berlinet, A. (1993). Hierarchies of higher order kernels. Probability Theory and Related Fields,
94, 489–504.
2. Berlinet, A., & Devroye, L. (1994). A comparison of kernel density estimates. Publications de
l’Institut de Statistique de l’Université de Paris, 38, 3–59.
3. Box, G. E. P., & Jenkins, G. M. (1970). Time series analysis: Forecasting and control. San
Francisco, CA, USA: Holden-Day (Second Edition 1976).
4. Castles, I. (1987). A guide to smoothing time series estimates of trend. Catalogue No. 1316.
Australian Bureau.
5. Chhab, N., Morry, M., & Dagum, E. B. (1999). Results on alternative trend-cycle estimators
for current economic analysis. Estadistica, 49–51, 231–257.
6. Cholette, P. A. (1981). A comparison of various trend-cycle estimators. In O. D. Anderson &
M. R. Perryman (Eds.), Time series analysis (pp. 77–87). Amsterdam: North Holland.
7. Cleveland, R. B., Cleveland, W. S., McRae, J. E., & Terpenning, I. (1990). STL: A seasonal
trend decomposition procedure based on LOESS. Journal of Official Statistics, 6(1), 3–33.
8. Dagum, E. B. (1988). The X11ARIMA/88 seasonal adjustment method-foundations and user’s
manual. Ottawa, Canada: Time Series Research and Analysis Centre, Statistics Canada.
9. Dagum, E. B. (1996). A new method to reduce unwanted ripples and revisions in trend-cycle
estimates from X11ARIMA. Survey Methodology, 22, 77–83.
10. Dagum, E. B., & Bianconcini, S. (2006). Local polynomial trend-cycle predictors in reproduc-
ing kernel Hilbert spaces for current economic analysis. Anales de Economia Aplicada, 1–22.
11. Dagum, E. B., & Bianconcini, S. (2008). The Henderson smoother in reproducing kernel
Hilbert space. Journal of Business and Economic Statistics, 26(4), 536–545.
12. Dagum, E. B., Chhab, N., & Chiu, K. (1996). Derivation and analysis of the X11ARIMA and
Census X11 linear filters. Journal of Official Statistics, 12(4), 329–347.
13. Dagum, E. B., & Laniel, N. (1987). Revisions of trend-cycle estimators of moving average
seasonal adjustment methods. Journal of Business and Economic Statistics, 5, 177–189.
14. Dagum, E. B., & Luati, A. (2000). Predictive performance of some nonparametric linear and
nonlinear smoothers for noisy data. Statistica, LX(4), 635–654.
15. Dagum, E. B., & Luati, A. (2004). A linear transformation and its properties with special
applications in time series filtering. Linear Algebra and its Applications, 338, 107–117.
16. Dagum, E. B., & Luati, A. (2009). A cascade linear filter to reduce revisions and turning points
for real time trend-cycle estimation. Econometric Reviews, 28(1–3), 40–59.
17. Dalton, P., & Keogh, G. (2000). An experimental indicator to forecast turning points in the Irish
business cycle. Journal of the Statistical and Social Inquiry Society of Ireland, XXIX, 117–176.
18. Darnè, O. (2002). La dèsaisonnalisation des chroniques économiques: analyse des conditions
conjuncturelles (Ph.D. dissertation, Department of Economics, University of Montpellier 1,
France).
19. Findley, D. F., & Martin, D. E. K. (2006). Frequency domain analysis of SEATS and X-11/X-
12-ARIMA seasonal adjustment filters for short and moderate length time series. Journal of
Official Statistics, 22, 1–34.
20. Findley, D. F., Monsell, B. C., Bell, W. R., Otto, M. C., & Chen, B. C. (1998). New capabilities
and methods of the X12ARIMA seasonal adjustment program. Journal of Business and
Economic Statistics, 16(2), 127–152.
21. Gray, A., & Thomson, P. (1996). Design of moving-average trend filters using fidelity and
smoothness criteria. In P. M. Robinson & M. Rosenblatt (Eds.), Time series analysis (in
memory of E.J. Hannan). Vol. II. Lecture notes in statistics (Vol. 115, pp. 205–219). New York:
Springer.
22. Henderson, R. (1916). Note on graduation by adjusted average. Transaction of Actuarial
Society of America, 17, 43–48.
23. Kenny, P., & Durbin, J. (1982). Local trend estimation and seasonal adjustment of economic
and social time series. Journal of the Royal Statistical Society A, 145, 1–41.
References 223

24. Ladiray, D., & Quenneville, B. (2001). Seasonal adjustment with the X-11 method. Lecture
notes in statistics (Vol. 158). New York: Springer.
25. Loader, C. (1999). Local regression and likelihood. New York: Springer.
26. Müller, H. G. (1984). Smooth optimum kernel estimators of regression curves, densities and
modes. Annals of Statistics, 12, 766–774.
27. Musgrave, J. (1964). A set of end weights to end all end weights. Working Paper. U.S. Bureau
of Census, Washington DC.
28. Priestley, M. B. (1981). Spectral analysis and time series. Probability and mathematical
statistics. New York: Academic Press.
29. Quenneville, B., Ladiray, D., & Lefrancois, B. (2003). A note on Musgrave asymmetrical trend-
cycle filters. International Journal of Forecasting, 19(4), 727–734.
30. Shiskin, J., Young, A. H., & Musgrave, J. C. (1967). The X-11 variant of the Census Method
II seasonal adjustment program. Technical Paper 15 (revised). US Department of Commerce,
Bureau of the Census, Washington, DC.
31. Wahba, G. (1990). Spline models for observational data. Philadelphia: SIAM.
32. Zellner, A., Hong, C., & Min, C. (1991). Forecasting turning points in international output
growth rates using Bayesian exponentially weighted autoregression, time-varying parameter
and pooling techniques. Journal of Econometrics, 48, 275–304.
Chapter 9
A Unified View of Trend-Cycle Predictors
in Reproducing Kernel Hilbert Spaces (RKHS)

Abstract A common approach for trend-cycle estimation is to use nonparametric

smoothing estimators which can be finite or infinite in length. The main methods
presented in this chapter are based on different criteria of fitting and smoothing,
namely: (1) density functions, (2) local polynomial fitting, (3) graduation theory, and
(4) smoothing spline regression. A unified approach for all of these different trend-
cycle nonparametric estimators is provided by means of the Reproducing Kernel
Hilbert Space (RKHS) methodology. It is shown how nonparametric estimators can
be transformed into kernel functions of order two, that are probability densities,
and from which corresponding higher order kernels are derived. This kernel
representation enables the comparison of estimators based on different smoothing
criteria, and has important consequences in the derivation of the asymmetric filters
which are applied to the most recent seasonally adjusted data for real time trend-
cycle estimation.

In many scientific fields the observations are measured with error and in such
cases the nonstationary mean of the time series is of great interest. Based on the
assumption that a time series can be decomposed as a sum of a signal plus an erratic
component, signal extraction deals with the problem of finding the “best” estimate
of the nonstationary mean given the observations corrupted by noise.
If the time series is without seasonality or seasonally adjusted, the signal
represents the trend and cyclical components, usually referred to as trend-cycle for
they are estimated jointly in medium and short length series. The determination of
a suitable inferential methodology will hinge on the assumptions made about the
signal.
A common approach is to use nonparametric smoothing estimators which can
be finite or infinite in length. They are based on different criteria of fitting and
smoothing and are: (1) density functions, (2) local polynomial fitting, (3) graduation
theory, and (4) smoothing spline regression.
A unified approach for all of these different nonparametric estimators is provided
by means of the Reproducing Kernel Hilbert Space (RKHS) methodology.

© Springer International Publishing Switzerland 2016 225

E. Bee Dagum, S. Bianconcini, Seasonal Adjustment Methods and Real Time
Trend-Cycle Estimation, Statistics for Social and Behavioral Sciences,
DOI 10.1007/978-3-319-31822-6_9
226 9 A Unified View of Trend-Cycle Predictors in Reproducing Kernel Hilbert. . .

The main theory and systematic development of reproducing kernels and

associated Hilbert spaces were laid out by Aronszajn [1], who showed that the
properties of RKHS are intimately bounded up with properties of nonnegative
definite functions. Parzen [26] was the first to apply these concepts to time series
problems by means of a strictly parametric approach. From a nonparametric
perspective, de Boor and Lynch [12] used this methodology in the context of cubic
spline approximation. Later, Kimeldorf and Wahba [21, 22] proved that minimum
norm interpolations and smoothing problems with quadratic constraints imply
an equivalent Gaussian stochastic process. Recently, reproducing kernel methods
have been prominent as a framework for penalized spline and quantile regression
(see, e.g., Wahba [33]), and in the support vector machine literature, as described in
[7, 14, 34], and [27].
In this chapter, we show how nonparametric estimators can be transformed
into kernel functions of order two, that are probability densities, and from which
corresponding higher order kernels are derived. The density function provides
the initial weighting form from which the higher order kernels inherit their
properties. This kernel representation enables the comparison of estimators based
on different smoothing criteria, and has important consequences in the derivation of
the asymmetric filters which are applied to the most recent seasonally adjusted data
for real time trend-cycle estimation. In particular, those obtained by means of RKHS
are shown to have superior properties from the view point of signal passing, noise
suppression, and revisions relative to the currently applied that we call the classical
ones.
The main methods presented here have been discussed in details by Dagum and
Bianconcini [11], and we refer the readers to such study for further information.

9.1 Nonparametric Estimators in RKHS

Let fyt ; t D 1; 2; : : : ; ng denote an n-dimensional vector of observed seasonally

adjusted (or without seasonality) time series, specified as follows:
Assumption 1 The time series fyt ; t D 1; 2; : : : ; ng, that is seasonally adjusted
or without seasonality, can be decomposed as the sum of a systematic component,
called the signal TCt (or nonstationary mean or trend-cycle), plus an erratic
component It , called the noise, such that

yt D TCt C It ; (9.1)

where It is purely random or follows an ARMA process of low order.

Assumption 2 The time series fyt ; t D 1; 2; : : : ; ng is a finite realization of
a stochastic process, whose trajectories fY.t/; t 2 Tg are square Lebesgue
integrable functions with respect to a density
R function f0 defined on T (e.g., T D
Œ1; 1 ; Œ0; C1/, or .1; 1/) that is, T jY.t/j2 f0 .t/dt < C1.
9.1 Nonparametric Estimators in RKHS 227

In other words, fY.t/; t 2 Tg belong to the Hilbert space L2 . f0 / endowed with the
inner product defined by
Z
< Y1 ; Y2 >L2 . f0 / D Y1 .t/Y2 .t/f0 .t/dt 8Y1 ; Y2 2 L2 . f0 /: (9.2)
T

f0 is a density function of time, that is a nonnegative definite function integrating to

unity over the time domain T. It weights each value to take into account its temporal
position. In particular, f0 can be seen as a probabilistic density function since it
allows to associate to each value in time a measure of its reliability as a function of
its distance to the time point of interest.
Assumption 3 The signal TC is globally, i.e., on the whole time span, a smooth
function of time.
Under Assumption 3, in virtue of the Taylor’s expansion, TC can be locally
approximated by a polynomial function, say of degree p, of the time distance j
between yt and its neighboring values ytCj , such that

TCtCj D a0 C a1 j C    C ap jp C "tCj ; j D m; : : : ; m; (9.3)

where a0 ; a1 ; : : : ; ap 2 R, and "t is a random component assumed to be mutually

uncorrelated with It . The size of the neighborhood of values around t taken into
account in the estimation of the polynomial coefficients a0 ; a1 ; : : : ; ap is equal to
2m C 1.
Given Eq. (9.3), the analysis of the signal can be performed in the space of
polynomials of degree at most p, Pp L2 . f0 /, being p a nonnegative integer.
The coefficients a0 ; a1 ; : : : ; ap of the local polynomial trend-cycle are estimated
by projecting the values in a neighborhood of yt on the subspace Pp , or equivalently
by minimizing the weighted least square fitting criterion
Z
min ky  TCkPp D .y.t  s/  TC.t  s//2 f0 .s/ds;
2
(9.4)
TC2Pp T

where kk2Pp denotes the Pp -norm. The weighting function f0 depends on the distance
between the target point t and each observation within a symmetric neighborhood
of points around t.
c t , for which a general charac-
The solution for aO 0 provides the estimate TC
terization and explicit representation can be provided by means of the RKHS
methodology.
Definition 9.1 Given a Hilbert space H , the reproducing kernel R is a function
R W T  T ! R, satisfying the following properties:
1. R.t; :/ 2 H , 8t 2 T;
2. < g.:/; R.t; :/ >D g.t/, 8t 2 T and g 2 H .
228 9 A Unified View of Trend-Cycle Predictors in Reproducing Kernel Hilbert. . .

Condition 2 is called the reproducing property: the value of the function TC at the
point t is reproduced by the inner product of TC with R.t; :/. R.t; :/ is called the
reproducing kernel since

< R.t; :/; R.:; s/ >D R.t; s/: (9.5)

Lemma 9.1 The space Pp is a reproducing kernel Hilbert space of polynomials on

some domain T, that is, there exists an element Rp .t; :/ 2 Pp , such that

P.t/ D< P.:/; Rp .t; :/ > 8t 2 T and 8P 2 Pp

The proof easily follows by the fact that any finite dimensional Hilbert space has a
reproducing kernel [3].
Theorem 9.1 Under the Assumptions 1–3, the minimization problem (9.4) has a
unique and explicit solution given by
Z
c
TC.t/ D y.t  s/KpC1 .s/ds; (9.6)
T

where KpC1 is a kernel function of order p C 1. Note that for p  2; Kp is said to be

of order p if
Z Z
Kp .s/ds D 1; and si Kp .s/ds D 0; (9.7)
T T

for i D 1; 2; : : : ; p1. In other words, it will reproduce a polynomial trend of degree

p  1 without distortion.
Proof By the projection theorem [29], each element of the Hilbert space L2 . f0 / can
be decomposed into the sum of its projection in a Hilbert subspace of L2 . f0 /, such
as the space Pp , plus its orthogonal complement as follows:

y.t  s/ D ˘Pp Œ y.t  s/ C fy.t  s/  ˘Pp Œ y.t  s/ g; (9.8)

where ˘Pp Œ y.t  s/ denotes the projection of the data y.t  s/ on Pp . By

orthogonality,

c D ˘Pp Œ y.t/ D< ˘Pp Œ y.t  s/ ; Rp .s; 0/ >D< y.t  s/; Rp .s; 0/ > :
TC.t/ (9.9)

c is given by
Thus, TC.t/
Z
c D
TC.t/ ˘Pp Œ y.t  s/ Rp .s; 0/f0 .s/ds (9.10)
T
Z
D y.t  s/Rp .s; 0/f0 .s/ds; (9.11)
T

where Rp is the reproducing kernel of the space Pp .

9.1 Nonparametric Estimators in RKHS 229

c can be seen as a local weighted average of the seasonally

Hence, the estimate TC.t/
adjusted values, where the weights are derived by a kernel function K of order p C1,
where p is the degree of the fitted polynomial.
Next we make use of the two following fundamental results.
First, as shown by Berlinet [2], kernels of order .p C 1/, p  1 , can be written as
products of the reproducing kernel Rp .t; :/ of the space Pp L2 . f0 / and a density
function f0 with finite moments up to order 2p. That is,

KpC1 .t/ D Rp .t; 0/f0 .t/: (9.12)

Second, using the Christoffel–Darboux formula, for any sequence .Pi /0ip of
.p C 1/ orthonormal polynomials in L2 . f0 /, we have

X
p
Rp .t; 0/ D Pi .t/Pi .0/: (9.13)
iD0

Therefore, Eq. (9.12) becomes

X
p
KpC1 .t/ D Pi .t/Pi .0/f0 .t/: (9.14)
iD0

Applied to seasonally adjusted data or data without seasonality, the kernel acts as a
locally weighted average or linear filter that for each target point t gives the estimate

X
n
ct D
TC wt;i yi ; t D 1; 2; : : : ; n; (9.15)
iD1

where wt;i denotes the weights to be applied to yi to get the estimate TC c t for each
point in time t. Once a symmetric span 2mC1 of the neighborhood has been selected,
the weights for the data corresponding to points falling out the neighborhood of any
target point are null or approximately null. Hence, the estimates of the n2m central
values are obtained by applying 2mC1 symmetric weights to the values neighboring
the target point. That is,

X
m
ct D
TC wj ytCj t D m C 1; : : : ; n  m:
jDm

The weights wj ; j D m; : : : ; m, depend on the shape of the nonparametric estimator

KpC1 and on the value of a bandwidth parameter b fixed to ensure a neighborhood
230 9 A Unified View of Trend-Cycle Predictors in Reproducing Kernel Hilbert. . .

amplitude equal to 2m C 1, such that

 
j
KpC1 b
wj D P  ; j D m; : : : ; m: (9.16)
m j
jDm KpC1 b

On the other hand, the missing estimates for the first and last m data points can be
obtained by applying asymmetric linear filters of variable lengths.
Several nonparametric estimators have been developed in the literature for
time series smoothing. One approach based on least squares includes: (1) kernel
estimators, (2) local polynomial fitting, and (3) graduation theory. A second
approach corresponds to smoothing spline regression. Smoothing splines introduce
a roughness penalty term in the minimization problem (9.4), searching for an
optimal solution between both fitting and smoothing of the data. This requires an
adapted RKHS. Here it is shown how an equivalent kernel representation for the
smoothing spline can be derived in the polynomial space Pp .
A unified perspective is provided for all of these different nonparametric esti-
mators according to which they are transformed into kernel functions. This enables
comparisons among smoothers of different order within the same hierarchy as well
as kernels of the same order, but belonging to different hierarchies. Therefore, for
every estimator the density function f0 and the corresponding reproducing kernel are
identified.

9.1.1 Polynomial Kernel Regression

Local kernel regression deals with the problem of fitting a polynomial trend to
ytCj ; j D m; : : : ; m, where the value of the fitted function at j D 0 defines the
smoothed observation TC c t . Representing the weights assigned to the residuals from
the local polynomial regression by a symmetric and nonnegative function K0 , the
problem is the minimization of

X  
ti  2
n
K0 yt  a0  a1 .t  i/      ap .t  i/p ; (9.17)
iD1
b

where the parameter b determines the bandwidth of the weighting function.

Kernel estimators, local polynomial regression smoothers, and filters derived by
the graduation theory differ in the degree of the fitted polynomial, in the shape of the
weighting function, and the neighborhood of the values taken into account. We shall
discuss here the properties of the most often applied trend-cycle smoothers, namely
9.1 Nonparametric Estimators in RKHS 231

Table 9.1 Second and third order kernel within the Gaussian, tricube (LOESS) and biweight
(Henderson) hierarchies
Estimator Density   Third order kernel    2
2 2
Gaussian f0G .t/ D p1 exp  t2 t 2 .1; 1/ K3G .t/ D p12 exp  t2  3t
2 2
kernel
70

3 70

3 539

LOESS f0T .t/ D 1  jtj3 IŒ1;1 .t/; K3T .t/ D 1  jtj3  3719 t2
81
15
81
15

293 638

Hender- f0B .t/ D 16

.1  t2 /2 IŒ1;1 .t/: K3B .t/ D 16
.1  jtj2 /2  74 21 2
 4t
son filter

the Gaussian kernel, the LOESS estimator [5], and the Henderson filter [19]. The
density function and third order kernel for each estimator are shown in Table 9.1.
We refer the reader to [4, 9, 10], for a detailed discussion on how these density
functions and reproducing kernels have been derived.

9.1.2 Smoothing Spline Regression

Extending previous work by Whittaker [35] and Whittaker and Robinson [36],
Schoenberg [30] showed that a natural smoothing spline estimator of order ` arises
by minimizing the loss function
Z Z
.`/
2
min ky  TCk2W ` D .y.t/  TC.t//2 dt C  TC .t/ dt; (9.18)
TC2W2` .T/ 2 T T

where TC.`/ denotes the `-th derivative of the function TC, and W2` is the Sobolev
space of functions TC satisfying the following properties: (1) TC.i/ ; i D 1; : : : ; `1,
are absolutely continuous, (2) TC.`/ 2 L2 .T/.
The parameter  regulates the balance between the goodness of fit and the
smoothness of the curve. As  ! 0 the solution approaches an interpolating spline,
whereas as  ! 1, the solution tends to the least squares line.
Kimeldorf and Wahba [21, 22] showed that the minimizer of Eq. (9.18) is a
polynomial spline of degree p D 2`1 with knots at the data points. It is determined
by the unique Green’s function G .t; s/, such that
Z
c D
TC.t/ G .t; s/y.s/ds: (9.19)
T

The derivation of G .t; s/ corresponding to a smoothing spline of order ` requires

the solution of a 2`  2` system of linear equations for each value of .
A simplification is provided by studying G .t; s/ as the reproducing kernel
R`; .t; s/ of the Sobolev space W2` .T/, where T is an open subset of R. When T D R,
the space W2` .T/ falls into the family of Beppo-Levi spaces described in Thomas-
Agnan [32]. The corresponding reproducing kernel is translation invariant, and can
232 9 A Unified View of Trend-Cycle Predictors in Reproducing Kernel Hilbert. . .

be expressed in terms of R`;1 as follows:

1 t  s
R`; .t; s/ D R`; .t  s/ D R`;1 : (9.20)
 
A general formula for R`;1 , denoted as R` , is given by [32],

X̀ exp jtjei 2` Ck `  2

R` .t/ D ; ` D 1; 2; 3 : : : (9.21)
2`e.2`1/.i 2` Ci ` /
 k
kD0

Equation (9.21) describes an equivalent kernel hierarchy for smoothing splines of

order `, ` D 1; 2; : : :. The standard Laplace density R1 is the second order kernel
when ` D 1, and hence p D 1. Higher order kernels are derived by multiplying
R1 by a combination of trigonometric polynomials which take into account for the
roughness penalty term in (9.18). The third order smoother R2 , obtained by selecting
` D 2, hence p D 3, is familiar to the nonparametric statisticians since it is the
asymptotically equivalent kernel to the cubic smoothing spline derived by Silverman
[31]. It should be noticed, however, that when the neighborhood of points for the
estimation is small, as in most socioeconomic cases, Eq. (9.21) poorly represents
the linear approximation of the classical cubic smoothing spline, described in terms
of the influential matrix A./ [18, 33]. A./ relates the estimated values TC c0 D
c c c 0
.TC1 ; TC2 ; : : : ; TCn / to the values y D .y1 ; y2 ; : : : ; yn / as follows:

c D A./y:
TC (9.22)

c t is a weighted linear combination of all the values yt , with weights given

Each TC
by the elements of the t-th row of A./. In this study, we assume  as given, and
thus approximate each cubic spline predictor with time invariant linear filters. This
enables us to analyze the properties of the estimators looking at the corresponding
transfer functions. To obtain a reproducing kernel representation of smoothing
splines coherent with that derived for local kernel regression estimators, we have
to find a density function f0 according to which higher order kernels are obtained
via multiplication of f0 with corresponding orthonormal
R polynomials. This density
has to take into account for the smoothing term  T .g.`/ .u//2 du in Eq. (9.18).
Starting from the results of Eq. (9.21), we first consider the standard Laplace
density multiplied by the corresponding orthonormal polynomials. To evaluate the
goodness of this approximation we calculate the Euclidean distance  between
the linear approximation of the classical cubic smoothing spline (CSS) and the
corresponding third order kernel within each hierarchy, denoted by K, both in terms
of weights and gain functions. That is,
v v
uX u 1=2
u m uX
weights D t j wCSSj  wKj j 2 gain D t j GCSS .!/  GK .!/ j2 ;
jDm !D0
9.1 Nonparametric Estimators in RKHS 233

Table 9.2 2-Norm distances between classical and reproducing kernel smoothing splines
Filter length
Third order kernel 9-term 13-term 23-term
Weights Gain function Weights Gain function Weights Gain function
Sobolev space R2 0:144 3:213 0:428 1:925 0:482 1:709
Standard Laplace 0:049 1:088 0:608 3:863 0:583 2:916
Logistic ˇ D 0:2 0:021 0:480 0:271 0:588 0:260 0:471

where ! denotes the frequency in cycles per unit of time and G.!/ is the gain of the
filter. For illustrative purposes, we compute these measures for filter spans generally
applied to monthly time series, that is, 9-, 13-, and 23-term, even if filter of any
length can be considered. Table 9.2 shows that the third order standard Laplace ker-
nel presents large discrepancies for each span indicating that asymptotically it does
not approach to the CSS. On the other hand, the equivalent kernel representation
derived in the Sobolev space R2 provides the worst approximation for the shortest
filter length, but it has a better performance as the span increases.
Given the poor performance of the standard Laplace estimators, we looked
for another density function. To be coherent with the results of Eq. (9.21), we
considered the log F2m1 ;2m2 distribution family introduced by Prentice [28] which
includes the standard Normal (m1 ; m2 ! 0) and Laplace (m1 ; m2 ! 1) as limiting
cases. Many densities belong to this class of distributions, among others, the logistic
(m1 D m2 D 1) and the exponential (m1 ¤ 0; m2 ! 0). We concentrate on
the former given its strong connection with the standard Laplace and other widely
applied density functions, as recently shown by Lin and Hu [23]. These authors
modified and extended previous work by Mudholkar and George [24], providing
a characterization of the logistic density in terms of sample median and Laplace
distribution, as well as in terms of the smallest order statistics and exponential
density [20]. Furthermore, George and Ojo [16], and George et al. [15] studied the
logistic distribution as approximation of Student’s t functions. The logistic density
is defined as
 1   
1 2 1 t˛
f0L .t/ D sech ; t 2 .1; 1/; (9.23)
4ˇ 2 ˇ

where ˛ is the mean, set equal to 0, ˇ > 0 is the dispersion parameter, and sech is
the hyperbolic secant function.
Table 9.2 shows the results for the third order kernel within the hierarchy derived
by f0L , where the dispersion parameter has been set equal to 0.2. It is given by
  
5 5 21 2085 2
K3L .t/ D sech2 t  t : (9.24)
4 2 16 878

This kernel representation closely approximates the CSS as shown in Table 9.2.
234 9 A Unified View of Trend-Cycle Predictors in Reproducing Kernel Hilbert. . .

Fig. 9.1 13-Term (a) second and (b) third order kernels

Fig. 9.2 Gain functions of symmetric 13-term third order kernels

The logistic kernel representation of smoothing splines enables to compare

directly all the derived hierarchies. Figure 9.1a shows the density functions or
second order kernels within each family, whereas Fig. 9.1b illustrates the third
order smoothers. The 13-term filters are exhibited because the logistic kernel
representation of smoothing splines enables to compare directly in the time domain
all the hierarchies we have considered. But similar results are derived for 9- and
23-term smoothers. These hierarchies reproduce and describe several temporal
dynamics by predicting polynomial trends of different degrees that solve several
minimization problems. This is shown in Fig. 9.2, where we illustrate the gain
functions of 13-term third order kernels.
9.2 Boundary Behavior 235

LOESS and Henderson kernels present similar properties in terms of trend-cycle

estimation. They eliminate a large amount of noise and pass all the power associated
to the signal frequency band, here defined by 0 < !  0:06. However, they do not
suppress power at the frequency ! D 0:10, corresponding to cycles of 10 months,
known as unwanted ripples. On the other hand, the third order Gaussian and spline
kernels lose part of their optimality prediction properties, leaving untouched the
signal but passing a lot of noise including a large number of unwanted ripples.

9.2 Boundary Behavior

The kernels derived by means of the RKHS methodology provide a unified view of
presently used nonparametric trend-cycle estimators. The third order kernels in the
tricube, biweight, and logistic hierarchies are equivalent kernels for the classical
LOESS of degree 2 (LOESS 2), Henderson filter, and cubic smoothing spline,
respectively. No comparisons can be made for the third order Gaussian estimator
which is already a kernel function, and for which no counterpart exists in the
literature.
The reproducing kernel representation has important consequences in the deriva-
tion of the corresponding asymmetric smoothers, which are of crucial importance
in current analysis where the aim is to obtain the estimate of the trend-cycle for the
most recent seasonally adjusted values. We derived the asymmetric smoothers by
adapting the kernel functions to the length of the last m asymmetric filters, such that
 
j
K3 b
wq;j D P   j D m; : : : ; qI q D 0; : : : ; m  1;
q j
jDm K3 b

where j denotes the distance to the target point t .t D n  m C 1; : : : ; n/, b is

the bandwidth parameter selected in view of ensuring a symmetric filter of length
2m C 1, and m C q C 1; q D 0; : : : ; m  1, is the asymmetric
Pq filter length.
These asymmetric filters only satisfy the condition jDm wq;j D 1, hence will
reproduce without distortion only a constant on the asymmetric support. Given the
small number of points generally available in the boundaries, this condition can be
considered sufficient in view of obtaining a better performance of the kernel in terms
of mean square error.
The main goal is to analyze the goodness of the reproducing kernel representa-
tions versus the classical last point asymmetric filters.
Cleveland [5] showed that, in the middle of the series, LOESS acts as a
symmetric moving average with window length 2m C 1. At the end of the series, its
window length remains 2m C 1, rather than decreasing to m C 1 as in the case of the
most widely applied asymmetric last point trend-cycle estimators. As discussed by
Gray and Thomson [17], this implies a heavier than expected smoothing at both ends
236 9 A Unified View of Trend-Cycle Predictors in Reproducing Kernel Hilbert. . .

Fig. 9.3 (a) Gain and (b) phase shift functions of the asymmetric (end point) weights of the third
order tricube kernel and the classical LOESS 2

Fig. 9.4 (a) Gain and (b) phase shift functions of the asymmetric (end point) weights of the third
order biweight kernel and the classical Henderson filter

of the series with respect to the middle, and represents a drawback, particularly for
economic time series where turning points are important to be identified. As shown
in Fig. 9.3a, the last point LOESS asymmetric kernel exhibits a gain function with
better properties of signal passing and noise suppression relative to the classical one.
This implies smaller filter revisions as new data are added to the series. The phase
shifts for both filters, shown in Fig. 9.3b, are smaller than one month in the signal
frequency band 0 < !  0:06, usually assumed to correspond to the trend-cycle
component.
Similar conclusions can be derived for the last point asymmetric Henderson filter
which were developed by Musgrave [25], as shown in Fig. 9.4. They are based on the
minimization of the mean squared revisions between the final estimates (obtained
by the application of the symmetric filter) and the preliminary ones (obtained by
the application of an asymmetric filter) subject to the constraint that the sum of the
weights is equal to one [13]. The assumption made is that at the end of the series, the
seasonally adjusted values follow a linear trend-cycle plus a purely random irregular
"t , such that "t  NID.0;  2 /.
The asymmetric filters for the natural cubic smoothing splines are obtained by
adding additional constraints, ensuring that the function is of degree one beyond
the boundary knots. In this study, the asymmetric classical splines are obtained by
fixing the  parameter to have a symmetric filter of 2m C 1 terms, and then selecting
the last m rows of the influential matrix A./. We illustrate the results for the last
point asymmetric weights corresponding to a 23-term symmetric filter because the
9.2 Boundary Behavior 237

Fig. 9.5 (a) Gain and (b) phase shift functions of the asymmetric (end point) weights of the third
order spline kernel and the CSS smoother

spline gives very poor results for short lengths such as 13 or less. Figure 9.5a shows
that the asymmetric kernel exhibits a gain function with better properties of signal
passing and noise suppression, without implying larger phase shifts. These latter are
smaller than 1 month for both filters (see Fig. 9.5b).

9.2.1 Empirical Evaluation

This section examines how the kernel estimators perform on real seasonally adjusted
data in comparison with the LOESS, Henderson filter, and cubic smoothing spline.
The last point filters are applied to a set of fifty time series taken from the Hyd-
man’s time series library (http://www-personal.buseco.monash.edu.au/~hyndman/
TSDL/). These series are seasonally adjusted and related to different fields (finance,
labor market, production, sales, transports, meteorology, hydrology, physics, and
health). The periods selected vary to sufficiently cover the various lengths published
for these series. For each series, the length of the Henderson and LOESS filters is
selected according to the I=C (noise to signal) ratio that ranges from 0.16 to 7.35,
hence filters of length 9-, 13-, and 23-term are applied.
On the other hand, for the cubic smoothing spline the generalized cross validation
(GCV) criteria [6] is used to estimate the  parameter, and consequently the length
of the filters to be applied by considering the number of non-null elements in a
central row of the matrix A./. The kernel estimator of the same length is then
calculated. The smoothing parameter  is known as hyperparameter in the Bayesian
terminology, and it has the interpretation of a noise to signal ratio: the larger the ,
the smoother the output. In our sample,  ranges from a minimum of 0.013, at which
corresponds a filter length equal to 7 terms, to a maximum of 15, which corresponds
to a 43-term smoother. This enables us to analyze the pattern of the two estimators
on series characterized by different degrees of variability.
238 9 A Unified View of Trend-Cycle Predictors in Reproducing Kernel Hilbert. . .

The comparisons are based on the relative filter revisions between the final
symmetric filter S and the last point asymmetric filter A, that is,

S t  At
Rt D ; t D 1; 2; : : : ; n (9.25)
St

For each series and for each estimator, it is calculated the ratio between the mean
square error (MSE) of the revisions corresponding to the third order kernel RK and
K/
to the corresponding classical filter RC , that is MSE.R
MSE.RC /
.
For all the estimators, the results illustrated in Table 9.3 indicate that the ratio
is always smaller than one, showing that the kernel last point predictors introduce
smaller revisions than the classical ones. This implies that the kernel estimates will
be more reliable and efficient than the ones obtained by the application of the
classical LOESS, Henderson filter, and cubic spline. In particular, for the LOESS
estimator, 96 % of the sample has a ratio less than 0.7 and, in general, it is never
greater than 0.806. For the Henderson filter, the ratio is less than 0.7 in 20 % of the
series and for the cubic spline in 34 % of the sample. In both cases, the highest value
of the ratio is close to 0.895.
For current economic analysis, it is also of great importance to assess the
performance of the several kernel estimators for turning point detection. A turning
point is generally defined to occur at time t if (downturn):

ytk      yt1 > yt  ytC1      ytCm

or (upturn)

ytk      yt1 < yt  ytC1      ytCm

for k D 3 and m D 1 given the smoothness of the trend-cycle data [8]. For each
estimator, the time lag to detect the true turning point is obtained by calculating the
number of months it takes for the revised trend series to signal a turning point in the
same position as in the final trend series.
True turning points are those recognized in the economy after occurring. For
instance, the Business Cycle Dating Committee of the National Bureau of Economic
Research (NBER) had identified the last trough in the US economy in June 2009.
To determine this date, the behavior of various monthly indicators of the economic
activity has been analyzed. Among others, the committee considered a measure of
monthly Gross Domestic Product (GDP) developed by the private forecasting firm
Macroeconomic Advisers (MAGDP), monthly GDP and Gross Domestic Income
(GDI) developed by James H. Stock and Mark W. Watson (SWGDP, SWGDI),
real personal income excluding transfers (INC), the payroll (PAY) and household
measures of total employment (HOU). Monthly data series for industrial production
(IIP) and manufacturing trade sales (MAN) have also been taken into account. The
committee designated June as the month of the trough for several of these indicators,
that is, MAGDP, SWGDP, SWGDI, MAN, and IIP. On the other hand, the trough
was identified in October 2009 for INC, and in December 2009 for PAY and HOU.
9.2 Boundary Behavior 239

Table 9.3 MSE revision ratio between kernel and classical last point predictors
Series Ratio
Health Number of cases of measles, New York city 0.806
Number of cases of measles, Baltimore 0.512
Body weight of rats 0.544
Number of chickenpox, New York city 0.597
Hydrology Temperature, copper mine 0.539
Flows, Colorado river Lees ferry 0.580
Lake Frie levels 0.568
Flows, Chang Jiang 0.696
Labour market Wisconsin employment series, fabricated metals 0.502
US male (20 years and over) unemployment figures 0.540
Number of employed persons in Australia 0.318
Macroeconomics Consumer price index 0.346
Meteorology Degree days per heating in Chicago 0.541
Microeconomics Gambling expenditure in Victoria, Australia 0.584
Logged flour price indices over the 9-years 0.558
Miscellaneous Average daily calls to directory assistance 0.607
Physics Zurich sunspot numbers 0.546
Critical radio frequencies in Washington DC 0.621
Mean thickness (Dobson units) ozone column Switzerland 0.685
Production Basic iron production in Australia 0.696
Production of chocolate confectionery in Australia 0.628
Production of Portland cement 0.687
Electricity production in Australia 0.783
Production of blooms and slabs in Australia 0.666
Production of blooms and slabs 0.677
Sales Sales of Tasty Cola 0.459
Unit sales, Winnebago Industries 0.594
Sales of new one-family houses sold in US 0.603
Sales of souvenir shop in Queensland, Australia 0.570
Demand for carpet 0.502
Transport and Tourism Portland Oregon average monthly bus ridership 0.511
US air passenger miles 0.546
International airline passengers 0.473
Weekday bus ridership, Iowa city (monthly averages) 0.567
Passenger miles flow domestic UK 0.703
Utilities Average residential gas usage Iowa 0.633
Total number of consumer 0.615
240 9 A Unified View of Trend-Cycle Predictors in Reproducing Kernel Hilbert. . .

Table 9.4 Time lag in detecting true turning points (TP) for classical and kernel estimators
LOESS Henderson CSS
Series True TP Length Classic Kernel Length Classic Kernel Length Classic Kernel
MAGDP 06/09 13 6 4 13 4 3 27 6 4
SWGDP 06/09 13 6 4 13 4 3 29 6 6
SWGDI 06/09 13 6 4 13 5 3 21 4 4
MAN 06/09 9 4 3 9 4 2 33 8 8
IIP 06/09 9 4 3 9 2 2 25 6 6
INC 10/09 13 6 4 13 4 3 17 6 6
PAY 12/09 9 4 3 9 4 4 13 6 5
HOU 12/09 9 4 3 9 2 1 21 6 6

The results obtained by applying the classical and kernel estimators to all of these
series are shown in Table 9.4.
It can be noticed that kernel estimators perform better than classical filters.
In particular, the third order biweight kernel always detects the turning points only
very few months after it has occurred, and faster than the classical Henderson filter,
which is very well-known for fast turning point detection. Tricube LOESS kernel
also presents good time lags, not greater than 4 months, that are smaller than those
observed for the classical LOESS estimator. On the other hand, the length selected
for the classical and kernel CSS estimators is generally greater than 20 terms and,
as expected, they are generally poor predictors of the true turning points, with a
slightly better performance for the logistic kernel than for the classical estimator.

References

1. Aronszajn, N. (1950). Theory of reproducing kernels. Transaction of the AMS, 68, 337–404.
2. Berlinet, A. (1993). Hierarchies of higher order kernels. Probability Theory and Related Fields,
94, 489–504.
3. Berlinet, A., & Thomas-Agnan, C. (2003). Reproducing kernel Hilbert spaces in probability
and statistics. Boston, MA: Kluwer Academic Publishers.
4. Bianconcini, S. (2006). Trend-Cycle Estimation in Reproducing Kernel Hilbert Spaces (Ph.D.
thesis, Department of Statistics, University of Bologna).
5. Cleveland, W. S. (1979). Robust locally regression and smoothing scatterplots. Journal of the
American Statistical Association, 74, 829–836.
6. Craven, P., & Wahba, G. (1979). Smoothing noisy data with spline functions. Numerical
Mathematics, 31, 377–403.
7. Cristianini, N., & Shawe-Taylor, J. (2000). An introduction to support vector machines and
other kernel-based learning methods. New York: Cambridge University Press.
8. Dagum, E. B. (1996). A new method to reduce unwanted ripples and revisions in trend-cycle
estimates from X11ARIMA. Survey Methodology, 22(1), 77–83.
9. Dagum, E. B., & Bianconcini, S. (2006). Local polynomial trend-cycle predictors in reproduc-
ing kernel Hilbert spaces for current economic analysis. Anales de Economia Aplicada, 1–22.
10. Dagum, E. B., & Bianconcini, S. (2008). The Henderson smoother in reproducing kernel
Hilbert space. Journal of Business and Economic Statistics, 26(4), 536–545.
References 241

11. Dagum, E. B., & Bianconcini, S. (2013). A unified probabilistic view of nonparametric
predictors via reproducing kernel Hilbert spaces. Econometric Reviews, 32(7), 848–867.
12. De Boor, C., & Lynch, R. (1966). On splines and their minimum properties. Journal of
Mathematics and Mechanics, 15, 953–969.
13. Doherty, M. (1992). The Surrogate Henderson Filters in X-11. Working Paper. Wellington,
New Zealand: Statistics New Zealand.
14. Evgeniou, T., Pontil, M., & Poggio, T. (2000). Regularization networks and support vector
machines. Advanced in Computational Mathematics, 13, 1–50.
15. George, E. O., El-Saidi, M., & Singh, K. (1986). A generalized logistic approximation of the
Student t distribution. Communications in Statistics B - Simulation Computing, 15(1), 261–277.
16. George, E. O., & Ojo, M. O. (1980). On a generalization of the logistic distribution. Annals of
the Institute of Statistical Mathematics, 32, 161–169.
17. Gray, A., & Thomson, P. (1996). On a family of moving-average trend filters for the ends of
series. In Proceedings of the Business and Economic Statistics Section. American Statistical
Association Annual Meeting, Chicago.
18. Green, P. J., & Silverman, B. W. (1994). Nonparametric regression and generalized linear
models. London: Chapman and Hall.
19. Henderson, R. (1916). Note on graduation by adjusted average. Transactions of the Actuarial
Society of America, 17, 43–48.
20. Jones, M. C. (2006). The logistic and the log F distributions. In N. Balakrishan (Ed.), Handbook
of logistic distribution. New York: Dekker.
21. Kimeldorf, G. S., & Wahba, G. (1970). A correspondence between Bayesian estimation
on stochastic processes and smoothing by splines. Annals of Mathematical Statistics, 41,
495–502.
22. Kimeldorf, G. S., & Wahba, G. (1970). Splines functions and stochastic processes. Sankhya,
Series A, 32, 173–180.
23. Lin, G. D., & Hu, C. Y. (2008). On the characterization of the logistic distribution. Journal of
Statistical Planning and Inference, 138(4), 1147–1156.
24. Mudholkar, G. S., & George, E. O. (1978). A remark on the shape of the logistic distribution.
Biometrika, 65, 667–668.
25. Musgrave, J. (1964). A Set of End Weights to End All End Weights. Working paper. Washington:
US Bureau of the Census.
26. Parzen, E. (1959). Statistical Inference on Time Series by Hilbert Space Methods. Technical
Report No. 53. , Stanford, CA: Statistics Department, Stanford University.
27. Pearce, N. D., & Wand, M. P. (2006). Penalized splines and reproducing kernel methods. The
American Statistician, 60(3), 223–240.
28. Prentice, R. L. (1976). A generalization of probit and logit methods for dose response curves.
Biometrics, 32, 761–768.
29. Priestley, M. B. (1981). Spectral analysis and time series. Probability and mathematical
statistics. New York: Academic.
30. Schoenberg, I. (1964). Monosplines and quadrature formulae. In T. Greville (Ed.), Theory and
applications of spline functions. Madison, WI: University of Wisconsin Press.
31. Silverman, B. (1994). Spline smoothing: The equivalent kernel method. Annals of Statistics,
12, 898–916.
32. Thomas-Agnan, C. (1991). Splines functions and stochastic filtering. Annals of Statistics, 9(3),
1512–1527.
33. Wahba, G. (1990). Spline models for observational data. Philadelphia: SIAM.
34. Wahba, G. (1999). Support vector machine, reproducing kernel Hilbert spaces, and randomized
GACV. In B. Scholkopf, C. Burges, & A. Smola (Eds.), Advanced in kernel methods: Support
vector learning (pp. 69–88). Cambridge, MA: MIT Press.
35. Whittaker, E. T. (1923). On a new method of graduation. In Proceedings of the Edinburgh
Mathematical Association (Vol. 78, pp. 81–89).
36. Whittaker, E. T., & Robinson, G. (1924). Calculus of observations: A treasure on numerical
calculations. London: Blackie and Son.
Chapter 10
Real Time Trend-Cycle Prediction

Abstract For real time trend-cycle analysis, official statistical agencies generally
produce estimates derived from asymmetric moving average techniques. However,
the use of the asymmetric filters introduces revisions as new observations are incor-
porated to the series and also delays in detecting true turning points. This chapter
deals with a reproducing kernel approach to obtain time-varying asymmetric trend-
cycle filters that converge fast and monotonically to the corresponding symmetric
ones. It shows that the bandwidth parameter that minimizes the distance between the
gain functions of the asymmetric and symmetric filters is to be preferred. Respect
to the asymmetric filters incorporated in the X12ARIMA software currently applied
by the majority of official statistical agencies, the new set of weights produces both
smaller revisions and time delay to detect true turning points. The theoretical results
are empirically corroborated with a set of leading, coincident, and lagging indicators
of the US economy.

The basic approach to the analysis of current economic conditions, known as

recession and recovery analysis, is that of assessing the real time trend-cycle of
major socioeconomic indicators (leading, coincident, and lagging) using percentage
changes, based on seasonally adjusted units, calculated for months and quarters
in chronological sequence. The main goal is to evaluate the behavior of the
economic indicators during incomplete phases by comparing current contractions
or expansions with corresponding phases in the past. This is done by measuring
changes of single time series (mostly seasonally adjusted) from their standing
at cyclical turning points with past changes over a series of increasing spans.
This differs from business-cycle studies where cyclical fluctuations are measured
around a long-term trend to estimate complete business cycles. The real time trend
corresponds to an incomplete business cycle and is strongly related to what is
currently happening on the business-cycle stage.
In recent years, statistical agencies have shown an interest in providing trend-
cycle or smoothed seasonally adjusted graphs to facilitate recession and recovery
analysis. Among other reasons, this interest originated from the recent crisis and
major economic and financial changes of global nature which have introduced
more variability in the data. The USA entered in recession in December 2007 till
June 2009, and this has produced a chain reaction all over the world, particularly,

© Springer International Publishing Switzerland 2016 243

E. Bee Dagum, S. Bianconcini, Seasonal Adjustment Methods and Real Time
Trend-Cycle Estimation, Statistics for Social and Behavioral Sciences,
DOI 10.1007/978-3-319-31822-6_10
244 10 Real Time Trend-Cycle Prediction

in Europe. There is no evidence of a fast recovery as in previous recessions:

the economic growth is sluggish and with high levels of unemployment. It has
become difficult to determine the direction of the short-term trend (or trend-cycle)
as traditionally done by looking at month to month (quarter to quarter) changes
of seasonally adjusted values, particularly to assess the upcoming of a true turning
point. Failure in providing reliable real time trend-cycle estimates could give rise
to dangerous drift of the adopted policies. Therefore, a consistent prediction is of
fundamental importance. It can be done by means of either univariate parametric
models or nonparametric techniques. The majority of the statistical agencies use
nonparametric seasonally adjusted software, such as the Census II-X11 method and
its variants X11/X12ARIMA and X13, and hence this chapter deals with the real
time trend-cycle estimation produced by the Musgrave filters [19] available in these
software.
As widely discussed in Chap. 8, the linear filter developed by Henderson [15]
is the most frequently applied and has the property that fitted to exact cubic
functions will reproduce their values, and fitted to stochastic cubic polynomials
it will give smoother results than those estimated by ordinary least squares. The
properties and limitations of the Henderson filters have been extensively discussed
in Chap. 8, where we have also provided its Reproducing Kernel Hilbert Space
(RKHS) representation. It consists in a kernel function obtained as the product of
the biweight density function and the sum of its orthonormal polynomials that is
particularly suitable when the length of the filter is rather short, say between 5 and
23 terms, which are those often applied by statistical agencies (see also, for details,
Dagum and Bianconcini [6]).
At the beginning and end of the sample period, the Henderson filter of length, say
2m C 1, cannot be applied to the first and last m data points, hence only asymmetric
filters can be used. The estimates of the real time trend are then subject to revisions
produced by the innovations brought by the new data entering in the estimation and
the time-varying nature of the asymmetric filters, in the sense of being different
for each of the m data points. The asymmetric filters applied to the first and last m
observations associated with the Henderson filter were developed by Musgrave [19]
on the basis of minimizing the mean squared revision between the final estimates,
obtained with the symmetric Henderson weights, and preliminary estimates from
the asymmetric weights, subject to the constraint that the sum of these weights is
equal to one. The assumption made is that at the end of the series, the seasonally
adjusted values do not follow a cubic polynomial, but a linear trend-cycle plus a
purely random irregular. Several authors studied the properties and limitations of
the Musgrave filters [2, 10–12, 14, 18, 20], and Dagum and Bianconcini [6, 7] were
the first to introduce an RKHS representation of them.
The RKHS approach followed in this book and presented here is strictly nonpara-
metric. As discussed in Chaps. 8 and 9, based on a fundamental result due to Berlinet
[1], a kernel estimator of order p can be always decomposed into the product of
a reproducing kernel Rp1 , belonging to the space of polynomials of degree at
most p  1, and a density function f0 with finite moments up to order 2p. Given
the density function, once the length of the symmetric filter is chosen, let us say,
10.1 Asymmetric Filters and RKHS 245

2m C 1, the statistical properties of the asymmetric filters are strongly affected by

the bandwidth parameter of the kernel function from which the weights are derived.
In Chap. 9, as suggested by Dagum and Bianconcini [6, 7], this bandwidth parameter
has been selected to be equal to m C 1, independently on the length of the associated
asymmetric filter. The corresponding asymmetric weights approximate very well
the classical Musgrave filters [19]. This chapter presents time-varying bandwidth
parameters because the asymmetric filters are time-varying. Three specific criteria
of bandwidth selection are chosen based on the minimization of
(1) the distance between the transfer functions of asymmetric and symmetric filters,
(2) the distance between the gain functions of asymmetric and symmetric filters,
and
(3) the phase shift function over the domain of the signal.
It deals only with the reduction of revisions due to filter changes that depends on
how close the asymmetric filters are respect to the symmetric one [5, 9] and do not
consider those introduced by the innovations in the new data. Another important
aspect dealt with is the capability of the asymmetric filters to signal the upcoming
of a true turning point which depends on the time delay for its identification. This is
obtained by calculating the number of months (quarters) it takes for the last trend-
cycle estimate to signal a true turning point in the same position of the final trend-
cycle data. An optimal asymmetric filter should have a time path that converges fast
and monotonically to the final estimate as new observations are added to the series.

10.1 Asymmetric Filters and RKHS

Let fyt ; t D 1; 2; : : : ; ng denote the input series, supposed to be seasonally adjusted

where trading day variations, moving holidays, and outliers, if present, have been
removed. It is assumed that it can be decomposed into the sum of a systematic
component (signal) TCt , representing the trend-cycle (usually estimated jointly)
plus an erratic component It (noise), such that

yt D TCt C It : (10.1)

The noise It is assumed to be white noise, WN.0; I2 /, or, more generally, a
stationary and invertible autoregressive moving average (ARMA) process. The
signal TCt ; t D 1; : : : ; n; is assumed to be a smooth function of time, such that it can
be represented locally by a polynomial of degree p in a variable j; which measures
the distance between yt and its neighboring observations ytCj ; j D m; : : : ; m.
This is equivalent to estimate the trend-cycle TC c t as a weighted moving average
as follows:

X
m
ct D
TC wj ytCj D w0 y t D m C 1; : : : ; n  m; (10.2)
jDm
246 10 Real Time Trend-Cycle Prediction

 
where w0 D wm    w0    wm contains the weights to be applied to the input
 
c t for each point in time.
data y0 D ytm    yt    ytCm to get the estimate TC
As discussed in Sect. 9.2, several nonparametric estimators, based on different
sets of weights w, have been developed in the literature. It is shown in Sect. 8.3.2
that the Henderson filter can be derived using the Reproducing Kernel Hilbert Space
methodology. In this context, its equivalent kernel representation is given by

3
X
K4 .t/ D Pi .t/Pi .0/f0B .t/ t 2 Œ1; 1 ; (10.3)
iD0

where f0B is the biweight density function f0B .t/ D 1516

.1  t2 /2 ; t 2 Œ1; 1 , and
Pi ; i D 0; 1; 2; 3; are the corresponding orthonormal polynomials. Equation (10.3)
provides a good approximation for Henderson filters of short length, say between 5
and 23 terms, which are those used by statistical agencies [6].
Dagum and Bianconcini [8] have shown that, when applied to real data, the
symmetric filter weights are given by
2  2 3
4  2 j  
6 b71 j
wj D 4 5 f0B j D m; : : : ; m; (10.4)
S0 4  S2 2 b b

where b is a time invariant global bandwidth parameter (same for all t D m C

1; : : : ; n  m), whereas 2 and 4 are the second and  fourth
r  moments
 of the
Pm 1 j
biweight density, respectively, being Sr D jDm b b
j
f0 b their discrete
approximations.
In this setting, once the length of the filter is selected, the choice of the bandwidth
parameter b is fundamental. It has to be chosen to ensure that only, say, 2m C 1
observations surrounding the target point will receive nonzero weights. It has also to
approximate, as close as possible, the continuous density function with the discrete
function as well as its moments. The bandwidth parameter selection is done to
guarantee specific inferential properties of the trend-cycle estimators. In this regard,
Dagum and Bianconcini [6, 7] used a time invariant global bandwidth b equal to
m C 1, which gave good results.
The derivation of the symmetric Henderson filter has assumed the availability
of 2m C 1 input values centered at t. However, at the end of the sample period,
that is t D n  .m C 1/; : : : ; n, only 2m; : : : ; m C 1 observations are available, and
asymmetric filters of the same length have to be considered. Hence, at the boundary,
the effective domain of the kernel function K4 is Œ1; q , with q << 1, instead
of Œ1; 1 as for any interior point. This implies that the symmetry of the kernel is
R q
lost, and it does not integrate to unity on the asymmetric support ( 1 K4 .t/dt ¤ 1).
R q
Furthermore, the moment conditions are not longer satisfied, that is, 1 ti K4 .t/dt ¤
10.1 Asymmetric Filters and RKHS 247

0 for i D 1; 2; 3. To overcome these limitations, several boundary kernels have been

proposed in the literature.
In the context of real time trend-cycle estimation, the condition that the kernel
function integrates to unity is essential, whereas the unbiasedness property can
only be satisfied with a great increase in the variance of the estimates. This is a
consequence of the well-known trade-off between bias and variance. This latter
becomes very large because most of the contribution to the real time trend-cycle
estimates comes from the current observation which gets the largest weight. Based
on these considerations, Dagum and Bianconcini [6, 7] have suggested to follow
the so-called cut and normalize method [13, 16], according to which the boundary
q
kernels K4 are obtained by cutting the symmetric kernel K4 to omit that part of the
function lying between q and 1, and by normalizing it on [1,q ]. That is,

q K4 .t/ det.H04 Œ1; t /f0B .t/

K4 .t/ D R q D t 2 Œ1; q (10.5)
det.H04 Œ1; q /
1 K4 .t/dt
 q q q  R q r
where q D q 0 1 2 3 with q
r D 1 t f0B .t/dt being proportional to
the moments of the truncated biweight density f0B on the support Œ1; q , which
from now on we simply refer to as truncated moments. H04 Œ1; t is the Hankel
matrix whose elements are themoments of f0B , and where the first column has been
substituted by the vector t0 D 1 t t2 t3 , whereas H04 Œ1; q / is the Hankel matrix
where the first column is substituted by q .
Applied to real data, the “cut and normalize” method gives the following formula
for the asymmetric weights:

q
K4 .j=bq / det.H04 Œ1; j=bq /.1=bq /f0B .j=bq /
wq;j D Pq q
D ; (10.6)
K4 .j=bq / det.Ha /
jDm

j D m; : : : ; qI q D 0; : : : ; m  1;

where H04 Œ1; j=bq is the Hankel matrix whose elements are the moments of
f0B , and where the first  column has been substituted by the vector j=bq 0 D
1 .j=bq / .j=bq /2 .j=bq /3 . On the other hand, Ha D H04 Œ1; Sq with Sq D
 q q q q 0 Pq  r  
1 j j
S0 S1 S2 S3 , being Srq D jDm bq bq f0B bq the discrete approximation
of r . Finally, bq ; q D 0; : : : ; m  1, is the local bandwidth, specific for each
q

asymmetric filter. It allows to relate the discrete domain of the filter, that is,
fm; : : : ; qg, for each q D 0; : : : ; m  1, to the continuous domain of the kernel
function, that is Œ1; q .
Proposition 1 Each asymmetric filter wq D Œwq;m    wq;q 0 of length .m C q C
1/; q D 0; : : : ; m  1, admits the following matrix representation:

w0q D e01 Ha 1 X0q Fq q D 0; : : : ; m  1; (10.7)

248 10 Real Time Trend-Cycle Prediction

where Xq is a matrix of dimensions .m C q C 1/   4, whose generic row is given

 
1 1 q
by j=bq ; j D m; : : : ; q, and Fq D diag bq f0B  bq ; : : : ; bq f0B bq . It can be
m

easily shown that the generic element of wq is

2  2 3
4  2 j  
6 bq 7 1 j
wq;j D4 q q 5 f0B (10.8)
S0 4  S2 2 bq bq

j D m; : : : ; qI q D 0; : : : ; m  1:

Proof Based on Eq. (10.6), we can write

   
det.H04 Œ1; j=bq / b1q f0B j
bq
det.H04 Œ1; j=bq / b1q f0B bjq
wq;j D P  D  h P  i
1
0
det H04 1; jDm j=bq b1q f0B bjq
q q
jDm det.H4 Œ1; j=bq / bq f0B bq
j

   
det.H04 Œ1; j=bq / b1q f0B bjq det.H04 Œ1; j=bq / b1q f0B bjq
D D : (10.9)
det.H04 Œ1; Sq / det.Ha /

The expression (10.9) is exactly the same we would obtain by solving for ˇO0 D TC
ct
the system of linear equations

Ha ˇ D X0q Fq y:

Indeed, setting c D X0q Fq y, the first coordinate of the solution vector is

Q 0 Œ1; c /
det.H det.H04 Œ1; c /
ˇO0 D 4
D :
det.Ha / det.Ha /
Pq    
1 j
Given that c D jDm
j
bq bq f0B bq ytCj , it follows that

X
q  
j 1 j
det.H04 Œ1; bq / D det.H04 Œ1; / f0B ytCj
jDm
bq bq bq

and therefore
 
X
q det.H04 Œ1; j=bq / b1q f0B j
bq
ct D
TC ytCj :
jDm
det.Ha /
10.1 Asymmetric Filters and RKHS 249

Hence,

ˇO0 D e01 Ha 1 X0q Fq y;

and it follows that

w0 D e01 Ha 1 X0q Fq :

10.1.1 Properties of the Asymmetric Filters

Since the trend-cycle estimates for the last m data points do not use 2m C 1
observations as for any interior point, but 2m; 2m  1; : : : ; m C 1 data, they are
subject to revisions due to new observations entering in the estimation and filters
change. As said before, we will concentrate on the reduction of revisions due to
filters change. The reduction of these revisions is an important property that the
asymmetric filters should possess together with a fast detection of true turning
points. In the specific case of the RKHS filters, Eq. (10.8) shows how the asymmetric
filter weights are related to the symmetric ones given in Eq. (10.4). It is clear that
the convergence depends on the relationship between the two discretized biweight
density functions, truncated and non-truncated, jointly with the relationship between
their respective truncated Srq and untruncated Sr discrete moments. The latter provide
an approximation of the continuous moments r , which improves as the asymmetric
filter length increases. Similarly, the convergence of Srq ; q D 0; : : : ; m, to the
corresponding non-truncated moment Sr depends on the length of the asymmetric
filter given by q, and on the local bandwidth bq . It should be noticed that bq plays
a very important role in the convergence property. For the last trend-cycle point
weight, q D 0, Eq. (10.8) reduces to

4 15
w0;0 D :
S00 4 0
 S2 2 16b0

It is apparent that the largest b0 , the smaller is the weight given to the last
trend-cycle point. Since the sum of all the weights of the last point asymmetric
filter, w0;m ; : : : ; w0;0 , must be equal to one, this implies that the weights for the
remaining points are very close to one another. This can be seen in Fig. 10.1 (right
side) that shows, for m D 6, the truncated continuous biweight density function and
its discretized version when b0 is equal to 12. The opposite is observed when b0
is smaller, as shown in the same figure (left side) for b0 equal to 7. Since a larger
weight is given to the last point, much smaller weights have to be assigned to the
remaining ones for all of them to add to one. Next, we introduce time-varying local
bandwidths to improve the properties of the asymmetric filters in terms of size of
revisions and time delay to signal the upcoming of true turning points.
250 10 Real Time Trend-Cycle Prediction

0∗ 0∗
0 0

0.8
0.8

b0 = 7 b0 = 12

0.4
0.4

0.0
0.0

−1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0
t t

Fig. 10.1 Behavior of S00 with m D 6, b0 D 7 (left) and b0 D 12 (right)

10.2 Optimal Bandwidth Selection

The main effects introduced by a linear filter on a given input can be fully described
in the frequency domain by its transfer function

X
m
 .!/ D wj exp.i2!j/ ! 2 Œ1=2; 1=2 ;
jDm

where, for better interpretation, the frequencies ! are given in cycles for unit of
time instead of radians.  .!/ represents the Fourier transform of the filter weights,
wj ; j D m; : : : ; m, and it relates the spectral density hy .!/ and hg .!/ of the input
and of the output, respectively, by

hg .!/ D  .!/hy .!/:

Thus, the transfer function  .!/ measures the effect of the filter on the total variance
of the input at different frequencies. It is generally expressed in polar coordinates

 .!/ D G.!/ exp.i2.!//; (10.10)

such that the impact of the filter on a particular (complex-valued) series yt D

exp.i2!t/; ! 2 Œ1=2; 1=2 , is given by

c t D  .!/ exp.i2!t/
TC
D G.!/ exp.i2.!// exp.i2!t/
D G.!/ expŒi2.!t  .!// :
10.2 Optimal Bandwidth Selection 251

G.!/ D j .!/j is called the gain of the filter and measures the amplitude of the
output for a sinusoidal input of unit amplitude, whereas .!/ is called the phase
function and shows the shift in phase of the output compared with the input. Hence,
the transfer function plays a fundamental role to measure that part of the total
revisions due to filters change.
A measure of total revisions introduced by Musgrave [19] is given by
2 32
X
q
X
m
E4 wq;j ytj  wj ytj 5 ; q D 0; : : : ; m  1; (10.11)
jDm jDm

where, in our specific case, wq;j and wj are given by Eqs. (10.8) and (10.4),
respectively. This criterion can be expressed in the frequency domain as follows:
2 32
X
q
X
m
 2
E4 wq;j ei2!.tj/  wj ei2!.tj/ 5 D E .q .!/   .!//ei2!t
jDm jDm
Z 1=2
D jq .!/   .!/j2 ei4!t hy .!/d!; (10.12)
1=2

where hy .!/ is the unknown spectral density of yt ; t D 1; : : : ; n, whereas q .!/ and

 .!/ are the transfer functions corresponding to the asymmetric and symmetric
filters, respectively. Similarly to (10.11), expression (10.12) shows that, as new
observations become available, revisions are produced by the new innovations
entering the input series and the change of the asymmetric filters. In order to
improve the current trend-cycle prediction based on the asymmetric Henderson
filters we study that part of the revisions due to asymmetric filters change. Because
the estimation of the real time trend-cycle is done concurrently, that is using all
of the data up to and including the most recent value, knowledge of the speed of
convergence of the last point trend-cycle filter to the central one gives valuable
information on how often the real time trend estimate should be revised.
The quantity jq .!/   .!/j2 accounts for the revisions due to filters change
[3, 4], and it can be decomposed using the law of cosines as follows:
 
jq .!/   .!/j2 D jGq .!/  G.!/j2 C 2Gq .!/G.!/ 1  cos.q .!// (10.13)
  ! 2
D jGq .!/  G.!/j2 C 4Gq .!/G.!/ sin q ;
2
where the phase shift for the symmetric filter is equal to 0 or ˙, and where 1 

2
cos.q .!// D 2 sin q !2 . Based on Eq. (10.13), the mean square filter revision
252 10 Real Time Trend-Cycle Prediction

error can be expressed as follows:

Z 1=2
2 jq .!/   .!/j2 d! (10.14)
0
Z 1=2 Z 1=2   ! 2
D2 jGq .!/  G.!/j2 d! C 8 Gq .!/G.!/ sin  d!:
0 0 2

The first component reflects the part of the total mean square filter error which is
attributed to the amplitude function of the asymmetric filter. On the other hand,
the second term measures the distinctive contribution of the phase shift. The term
Gq .!/G.!/ is a scaling factor which accounts for the fact that the phase function is
dimensionless, i.e., it does not convey level information [21].
As previously discussed, once the length of the filter is fixed, the properties of
the asymmetric filters derived in RKHS are strongly affected by the choice of the
time-varying local bandwidths bq ; q D 0; : : : ; m  1. A filter is said to be optimal if
it minimizes both revisions and time delay to detect a true turning point.
The LHS of Eq. (10.14) is a measure of total filter revision that provides the best
compromise between the amplitude function of the asymmetric filter (gain) and its
phase function (time displacement) [3, 4, 9]. Optimal asymmetric filters in this sense
can be derived using local bandwidth parameters selected according to the following
criterion:
s
Z 1=2
bq; D min 2 jq .!/   .!/j2 d!: (10.15)
bq 0

Based on the decomposition of the total filter revision error provided in Eq. (10.14),
further bandwidth selection criteria can be defined by emphasizing more the gain or
phase shift effects, and/or by attaching varying importance to the different frequency
components, depending on whether they appear in the spectrum of the input time
series or not. In the context of smoothing a monthly input, the frequency domain
˝ D f0  !  0:50g can be partitioned into two main intervals: (1) ˝S D f0 
!  0:06g associated with cycles of 16 months or longer attributed to the signal
(trend-cycle) of the series and (2) ˝N S D f0:06 < !  0:50g corresponding to short
cyclical fluctuations attributed to the noise.
A class of optimal asymmetric filters based on bandwidth parameters bq ; q D
0; : : : ; m  1, is selected as follows:
s
Z 1=2
bq;G D min 2 jGq .!/  G.!/j2 d!; (10.16)
bq 0

and
s Z
 
bq; D min 2 Gq .!/G.!/ 1  cos.q .!// : (10.17)
bq ˝S
10.2 Optimal Bandwidth Selection 253

Table 10.1 Optimal bandwidth values selected for each of the biweight asymmetric filters
corresponding to the 9-, 13-, and 23-term Henderson symmetric filters
q 0 1 2 3
bq; 6.47 5.21 4.90 4.92
bq;G 8.00 5.67 4.87 4.90
bq; 4.01 4.45 5.97 6.93
q 0 1 2 3 4 5
bq; 9.54 7.88 7.07 6.88 6.87 6.94
bq;G 11.78 9.24 7.34 6.85 6.84 6.95
bq; 6.01 6.01 7.12 8.44 9.46 10.39
q 0 1 2 3 4 5 6 7 8 9 10
bq; 17.32 15.35 13.53 12.47 12.05 11.86 11.77 11.77 11.82 11.91 11.98
bq;G 21.18 18.40 16.07 13.89 12.44 11.90 11.72 11.73 11.83 11.92 11.98
bq; 11.01 11.01 11.01 11.01 11.41 13.85 15.13 16.21 17.21 18.15 19.05

It has to be noticed that the minimization of the phase error in Eq. (10.17) is very
close to minimizing the average phase shift in month for the signal, that is
 Z 
1 .!/
bq; D min d! : (10.18)
bq 0:06 ˝S 2!

Table 10.1 illustrates the bandwidth parameters bq; , bq;G ; bq; ; q D 0; : : : ; m  1;

derived as minimizers of Eqs. (10.15), (10.16), and (10.18), respectively, corre-
sponding to the 9-, 13-, and 23-term symmetric filters. It can be noticed that, as q
approaches m, the bandwidth parameters selected to optimize the criteria (10.15)
and (10.16) get closer to m C 1, which is the global bandwidth considered for
the symmetric Henderson filter. Hence, based on the relationships between trun-
cated and untruncated discrete biweight density functions and respective discrete
moments previously discussed, the asymmetric filters based on bq; and bq;G ,
q D 0; : : : ; m  1, should be characterized by a fast convergence to the symmetric
filter. This is confirmed by Fig. 10.2 that illustrates, as an example, the time path of
these filters corresponding to the 13-term symmetric one. Similar conclusions can
be drawn for different filter lengths.
The asymmetric filters based on bq; and bq;G , q D 0; : : : ; m  1, converge
very fast to the symmetric filter, particularly after the previous to the last point,
with the main differences observed for the last point filters. For these latter, the
different behavior is analyzed in the frequency domain in Fig. 10.4, that shows the
corresponding gain and phase shift functions. It can be noticed that, as expected,
the filter whose bandwidth b0;G is derived as minimizer of Eq. (10.16) shows a gain
function closer to that of the symmetric Henderson filter than the one based on b0; ,
suppressing more noise at the highest frequencies, and it reproduces very well the
signal in the lower frequency band.
In terms of phase shift or time delay, the filters that behave better are the ones
based on the bandwidth parameters selected to minimize the average phase shift
254 10 Real Time Trend-Cycle Prediction
0.8

0.8
S q1 q3 q5 S q1 q3 q5

q0 q2 q4 q0 q2 q4
0.6

0.6
0.4

0.4
0.2

0.2
0.0

0.0
−6 −4 −2 0 2 4 6 −6 −4 −2 0 2 4 6
j j

Fig. 10.2 Time path of the asymmetric filters based on bq; (left), bq;G (right) corresponding to
the 13-term symmetric filter
0.8

0.8

S q1 q3 q5 S q1 q3 q5

q0 q2 q4 q0 q2 q4
0.6

0.6
0.4

0.4
0.2

0.2
0.0

0.0

−6 −4 −2 0 2 4 6 −6 −4 −2 0 2 4 6
j j

Fig. 10.3 Time path of the asymmetric filters based on bq; (left) and of the Musgrave asymmetric
filters (right) corresponding to the 13-term symmetric filter

in months over the signal domain. However, as shown in Fig. 10.3, their time path
is only very close to that of the filters derived by Musgrave [19] up to q D 2 but
there is no monotonic convergence of these asymmetric filters to their final ones.
As already said, the Musgrave filters are based on the minimization of the mean
squared revision between the final estimates, obtained by the application of the
symmetric filter, and the preliminary estimates, obtained by the application of an
asymmetric filter, subject to the constraint that the sum of the weights is equal to
one [12, 18]. These filters have the good property of fast detection of turning points.
This property is reflected in their phase shift function that, for the last point filter,
is illustrated in Fig. 10.4. As can be seen, both the last point Musgrave filter and
10.3 Empirical Application 255

1.2

1.5
Sym b0,G b0, b0, Mus

b0,G

0.5
0.8

b0,

( )
G( )

b0,

−0.5
0.4

Mus

−2.0
0.0

0.0 0. 1 0. 2 0.3 0.4 0.5 0.0 0. 1 0.2 0.3 0.4 0.5

Fig. 10.4 Gain (left) and phase shift (right) functions for the last point asymmetric filters based
on b0; , b0;G , and b0; compared with the last point Musgrave filter

the one based on b0; produce almost one half of the phase shift introduced by the
filter based on b0; and a quarter of the one introduced by the filter based on b0;G at
the signal frequency band. However, the reduced phase shift produced by these two
filters is compensated by larger revisions introduced in the final estimates. Indeed,
as shown by the corresponding gain functions, the last point Musgrave filter and
the one based on b0; suppress much less noise than the filters obtained through
minimization of Eqs. (10.15) and (10.16). Furthermore, the Musgrave filter has the
worst performance since it introduces a large amplification of the power attributed
to the trend and suppresses less noise.

10.3 Empirical Application

The asymmetric filters previously derived can be applied in many fields, such
as, macroeconomic, finance, health, hydrology, meteorology, criminology, physics,
labor markets, and utilities. In fact, in any time series where the impact of the trend-
cycle is of relevance. A set of leading, coincident, and lagging indicators of the US
economy is chosen to corroborate the theoretical conclusions discussed before. The
leading indicators are time series that have a turning point before the economy as
a whole changes, whereas the coincident indicators change direction approximately
at the same time as the whole economy, thereby providing information about the
current state of the economy. On the other hand, the lagging indicators are those that
usually change direction after the economy as a whole does. The composite indexes
are typically reported in the financial and trade press, and the data analyzed in this
study are from the St. Louis Federal Reserve Bank database, the Bureau of Labor
256 10 Real Time Trend-Cycle Prediction

Statistics, and the National Bureau of Economic Research (NBER). In particular,

the analyzed leading indicators are
– Average weekly overtime hours, manufacturing
– New orders for durable goods
– New orders for nondefense capital goods
– New private housing units authorized by building permits
– Stock prices, S&P common stocks
– Money supply, M2
– Interest rate spread, 10-year treasury bonds less federal funds
– Index of consumer expectation (University of Michigan).
We also consider the following coincident indicators:
– Employees on nonagricultural payrolls
– Personal income less transfer payments
– Industrial production index
– Manufacturing and trade sales.
Finally, the lagging indicators treated are
– Average duration of unemployment, weeks
– Ratio, manufacturing, and trade inventory to sale
– Change in labor cost per unit of output, manufacturing
– Commercial and industrial loans outstanding.
The asymmetric filters derived following the RKHS methodology versus the
Musgrave filters, applied in conjunction with the symmetric Henderson filter, are
evaluated as follows.

10.3.1 Reduction of Revision Size in Real Time Trend-Cycle

Estimates

The reduction of revisions in real time trend-cycle estimates is very important

because the estimates are preliminary and used to assess the current stage of the
economy. Statistical agencies and major users of these indicators are reluctant to
large revisions because these can lead to erroneous statement concerning the current
economic situation. The series considered are all seasonally adjusted, where also
trading day variations, moving holidays, and extreme values have been removed, if
present. The socioeconomic indicators are series of different length but the periods
selected sufficiently cover the various lengths published for these series.
Here, we study how the filters derived in RKHS and the classical Musgrave
estimators respond to the variability of the data. For each series, the length of the
filters is selected according to the I=C (noise to signal) ratio, as classically done in
10.3 Empirical Application 257

Table 10.2 Ratio of the mean square percentage revision errors of the last point asymmetric filters
based on b0;G , b0; , and b0; , and the last point Musgrave filter
b0;G b0; b0;
Macro-area Series Mus Mus Mus
Leading Average weekly overtime hours: manufacturing 0.492 0.630 0.922
New orders for durable goods 0.493 0.633 0.931
New orders for nondefense capital goods 0.493 0.633 0.931
New private housing units authorized by building permits 0.475 0.651 0.927
S&P 500 stock price index 0.454 0.591 0.856
M2 money stock 0.508 0.655 0.932
10-year treasury constant maturity rate 0.446 0.582 0.849
University of Michigan: consumer sentiment 0.480 0.621 0.912
Coincident All employees: total nonfarm 0.517 0.666 0.951
Real personal income excluding current transfer receipts 0.484 0.627 0.903
Industrial production index 0.477 0.616 0.884
Manufacturing and trade sales 0.471 0.606 0.869
Lagging Average (mean) duration of unemployment 0.509 0.649 0.937
Inventory to sales ratio 0.483 0.618 0.894
Index of total labor cost per unit of output 0.515 0.663 0.983
Commercial and industrial loans at all commercial banks 0.473 0.610 0.871

the X11/X12ARIMA procedure [17]. In the sample, the ratio ranges from 0.20 to
1.98, hence filters of length 9 and 13 terms are applied.
The comparisons are based on the relative filter revisions between the final
symmetric filter S and the last point asymmetric filter A, that is,

S t  At
Rt D ; t D 1; 2; : : : ; n: (10.19)
St

For each series and for each estimator, we calculate the ratio between the mean
square percentage error (MSPE) of the revisions corresponding to the filters
derived following the RKHS methodology and those corresponding to the last
point Musgrave filter. For all the estimators, the results illustrated in Table 10.2
indicate that the ratio is always smaller than one, indicating that the kernel last
point predictors, based on time-varying bandwidth parameters, introduce smaller
revisions than the Musgrave filter. This implies that the estimates obtained by the
former will be more accurate than those derived by the application of the latter. In
particular, as expected, the best performance is shown by the filter based on the
optimal bandwidth b0;G obtained by minimizing the criterion (10.16). In almost all
the series its ratio with the last point Musgrave filter is less than one half and, on
average, around 0.480. This implies that when applied to real data, the filter based
on b0;G produces a reduction of almost 50 % of the revisions introduced in the real
time trend-cycle estimates given by the Musgrave filter. The filter based on b0; ,
derived to minimize the size of total filter revisions as defined by Eq. (10.15), also
258 10 Real Time Trend-Cycle Prediction

performs very well with more than the 30 % of revision reduction with respect to
the Musgrave filter. In this case, the ratio is greater than the one corresponding to
the filter based on b0;G , but always less than 0.7 for all the series, being, on average,
around 0.636. The filter whose bandwidth parameter is selected to minimize the
average phase shift over the signal domain performs more similar to the last point
Musgrave filter but still shows revisions reduction, on average, around 10 %.

10.3.2 Turning Point Detection

It is important that the reduction of revisions in real time trend-cycle estimates is not
achieved at the expense of increasing the time lag to detect the upcoming of a true
turning point. A turning point is generally defined to occur at time t if (downturn):

ytk      yt1 > yt  ytC1      ytCm ;

or (upturn)

ytk      yt1 < yt  ytC1      ytCm :

Following Zellner et al. [22], it is selected k D 3 and m D 1 given the smoothness

of the trend-cycle data. For each estimator, the time lag to detect the true turning
point is affected by the convergence path of its asymmetric filters wq ; q D 0; : : : ; m
1, to the symmetric one w.
To determine the time lag needed by an indicator to detect a true turning point
it is calculated the number of months it takes for the real time trend-cycle estimate
to signal a turning point in the same position as in the final trend-cycle series. For
the series analyzed in this chapter, the time delays for each estimator are shown
in Table 10.3. It can be noticed that the filters based on the bandwidth bq; take
almost 2 months (on average), similar to the Musgrave filters, to detect the turning
point. This is due to the fact that, even if bq; filters are designed to be optimal in
timeliness, their convergence path to the symmetric filter is slower and moreover
not monotonic.
On the other hand, the filters based on bq; ; q D 0; : : : ; m  1, and bq;G ; q D
0; : : : ; m  1, perform strongly better. In particular, whereas the former detect the
turning point with an average time delay of 1.67 months, the latter takes 1.27
months.
10.3 Empirical Application 259

Table 10.3 Time lag in detecting true turning points for the asymmetric filters based on bq;G , bq; ,
and bq; , and the Musgrave filters
Macro-area Series bq;G bq; bq; Musgrave
Leading Average weekly overtime hours: manufacturing 1 1 1 1
New orders for durable goods 1 2 3 2
New orders for nondefense capital goods 1 2 2 3
New private housing units authorized by building 2 2 3 3
permits
S&P 500 stock price index 1 2 2 2
10-Year treasury constant maturity rate 1 1 1 2
University of Michigan: consumer sentiment 1 1 1 1
Coincident All employees: total nonfarm 1 1 1 2
Real personal income excluding current transfer 1 1 1 1
receipts
Industrial production index 1 1 1 1
Manufacturing and trade sales 1 2 3 3
Lagging Average (mean) duration of unemployment 3 3 4 3
Inventory to sales ratio 1 1 1 2
Index of total labor cost per unit of output 2 2 3 2
Commercial and industrial loans at all commercial 1 1 1 1
banks
Average time lag in months 1.27 1.67 1.93 2.00

The fastest the upcoming of a turning point is detected, the fastest new policies
can be applied to counteract the impact of the business-cycle stage. Failure to
recognize the downturn in the cycle or taking a long time delay to detect it may lead
to the adoption of policies to curb expansion when in fact, a recession is already
underway.
The filters based on local bandwidth parameters selected to minimize crite-
rion (10.16) are optimal, since they drastically reduce the total revisions by one half
with respect to the Musgrave filters and, similarly, almost by one third the number
of months needed to pick up the June 2009 turning point. The real time trend-cycle
filter calculated with the bandwidth parameter that minimizes the distance between
the asymmetric and symmetric filters gain functions is to be preferred. This last point
trend-cycle filter reduced by one half the size of the revisions and by one third the
time delay to detect the June 2009 turning point relative to the Musgrave filters. For
illustrative purposes, Tables 10.4 and 10.5 give the weight systems of these filters
for 9- and 13-term symmetric filters.
 260

Table 10.4 Weights corresponding to the 9-term symmetric filter

0:31218 0:28804 0:22278 0:13330 0:044036 0:00000 0:00000 0:00000 0:00000
0:27101 0:31845 0:27101 0:15289 0:02630 0:03965 0:00000 0:00000 0:00000
0:10931 0:25652 0:32009 0:25652 0:10931 0:01705 0:03470 0:00000 0:00000
0:01544 0:11250 0:25957 0:32281 0:25957 0:11250 0:01544 0:03605 0:00000
0:03907 0:01074 0:12023 0:26574 0:32767 0:26574 0:12023 0:01074 0:03907
0:00000 0:03605 0:01544 0:11250 0:25957 0:32281 0:25957 0:11250 0:01544
0:00000 0:00000 0:03470 0:01705 0:10931 0:25652 0:32009 0:25652 0:10931
0:00000 0:00000 0:00000 0:03965 0:02630 0:15289 0:27101 0:31845 0:27101
0:00000 0:00000 0:00000 0:00000 0:04404 0:13330 0:22278 0:28804 0:31184
10 Real Time Trend-Cycle Prediction
Table 10.5 Weights corresponding to the 13-term symmetric filter
10.3 Empirical Application

0:22362 0:21564 0:19266 0:157478 0:11444 0:06902 0:02714 0:00000 0:00000 0:00000 0:00000 0:00000 0:00000
0:21065 0:22352 0:21065 0:174523 0:12230 0:06460 0:01357 0:01982 0:00000 0:00000 0:00000 0:00000 0:00000
0:15391 0:21013 0:23100 0:21013 0:15391 0:08000 0:01250 0:02600 0:025570 0:00000 0:00000 0:00000 0:00000
0:06338 0:14212 0:20452 0:22808 0:20452 0:14212 0:06338 0:00217 0:02978 0:01617 0:00000 0:00000 0:00000
0:00258 0:06319 0:14245 0:20533 0:22909 0:20533 0:14245 0:06319 0:00258 0:02996 0:01593 0:00000 0:00000
0:02983 0:00060 0:06762 0:14651 0:20848 0:23179 0:20848 0:14651 0:06762 0:00060 0:02983 0:01855 0:00000
0:01986 0:02982 0:00217 0:07010 0:14921 0:21106 0:23429 0:21106 0:149208 0:07010 0:00217 0:02982 0:01986
0:00000 0:01855 0:02983 0:00066 0:06762 0:14651 0:20848 0:23179 0:20848 0:14651 0:06762 0:00060 0:02983
0:00000 0:00000 0:01593 0:02996 0:00258 0:06319 0:14245 0:20533 0:22909 0:20533 0:14245 0:06319 0:00258
0:00000 0:00000 0:00000 0:01617 0:02978 0:00217 0:06338 0:14212 0:20452 0:22808 0:20452 0:14212 0:06338
0:00000 0:00000 0:00000 0:00000 0:02557 0:02600 0:01250 0:08000 0:15391 0:21013 0:23100 0:21013 0:15391
0:00000 0:00000 0:00000 0:00000 0:00000 0:01982 0:01357 0:06460 0:12230 0:17452 0:21065 0:22352 0:21065
0:00000 0:00000 0:00000 0:00000 0:00000 0:00000 0:02714 0:06902 0:11444 0:15748 0:19267 0:21564 0:22362
261
262 10 Real Time Trend-Cycle Prediction

References

1. Berlinet, A. (1993). Hierarchies of higher order kernels. Probability Theory and Related Fields,
94, 489–504.
2. Bianconcini, S., & Quenneville, B. (2010). Real time analysis based on reproducing kernel
Henderson filters. Estudios de Economia Aplicada, 28(3), 553–574.
3. Dagum, E. B. (1982). Revisions of time varying seasonal filters. Journal of Forecasting, 1,
173–187.
4. Dagum, E. B. (1982). The effects of asymmetric filters on seasonal factor revisions. Journal of
American Statistical Association, 77, 732–738.
5. Dagum, E. B. (1996). A new method to reduce unwanted ripples and revisions in trend-cycle
estimates from X11ARIMA. Survey Methodology, 22, 77–83.
6. Dagum, E. B., & Bianconcini, S. (2008). The Henderson smoother in reproducing kernel
Hilbert space. Journal of Business and Economic Statistics, 26(4), 536–545.
7. Dagum, E. B., & Bianconcini, S. (2013). A unified probabilistic view of nonparametric
predictors via reproducing kernel Hilbert spaces. Econometric Reviews, 32(7), 848–867.
8. Dagum, E. B., & Bianconcini, S. (2015). A new set of asymmetric filters for tracking the short-
term trend in real time. The Annals of Applied Statistics, 9, 1433–1458.
9. Dagum, E. B., & Laniel, N. (1987). Revisions of trend-cycle estimators of moving average
seasonal adjustment methods. Journal of Business and Economic Statistics, 5, 177–189.
10. Dagum, E. B., & Luati, A. (2009). A cascade linear filter to reduce revisions and turning points
for real time trend-cycle estimation. Econometric Reviews, 28(1–3), 40–59.
11. Dagum, E. B., & Luati, A. (2012). Asymmetric filters for trend-cycle estimation. In W. R.
Bell, S. H. Holan, & T. S. McElroy (Eds.), Economic time series: Modeling and seasonality
(pp. 213–230). Boca Raton, FL: Chapman&Hall.
12. Doherty, M. (2001). The surrogate Henderson filters in X-11. Australian and New Zealand
Journal of Statistics, 43(4), 385–392.
13. Gasser, T., & Muller, H. G. (1979). Kernel estimation of regression functions. In Lecture notes
in mathematics (Vol. 757, pp. 23–68). New York: Springer.
14. Gray, A. G., & Thomson, P. J. (2002). On a family of finite moving-average trend filters for the
ends of series. Journal of Forecasting, 21, 125–149.
15. Henderson, R. (1916). Note on graduation by adjusted average. Transaction of Actuarial
Society of America, 17, 43–48.
16. Kyung-Joon, C., & Schucany, W. R. (1998). Nonparametric Kernel regression estimation near
endpoints. Journal of Statistical Planning and Inference, 66(2), 289–304.
17. Ladiray, D., & Quenneville, B. (2001). Seasonal adjustment with the X-11 method. New York:
Springer-Verlag.
18. Laniel, N. (1985). Design Criteria for the 13-term Henderson End Weights. Working paper.
Ottawa: Methodology Branch. Statistics Canada.
19. Musgrave, J. (1964). A Set of End Weights to End All End Weights. Working paper.
Washington, DC: U.S. Bureau of Census.
20. Quenneville, B., Ladiray, D., & Lefrancois, B. (2003). A note on Musgrave asymmetrical trend-
cycle filters. International Journal of Forecasting, 19(4), 727–734.
21. Wildi, M. (2008). Real-time signal extraction: Beyond maximum likelihood principles. Berlin:
Springer.
22. Zellner, A., Hong, C., & Min, C. (1991). Forecasting turning points in international output
growth rates using Bayesian exponentially weighted autoregression, time-varying parameter,
and pooling techniques. Journal of Econometrics, 49(1–2), 275–304.
Chapter 11
The Effect of Seasonal Adjustment
on Real-Time Trend-Cycle Prediction

Abstract Asymmetric nonparametric trend-cycle filters obtained with local time-

varying bandwidth parameters via the Reproducing Kernel Hilbert Space (RKHS)
methodology reduce significantly revisions and turning point detection respect to
the currently used by statistical agencies. The best choice of local time-varying
bandwidth is the one obtained by minimizing the distance between the gain
functions of the RKHS asymmetric and the symmetric filter to which it must
converge. Since the input to these kernel filters is seasonally adjusted series, it is
important to evaluate the impact that the seasonal adjustment method can have. The
purpose of this chapter is to assess the effects of the seasonal adjustment methods
when the real time trend is predicted with such nonparametric kernel filters. The
seasonal adjustments compared are the two officially adopted by statistical agencies:
X12ARIMA and TRAMO-SEATS applied to a sample of US leading, coincident,
and lagging indicators.

The RKHS trend-cycle filters discussed in Chap. 10 are applied to seasonally

adjusted data, hence it is important to evaluate how their statistical properties are
affected by the seasonal adjustment method used to produce the input data. In this
regard, we analyze the behavior of the Dagum and Bianconcini [9] asymmetric
filters in terms of size of revisions in the estimates as new observations are added
to the series, and time lag to detect true turning points. We look at seasonally
adjusted data from X12ARIMA and TRAMO-SEATS which are officially adopted
by statistical agencies, with outliers replaced if present.
Seasonal adjustment means the removal of seasonal variations in the original
series jointly with trading day variations and moving holiday effects. The main
reason for seasonal adjustment is the need of standardizing socioeconomic series
because seasonality affects them with different timing and intensity. Hence, the
seasonally adjusted data reflect variations due only to the remaining components,
namely trend-cycle and irregulars. The information given by seasonally adjusted
series has always played a crucial role in the analysis of current economic conditions
and provides the basis for decision making and forecasting, being of major
importance around cyclical turning points.
As discussed in Chap. 3, seasonal adjustment methods can be classified as
deterministic or stochastic depending on the assumptions made concerning how

© Springer International Publishing Switzerland 2016 263

E. Bee Dagum, S. Bianconcini, Seasonal Adjustment Methods and Real Time
Trend-Cycle Estimation, Statistics for Social and Behavioral Sciences,
DOI 10.1007/978-3-319-31822-6_11
264 11 The Effect of Seasonal Adjustment on Real-Time Trend-Cycle Prediction

seasonality evolves through time. Deterministic methods assume that the seasonal
pattern can be predicted with no error or with variance of the prediction error null.
On the contrary, stochastic methods assume that seasonality can be represented
by a stochastic process, that is, a process governed by a probability law and,
consequently, the variance of the prediction error is not null. The best known
seasonal adjustment methods belong to the following types:
(i) regression methods which assume global or local simple functions of time,
(ii) stochastic model-based methods which assume simple ARIMA models, and
(iii) moving average methods which are based on linear filtering and hence do not
assume explicit parametric models.
Only methods (i), which assume global simple functions for each component, are
deterministic; the others are considered stochastic. Moving averages and ARIMA
model-based (AMB) methods are those mainly applied by statistical agencies to
produce officially seasonally adjusted series. A brief description is given in the
following and we refer the readers to Chaps. 4 and 5, where each method is
discussed in detail.

11.1 Seasonal Adjustment Methods

11.1.1 X12ARIMA

The X12ARIMA is today the most often applied seasonal adjustment method by
statistical agencies. It was introduced by Findley et al. [10] and is an enhanced
version of the X11ARIMA method developed at Statistics Canada by Estela Bee
Dagum [5].
The major modifications concern: (1) extending the automatic identification and
estimation of ARIMA models for the extrapolation option to many more than the
three models available in the X11ARIMA, and (2) estimating trading day variations,
moving holidays, and outliers in what is called regARIMA model. The latter
consists of regression models with ARIMA errors. More precisely, they are models
in which the mean function of the time series (or its logs) is described by a linear
combination of regressors, and the covariance structure of the series is that of an
ARIMA process. If no regressors are used, indicating that the mean is assumed to
be zero, the regARIMA model reduces to an ARIMA model.
Whether or not special problems requiring the use of regressors are present
in the series to be adjusted, an important use of regARIMA models is to extend
the series with forecasts (and backcasts) in order to improve the seasonal adjust-
ments of the most recent (and the earliest) data. Doing this reduces problems
inherent in the trend estimation and asymmetric seasonal averaging processes of
the type used by the X11 method near the ends of the series. The provision of
this extension was the most important improvement offered by the X11ARIMA
11.1 Seasonal Adjustment Methods 265

program. Its theoretical and empirical benefits have been documented in many
publications, such as Dagum [6], Bobbit and Otto [3], and the references therein.
The X12ARIMA method has all the seasonal adjustment capabilities of the
X11ARIMA variant. The same seasonal and trend moving averages are available,
and the program still offers the X11 calendar and holiday adjustment routines
incorporated in X11ARIMA. But several important new options have been included.
The modeling module is designed for regARIMA model building with sea-
sonal socioeconomic time series. To this end, several categories of predetermined
regression variables are available, including trend constants or overall means, fixed
seasonal effects, trading day effects, holiday effects, pulse effects (additive outliers),
level shifts, temporary change outliers, and ramp effects. User-specified regression
variables can also be included in the models. The specification of a regARIMA
model requires specification of both the regression variables to be included in the
model and the type of ARIMA model for the regression errors (i.e., the order
.p; d; q/.P; D; Q/s).
Specification of the regression variables depends on user knowledge about
the series being modeled. Identification of the ARIMA model for the regression
errors follows well-established procedures based on examination of various sample
autocorrelation and partial autocorrelation functions produced by the X12ARIMA
program.
Once a regARIMA model has been specified, X12ARIMA estimates its param-
eters by maximum likelihood using an Iterative Generalized Least Squares (IGLS)
algorithm. Diagnostic checking involves examination of residuals from the fitted
model for signs of model inadequacy. X12ARIMA produces several standard
residual diagnostics for model checking, as well as provides sophisticated methods
for detecting additive outliers and level shifts. Finally, X12ARIMA can produce
point forecasts, forecast standard errors, and prediction intervals from the fitted
regARIMA model.
Trading day effects occur when a series is affected by the different day-of-the-
week compositions of the same calendar month in different years. Trading day
effects can be modeled with seven variables that represent (no. of Mondays), . . . ,
(no. of Sundays) in month t. Bell and Hillmer [1] proposed a better parameterization
of the same effects using six variables defined as (no. of Mondays/  .no.
of Sundays), . . . , (no. of Saturdays/  .no. of Sundays), along with a seventh
variable for Length of Month (LOM) or its deseasonalized version, the leap-
year regressor (lpyear). In X12ARIMA the six variables are called the tdnolpyear
variables. Instead of using a seventh regressor, a simpler and often better way to
handle multiplicative leap-year effects is to re-scale the February values of the
original time series before transformation to m N Feb yt =mt , where yt is the original
time series before transformation, mt is the length of month t (28 or 29), and
mN Feb D 28:25 is the average length of February. If the regARIMA model includes
seasonal effects, these can account for the length of month effect except in
Februaries, so the trading day model only has to deal with the leap-year effect.
When this is done, only the tdnolpyear variables need be included in the model.
X12ARIMA allows explicit choice of either approach, as well as an option (td)
266 11 The Effect of Seasonal Adjustment on Real-Time Trend-Cycle Prediction

that makes a default choice of how to handle length of month effects. When
the time series being modeled represents the aggregation of some daily series
(typically unobserved) over calendar months they are called monthly flow series.
If the series instead represents the value of some daily series at the end of
the month, called a monthly stock series, then different regression variables are
appropriate.
Holiday effects in a monthly flow series arise from holidays whose dates vary
over time if: (1) the activity measured by the series regularly increases or decreases
around the date of the holiday, and (2) this affects two (or more) months depending
on the date the holiday occurs each year. Effects of holidays with a fixed date, such
as Christmas, are indistinguishable from fixed seasonal effects. Easter effects are
the most frequently found holiday effects in American and European economic
time series, since the date of Easter Sunday varies between March 22 and April
25.
X12ARIMA provides four other types of regression variables to deal with abrupt
changes in the level of a series of a temporary or permanent nature: additive outliers
(AO), level shifts (LS), temporary changes (TC), and ramps. Identifying the location
and nature of potential outliers is the object of the outlier detection methodology
implemented. This methodology can be used to detect AOs, TCs, and LSs (not
ramps); any outlier that is detected is automatically added to the model as regression
variable. Prespecified AOs, LSs, TCs, and ramps are actually simple forms of
interventions as discussed by Box and Tiao [4].
The regARIMA modeling routine fulfills the role of prior adjustment and forecast
extension of the time series. That is, it first estimates some deterministic effects,
such as calendar effects and outliers using predefined built-in options or user-
defined regressors, and it removes them from the observed series. An appropriate
ARIMA model is identified for the pre-adjusted series in order to extend it
with forecasts. Then, this extended pre-adjusted series is decomposed into the
unobserved components of the series using moving averages, also accounting for
the presence of extreme values, as follows:
1. a preliminary estimate of the trend-cycle is obtained using a centered thirteen
(five) term weighted moving average of the monthly (quarterly) original series;
2. this trend-cycle series is removed from the original one to give an initial estimate
of the seasonal and irregular components, often called SI ratios;
3. initial preliminary seasonal factors are estimated from these initial SI ratios by
applying weighted seasonal moving averages for each month over several years,
separately;
4. an initial estimate of the irregular component is obtained by removing the initial
preliminary seasonal factors from the initial seasonal irregular (SI) ratios;
5. extreme values are identified and (temporarily) adjusted to replace the initial SI
ratios with the modified ones;
6. steps 3. and 4. are repeated on the modified SI ratios, initial seasonal factors are
derived by normalizing the initial preliminary seasonal factors;
11.1 Seasonal Adjustment Methods 267

7. an initial estimate of the seasonally adjusted series is obtained by removing the

initial seasonal factors from the original series;
8. a revised trend-cycle estimate is obtained by applying Henderson moving
averages to the initial seasonally adjusted series;
9. steps 2.–8. are again repeated twice to produce the final estimates of the trend-
cycle, seasonal, and irregular components.
Various diagnostics and quality control statistics are computed, tabulated, and
graphed in order to assess the seasonal adjustment results. If they are acceptable
based upon the diagnostics and quality measures, then the process is terminated.
If this is not the case, the above steps have to be repeated to search for a more
satisfactory seasonal adjustment of the series. The procedure is summarized in the
LHS of Fig. 11.1.

Fig. 11.1 X12ARIMA (left) and TRAMO-SEATS (right) procedures

268 11 The Effect of Seasonal Adjustment on Real-Time Trend-Cycle Prediction

11.1.2 TRAMO-SEATS

Gomez and Maravall [12] developed at the Bank of Spain, a seasonal adjustment
software called TRAMO-SEATS which is currently applied mainly by European
statistical agencies. TRAMO stands for “Time series Regression with ARIMA
noise, Missing observations and Outliers,” and SEATS for “Signal Extraction in
ARIMA Time Series.” First, TRAMO estimates via regression the deterministic
components, which are after removed from the input data. In a second round, SEATS
estimates the seasonal and trend-cycle components from the ARIMA model fitted
to the data where the deterministic components are removed. SEATS uses the filters
derived from the linearized ARIMA model that describes the stochastic behavior of
the time series.
It should be mentioned that Eurostat, in collaboration with the National Bank
of Belgium, the US Bureau of the Census, the Bank of Spain, and the European
Central Bank, has developed an interface of TRAMO-SEATS and X12ARIMA
called DemetraC. In the Bank of Spain and Eurostat websites it is also possible
to find a considerable number of papers relevant to TRAMO-SEATS as well as in
the European Statistical System (ESS) Guideline.
TRAMO is a regression method that performs the estimation, forecasting, and
interpolation of missing observations with ARIMA errors, in the presence of possi-
bly several types of outliers. The ARIMA model can be identified automatically or
by the user. It consists of the following steps:
1. automatic detection of several types of outliers and, if necessary, estimation of
other regression variables, such as calendar effects;
2. automatic identification of an ARIMA model;
3. then, TRAMO removes all the estimated components (trading days, moving
holidays, and outliers) from the original series and passes the linearized series
to SEATS, where the actual decomposition is performed.
SEATS belongs to the class of procedures based on ARIMA models for the
decomposition of time series into unobserved components and consists of the
following steps:
1. ARIMA model estimation. SEATS starts by fitting an ARIMA model to a series
not affected by deterministic components, such as trading day variations, moving
holidays, and outliers. Let yt denote this linearized series, and consider an
additive decomposition model (multiplicative if applied to the log transformation
of yt ), such that

zt D .1  B/d .1  Bs /D yt (11.1)

represent the “differenced” series. The corresponding model for the differenced
linearized series zt can be written as

p .B/˚P .Bs /.zt  zN/ D q .B/Q .Bs /at ; (11.2)

11.1 Seasonal Adjustment Methods 269

where zN is the mean of zt , at is a series of innovations, normally distributed with

zero mean and variance a2 , p .B/˚P .Bs /, and q .B/Q .Bs / are autoregressive
and moving average polynomials in B, respectively, which are expressed in
multiplicative form as the product of a regular polynomial in B and a seasonal
polynomial in Bs . The complete model can be written in detailed form as

p .B/˚P .Bs /.1  B/d .1  Bs /D yt D q .B/Q .Bs /at C c; (11.3)

and, in concise form, as

 .B/yt D .B/at C c; (11.4)

where c is equal to  .B/Ny, being yN the mean of the linearized series yt . In words,
the model that SEATS assumes is a linear time series with Gaussian innovations.
When used with TRAMO, estimation of the ARIMA model is made by the exact
maximum likelihood method described in Gomez and Maravall [11].
2. Derivation of the ARIMA models for each component. The program proceeds
by decomposing the series that follows the ARIMA model (11.4) into several
components. The decomposition can be multiplicative or additive. Next we shall
discuss the additive model, since the multiplicative relation can be taken care
with the log transformation of the data. That is,

yt D Tt C Ct C St C It ; (11.5)

where Tt denotes the trend component, Ct the cycle, St represents the seasonal
component, and It the irregulars. The decomposition is done in the frequency
domain. The spectrum (or pseudospectrum) of yt is partitioned into additive
spectra, associated with the different components which are determined, mostly,
from the AR roots of the model. The trend component represents the long-term
evolution of the series and displays a spectral peak at frequency 0, whereas the
seasonal component captures the spectral peaks at seasonal frequencies (e.g.,
for monthly data these are 0.524, 1.047, 1.571, 2.094, 2.618, and 3.142). The
cyclical component captures periodic fluctuations with period longer than a year,
associated with a spectral peak for a frequency between 0 and (2=s), and short-
term variation associated with low order MA components and AR roots with
small moduli. Finally, the irregular component captures white noise behavior, and
hence has a flat spectrum. The components are determined and fully derived from
the structure of the (aggregate) ARIMA model (11.4) for the linearized series
directly identified from the data. The program is aimed at monthly or quarterly
frequency data and the maximum number of observations that can be processed is
600. One important assumption is that of orthogonality among the components,
and each one will have in turn an ARIMA model. In order to identify the
components, the canonical decomposition is used which implies that the variance
of the irregulars is maximized, whereas the trend, seasonal, and cycle are as
stable as possible (compatible with the stochastic nature of model (11.2)). The
270 11 The Effect of Seasonal Adjustment on Real-Time Trend-Cycle Prediction

canonical condition on the trend, seasonal, and cyclical components identifies

a unique decomposition, from which the ARIMA models for the components
are obtained (including the component innovation variances). The trend-cycle
and seasonal components are estimated based on minimum mean squared error
(MMSE) criterion and using Wiener–Kolmogorov (WK) filters; the detected
outliers and some special effects are finally reintroduced into the components.
As in the case of X12ARIMA, diagnostic checking and analyses of the
decomposition accuracy and adequacy are performed using graphical, descriptive,
nonparametric, and parametric measures included in the output of the program. The
procedure is summarized in the RHS of Fig. 11.1.

11.2 Trend-Cycle Prediction in Reproducing Kernel Hilbert

Space (RKHS)

Due to major economic and financial changes of global nature, seasonally adjusted
data are not smooth enough to be able to provide a clear signal of the short-term
trend. Hence, further smoothing is required, but the main objections for trend-cycle
estimation are
1. the size of the revisions of the most recent values (generally much larger than for
the corresponding seasonally adjusted estimates), and
2. the time lag in detecting true turning points.
With the aim to overcome such limitations, Dagum and Bianconcini [7–9]
have recently developed nonparametric trend-cycle filters that reduce significantly
revisions and turning point detection with respect to the Musgrave [13] filters which
are applied by X12ARIMA. These authors rely on the assumption that the input
series fyt ; t D 1; 2; : : : ; ng is seasonally adjusted (where trading day variations and
outliers, if present, have been also removed), such that it can be decomposed into
the sum of a systematic component (signal) TCt , that represents the trend and cycle
usually estimated jointly, plus an irregular component It , called the noise, as follows:

yt D TCt C It ; t D 1; : : : ; n: (11.6)

The noise It is assumed to be either white noise, WN.0; I2 / or, more generally, a
stationary and invertible ARMA process.
On the other hand, the signal TCt ; t D 1; : : : ; n; is assumed to be a smooth
function of time that can be estimated as a weighted moving average as follows:

X
m
ct D
TC wj ytCj ; t D m C 1; : : : ; n  m; (11.7)
jDm
11.2 Trend-Cycle Prediction in Reproducing Kernel Hilbert Space (RKHS) 271

where fwj ; j D m; : : : ; mg are the weights applied to the (2m C 1/ observations
surrounding the target observation yt . Using the reproducing kernel Hilbert space
methodology [2], the weights can be derived from the following kernel function:

3
X
K4 .s/ D Pi .s/Pi .0/f0 .s/; s 2 Œ1; 1 ; (11.8)
iD0

where f0 .s/ D 15 16
.1  s2 /2 ; s 2 Œ1; 1 , is the biweight density function, and Pi ; i D
0; 1; 2; 3; are the corresponding orthonormal polynomials. In particular, the generic
weight wj ; j D m; : : : ; m, is given by
2  2 3
4  2 j  
6 71
b j
wj D 4 5 f0 ; (11.9)
S0 4  S2 2 b b

R1 P  r  
1 j
where r D 1 sr f0 .s/ds are the moments of f0 , and Sr D m jDm b b f0 bj
their discrete approximation that depends on the bandwidth parameter b whose
choice is of fundamental importance to guarantee specific inferential properties
to the trend-cycle estimator. It has to be selected to ensure that only (2m C 1)
observations surrounding the target point will receive nonzero weight and to
approximate, as close as possible, the continuous density function f0 , as well as
the moments r with Sr . Dagum and Bianconcini ([7] and [8]) have suggested a
time invariant global bandwidth b equal to m C 1, which gave excellent results.
At the end of the sample period, that is, t D n  .m C 1/; : : : ; n, only
2m; : : : ; m C 1 observations are available, and time-varying asymmetric filters have
to be considered. At the boundary, the effective domain of the kernel function K4 is
Œ1; q , where q << 1, instead of Œ1; 1 as for any interior point. This implies
that the symmetry of the kernel is lost, and it does not integrate to unity on the
R q
asymmetric support ( 1 K4 .s/ds ¤ 1). Based on these considerations, Dagum
and Bianconcini [8] have suggested to derive the asymmetric weights by “cutting
and normalizing” the symmetric kernel K4 , that means by omitting that part of the
function lying between q and 1, and by normalizing it on [1,q]. Hence, the
corresponding asymmetric weights result
2  2 3
4  2 j  
6 bq 7 1 j
wq;j D4 q q 5 f0 ; (11.10)
S0 4  S2 2 bq bq

j D m; : : : ; qI q D 0; : : : ; m  1:
272 11 The Effect of Seasonal Adjustment on Real-Time Trend-Cycle Prediction

Pq  r  
1
where Srq D jDm bq bq
j
f0 bjq is the discrete approximation of q r D
R q r
1 s f0 .s/ds that are proportional to the moments of the truncated biweight density
f0 on the support Œ1; q .
Dagum and Bianconcini [9] have analyzed in detail the statistical properties
of these asymmetric filters as function of the bandwidth parameters bq ; q D
0; : : : ; m  1. They have shown, both theoretically and empirically, that filters
based on local time-varying bandwidth parameters bq ; q D 0; : : : ; m  1, selected
to minimize the distance between the gain functions of each asymmetric filter
fwq;j ; j D 0; : : : ; q; q D 0; : : : ; mg and the symmetric one fwj ; j D m; : : : ; mg have
excellent properties. In particular, the last point trend-cycle filter reduces around
a half the size of the total revisions as well as the time delay to detect a true
turning point with respect to the Musgrave [13] filter. For a detailed discussion of
the properties of these filters and of the corresponding weight systems, we refer the
reader to Chap. 10.

11.3 Empirical Application

Since the RKHS trend-cycle filters discussed in the previous section are applied to
seasonally adjusted data, it is important to evaluate how their statistical properties
are affected by the seasonal adjustment method used to produce the input data.
In this regard, in this section, it is analyzed the behavior of the Dagum and
Bianconcini [9] asymmetric filters in terms of size of revisions in the estimates as
new observations are added to the series, and time lag to detect true turning points.
We looked at seasonally adjusted data with outliers replaced if present.
The filters are applied to a sample of series that cover various sectors, such as
labor, imports, exports, housing, industrial production, and inventories. The series
have been observed on different time periods. Specifically, we chose a set of leading,
coincident, and lagging indicators of the US economy. The leading indicators are
time series that have a turning point before the economy as a whole changes,
whereas the coincident indicators change direction approximately at the same time
as the whole economy, thereby providing information about the current state of the
economy. On the other hand, the lagging indicators are those that usually change
direction after the economy as a whole does. The composite indexes are typically
reported in the financial and trade press, and the data analyzed in this study are from
the St. Louis Federal Reserve Bank database, the Bureau of Labor Statistics, and the
National Bureau of Economic Research (NBER). In particular, the analyzed leading
indicators are
– Average weekly overtime hours, manufacturing
– New orders for durable goods
– New orders for nondefense capital goods
– New private housing units authorized by building permits
11.3 Empirical Application 273

– Stock prices, S&P common stocks

– Money supply, M2
– Interest rate spread, 10-year treasury bonds less federal funds
– Index of consumer expectation (University of Michigan).
We also consider the following coincident indicators:
– Employees on nonagricultural payrolls
– Personal income less transfer payments
– Industrial production index
– Manufacturing and trade sales.
Finally, the lagging indicators treated are
– Average duration of unemployment, weeks
– Ratio, manufacturing, and trade inventory to sale
– Change in labor cost per unit of output, manufacturing
– Commercial and industrial loans outstanding.
As an example, Figs. 11.2 and 11.3 illustrate the original and seasonally adjusted
(SA) monthly series of US Unemployment Rate for Males (16 years and over) for
the period January 1992–December 2013. This series is provided monthly by the
Bureau of Labor Statistics of the US Department of Labor. The unemployment rate
is a measure of unemployment and it is calculated as a percentage by dividing the

Original series
12

X12ARIMA SA series
10
8
6
4

1995 2000 2005 2010

Fig. 11.2 Original and X12ARIMA (default option) seasonally adjusted US Unemployment Rate
for Males (16 years and over) from January 1992 to December 2013
274 11 The Effect of Seasonal Adjustment on Real-Time Trend-Cycle Prediction

12
Original series
TRAMO-SEATS SA series
10
8
6
4

1995 2000 2005 2010

Fig. 11.3 Original and TRAMO-SEATS (default option) seasonally adjusted US Unemployment
Rate for Males (16 years and over) from January 1992 to December 2013

number of unemployed individuals by all individuals currently in the labor force,

employed plus unemployed. The unemployment rate for males is a series of the
Index of Coincident Economic Indicators, produced by the Conference Board to
provide an indication of the future direction of the US Economy. Indeed, if the
economy is just beginning to grow out of a recession, businesses will tend to hold
off on hiring new workers until they are more confident that economic growth is
improving. As the economy continues to improve, eventually businesses will be
forced to add more workers. This increase in employment will reinforce the positive
trend in economic growth and the unemployment rate will decrease. In contrast,
if the economy is just beginning to slow down, employers wishing to maintain
employee loyalty will try to keep their workers by reducing hours worked, rather
than immediately laying off workers. If the slowdown deepens into a recession,
then eventually businesses are forced to lay off workers, which reinforces the
negative trend in economic growth and the unemployment rate will increase. Thus
unemployment rate is a coincident indicator changing direction approximately at
the same time as the whole economy, thereby providing information about the
current state of the economy.
The seasonal adjustment is done using the default options of the officially
adopted X12ARIMA and TRAMO-SEATS. The outliers in the seasonally adjusted
data are replaced, if present. Figures 11.2 and 11.3 exhibit both original and
seasonally adjusted data. It is apparent that both seasonally adjusted series are very
close to one another. The total male unemployment rate shows a peak at the middle
of 2009, and underwent thenceforth a fast decline in the subsequent years.
11.3 Empirical Application 275

11.3.1 Reduction of Revisions in Real Time Trend-Cycle

Estimates

The reduction of revisions in real time trend-cycle estimates is very important

because the estimates are preliminary and used to assess the current stage of the
economy. For the seasonally adjusted series by X12ARIMA, the length of the filters
is selected according to the I=C (noise to signal) ratio, as classically done in the
X12ARIMA software. The I=C ranges from 0.19 to 3.22, hence symmetric filters
of 9- and 13-term and corresponding asymmetric filters have been applied. The
comparisons are based on the relative filter revisions between the final symmetric
filter S and the last point asymmetric filter A, that is,

S t  At
Rt D ; t D 1; 2; : : : ; n: (11.11)
St

For
q each series, it is calculated the root mean square percentage error (RMSPE),
100 PN 2
N tD1 Rt , and the mean absolute percentage error (MAPE) of the revisions,
100 P N
N tD1 jRt j, corresponding to the filters derived following the RKHS methodol-
ogy. Table 11.1 gives the RMSPE and MAPE of the revisions introduced by the
RKHS filters when applied to the SA data produced by X12ARIMA and TRAMO-
SEATS. The size of the revisions is small and very close to each other. An indication
is that the seasonal adjustment method chosen has no impact on the final results from
the revisions viewpoint.

11.3.2 Turning Point Detection

It is important that the reduction of revisions in real time trend-cycle estimates is

not achieved at the expense of increasing the time lag to detect true turning points.
A turning point is generally defined to occur at time t if (downturn):

ytk      yt1 > yt  ytC1      ytCm ;

or (upturn)

ytk      yt1 < yt  ytC1      ytCm :

Following Zellner et al. [14], k D 3 and m D 1 are chosen given the smoothness
of the trend-cycle data. For the values given by the RKHS method, the time lag to
detect the true turning point is affected by the convergence path of its asymmetric
filters fwq;j ; j D m; : : : ; qI q D 0; : : : ; m  1g to the symmetric one fwj ; j D
m; : : : ; mg.
 276

Table 11.1 Values of RMSPE and MAPE of the revisions induced by RKHS filters applied on different SA series
RMSPE MAPE
Macro-area Series X12ARIMA TRAMO-SEATS X12ARIMA TRAMO-SEATS
Leading Average weekly overtime hours: manufacturing 0.489 0.501 3.135 3.160
New orders for durable goods 0.438 0.450 3.075 3.210
New orders for nondefense capital goods 0.497 0.508 3.844 3.979
New private housing units authorized by building permits 0.577 0.586 4.160 4.294
S&P 500 stock price index 0.936 0.945 6.169 6.674
10-Year treasury constant maturity rate 0.986 1.046 7.685 8.055
University of Michigan: consumer sentiment 0.544 0.583 4.297 4.431
Coincident All employees: total nonfarm 0.049 0.049 0.419 0.420
Unemployment rate for males (16 years and over) 0.604 0.635 2.241 2.420
Real personal income excluding current transfer receipts 0.340 0.356 2.866 3.000
Industrial production index 0.128 0.125 0.962 0.959
Manufacturing and trade sales 0.319 0.281 2.497 2.354
Lagging Average (mean) duration of unemployment 0.364 0.390 2.778 3.021
Index of total labor cost per unit of output 0.150 0.183 1.202 1.336
Commercial and industrial loans at all commercial banks 0.242 0.240 2.127 2.102
Mean 0.395 0.406 2.858 3.019
11 The Effect of Seasonal Adjustment on Real-Time Trend-Cycle Prediction
References 277

Table 11.2 Ratio in the time delay in identifying true turning points for RKHS filters applied on
seasonally adjusted series
Time lag
Macro-area Series X12ARIMA TRAMO-SEATS
Leading Average weekly overtime hours: manufacturing 2.60 3.00
New orders for durable goods 2.33 2.50
New orders for nondefense capital goods 2.33 2.43
New private housing units authorized 2.67 3.00
by building permits
S&P 500 stock price index 2.50 2.75
10-Year treasury constant maturity rate 1.00 2.00
University of Michigan: consumer sentiment 1.00 1.00
Coincident All employees: total nonfarm 1.00 1.00
Unemployment rate for males (16 years and over) 1.61 1.94
Industrial production index 1.80 2.33
Manufacturing and trade sales 1.75 2.33
Lagging Average (mean) duration of unemployment 1.00 1.00
Index of total labor cost per unit of output 2.67 3.29
Commercial and industrial loans 1.20 1.50
at all commercial banks
Mean 1.759 2.177

To determine the time lag to detect a true turning point, we calculate the number
of months it takes for the real time trend-cycle estimate to signal a turning point in
the same position as in the final trend-cycle series. We have used a very large sample
of series and, but we illustrate here only the results for the leading, coincident, and
lagging indicators shown in Table 11.2. For the three sets in total 40 turning points
were detected and the average delay in month is equal to 1.759 when the input is
seasonally adjusted with X12ARIMA and 2.177 when done with TRAMO-SEATS.
This latter takes longer to identify a true turning point. In fact, two series of the
leading indicator set and one of the lagging indicator set show time delays of 3 and
3.29 months.
We can conclude that the seasonally adjusted series with X12ARIMA is to be
preferred relative to that from TRAMO-SEATS concerning the time delay to detect
a true turning point.

References

1. Bell, W. H., & Hillmer, S. C. (1984). Issues involved with the seasonal adjustment of economic
time series. Journal of Business and Economic Statistics, 2, 291–320.
2. Berlinet, A. (1993). Hierarchies of higher order kernels. Probability Theory and Related Fields,
94, 489–504.
278 11 The Effect of Seasonal Adjustment on Real-Time Trend-Cycle Prediction

3. Bobbit, L., & Otto, M. C. (1990). Effects of forecasts on the revisions of seasonally
adjusted values using the X11 seasonal adjustment procedure. In Proceedings of the American
Statistical Association Business and Economic Statistics Session (pp. 449–453).
4. Box, G. E. P., & Tiao, G. C. (1975). Intervention analysis with applications to economics and
environmental problems. Journal of the American Statistical Association, 70, 70–79.
5. Dagum, E. B. (1978). Modelling, forecasting and seasonally adjusting economic time series
with the X-11 ARIMA method. The Statistician, 27(3), 203–216.
6. Dagum, E. B. (1988). The X-11 ARIMA/88 seasonal adjustment method-foundations and user’s
manual. Ottawa, Canada: Time Series Research and Analysis Centre, Statistics Canada.
7. Dagum, E. B., & Bianconcini, S. (2008). The Henderson smoother in reproducing kernel
Hilbert space. Journal of Business and Economic Statistics, 26(4), 536–545.
8. Dagum, E. B., & Bianconcini, S. (2013). A unified probabilistic view of nonparametric
predictors via reproducing kernel Hilbert spaces. Econometric Reviews, 32(7), 848–867.
9. Dagum, E. B., & Bianconcini, S. (2015). A new set of asymmetric filters for tracking the short-
term trend in real time. The Annals of Applied Statistics, 9, 1433–1458.
10. Findley, D. F., Monsell, B. C., Bell, W. R., Otto, M. C., & Chen, B. C. (1998). New capabilities
and methods of the X12ARIMA seasonal adjustment program. Journal of Business and
Economic Statistics, 16(2), 127–152.
11. Gomez, V., & Maravall, A. (1994). Estimation, prediction and interpolation for nonstationary
series with the Kalman filter. Journal of the American Statistical Association, 89, 611–624.
12. Gomez, V., & Maravall, A. (1996). Program TRAMO and SEATS: Instructions for users.
Working Paper 9628. Service de Estudios, Banco de Espana.
13. Musgrave, J. (1964). A set of end weights to end all end weights. Working Paper. U.S. Bureau
of Census, Washington, DC.
14. Zellner, A., Hong, C., & Min, C. (1991). Forecasting turning points in international output
growth rates using Bayesian exponentially weighted autoregression, time-varying parameter,
and pooling techniques. Journal of Econometrics, 49(1–2), 275–304.
Glossary

ARIMA Model A model that belongs to the ARIMA (Autoregressive integrated

moving average) process which is characterized by being homogeneous station-
ary after a finite difference operator has being applied.
Backshift Operator It is generally denoted by B or L (as lag operator) and when
applied to an element of a time series will produce the previous element, that is,
Byt D yt1 .
Business Cycle It is a quasi-periodic oscillation characterized by period of expan-
sion and contraction of the economy, lasting on average from 3 to 5 years.
Census II-X11 Variant (X11) Method Seasonal adjustment software based on
linear filters or moving averages applied in an iterative manner.
Deterministic Process It is a system in which no randomness is involved in the
development of future states. A deterministic model will thus always produce
the same output from a given starting condition or initial state.
Difference Operator The first difference operator .1  B/ is a special case of lag
polynomial, such that when applied to an element of a time series it produces the
difference between the element and the previous one, .1  B/yt D yt  yt1 .
Extreme Value The smallest (minimum) or largest (maximum) value of a function,
either in an arbitrarily small neighborhood of a point in the function domain or
on a given set contained in the domain. In a time series, it identifies a value the
probability of which falls in either tail of the distribution. It follows the same
probability distribution of all the other elements of the time series.
Flow Series A series where the observations are measured via the accumulation of
daily activities. Specifically, a flow variable is measured over an interval of time,
for example, a month or a quarter.
Gain Function A function that relates the spectrum of the original series to the
spectrum of the output obtained with a linear time invariant filter. It measures the
amplitude of the output for a sinusoidal input of unit amplitude.

© Springer International Publishing Switzerland 2016 279

E. Bee Dagum, S. Bianconcini, Seasonal Adjustment Methods and Real Time
Trend-Cycle Estimation, Statistics for Social and Behavioral Sciences,
DOI 10.1007/978-3-319-31822-6
280 Glossary

Irregular Component It represents variations related to unpredictable events of

all kinds. Most irregular values have a stable pattern in the sense of a constant
mean and finite variance but some extreme values may be present. The irregulars
are due to unforeseeable events of all kind and often arise from sampling error,
nonsampling error, unseasonable weather, natural disasters, strikes, etc.
Kalman Filter It is an algorithm that uses a series of measurements observed
over time, containing noise and other inaccuracies, and produces estimates of
unknown variables that tend to be more precise than those based on a single
measurement alone. The algorithm works in a two-step process. In the prediction
step, the Kalman filter produces estimates of the current state variables, along
with their uncertainties. Once the outcome of the next measurement (necessarily
corrupted with some amount of error, including random noise) is observed, these
estimates are updated using a weighted average, with more weight being given
to estimates with higher certainty. The algorithm is recursive. It can run in real
time, using only the present input measurements and the previously calculated
state and its uncertainty matrix; no additional past information is required.
Linear Filter see Moving average
Moving Average It consists of weighted or unweighted averages usually with
a fixed number of terms. It is called moving because the average is applied
sequentially by adding and subtracting one term at a time.
Moving Holiday Effects The moving holiday or moving festival component is
attributed to calendar variations, namely due to the fact that some holidays
change date in successive years. Example of moving holidays are Easter that
moves between months of March or April, the Chinese New Year date that
depends on the lunar calendar, and Ramadan that falls 11 days earlier from year
to year.
Moving Seasonality Seasonality characterized by changes in the seasonal ampli-
tude and/or phase.
Noise to Signal Ratio The ratio of the average absolute month-to-month (quarter-
to-quarter) change in the noise to that in the signal.
Outlier An observation that is well outside the expected range of values in a study
or experiment. They are similar to extreme values but the value treated as outlier
is characterized by a different distribution function from the rest of the data.
Outliers can often be traced to identifiable causes, for example, strikes, droughts,
floods, and data processing errors. Some outliers are the result of displacement
of activity from one month to the other.
Phase Shift Function that shows the shift in phase of the output compared with the
input.
Recession and Recovery Analysis Technique to assess the real time trend-cycle of
major socioeconomic indicators using percentage changes based on seasonally
adjusted units calculated for months and quarters in chronological sequence.
RegARIMA Models Models that combine a regression model for deterministic
variables such as moving holidays, trading day variations, and/or outliers with a
seasonal ARIMA model for the error term.
Glossary 281

Reproducing Kernel Hilbert Space It is a Hilbert space characterized by a kernel

function that reproduces, via inner product, every function of the space or,
equivalently, a Hilbert space of real-valued functions with the property that every
point evaluation functional is bounded and linear.
Seasonal Adjustment The process of identifying, estimating, and removing the
seasonal variations (and other calendar-related effects, such as trading day and
moving holidays) from a time series. The basic goal of seasonal adjustment is to
standardize socioeconomic time series to assess the stage of the cycle at which
the economy stands.
Seasonal Effects Variations due to seasonal causes that are reasonably stable
in terms of timing, direction, and magnitude from year to year. Seasonality
originates from climate seasons and conventional events of religious, social, and
civic nature, which repeat regularly from year to year.
SEATS It stands for Signal Extraction in ARIMA Time Series and is a seasonal
adjustment software that uses ARIMA models to estimate the trend-cycle and
seasonality of a time series.
Second Order (or Weak) Stationary Process Stochastic process whose first two
order moments are not time dependent, that is, the mean and the variance are
constant and the autocovariance function depends only on the time lag and not
on the time origin.
Spectrum of a Time Series It is the distribution of variance of the series as a
function of frequencies. The object of spectral analysis is to estimate and study
the spectrum. The spectrum contains no new information beyond that in the
autocovariance function (ACVF), and in fact the spectrum can be computed
mathematically by transformation of the ACVF. The ACVF summarizes the
information in the time domain and the spectrum in the frequency domain.
Stable Seasonality Seasonality that can be represented by a strictly periodic
function of time.
State Space A state space representation is a mathematical model of a system
as a set of input, output, and state variables related by first order differential
equations. State space refers to the space whose axes are the state variables. The
state of the system can be represented as a vector within that space. To abstract
from the number of inputs, outputs, and states, these variables are expressed
as vectors. Additionally, if the dynamical system is linear, time invariant and
finite-dimensional, then the differential and algebraic equations may be written
in matrix form.
Stochastic Process In probability theory, a stochastic process, or often random
process, is a collection of random variables, representing the evolution of some
system of random values over time. This is the probabilistic counterpart to a
deterministic process (or deterministic system). It is a set of random variables
indexed in time.
Stock Series A stock variable is measured at one specific time, and represents a
quantity existing at that point in time (e.g., December 31, 2015).
282 Glossary

Structural Time Series Model Model that consists of a number of linear stochas-
tic processes that stand for the trend, cycle, seasonality, and remaining stationary
dynamic features in an observed times series.
Time Series Set of observations on a given phenomenon that are ordered in time.
Formally, it is defined as a sample realization of a stochastic process, that is, a
process that evolves according to a probability law.
Trading Day Effects Variations associated with the composition of the calendar
which are due to the fact that the activities of some days of the week are
more important than others. Trading day variations imply the existence of a
daily pattern analogous to the seasonal pattern. However, these daily factors are
usually referred to as daily coefficients. This occurs because only non-leap year
Februaries have four of each day—four Mondays, four Tuesdays, etc. All other
months have an excess of some types of days. If an activity is higher on some
days compared to others, then the series can have a trading day effect.
TRAMO-SEATS It is a seasonal adjustment software composed of two main
parts: (1) TRAMO that automatically selects a regARIMA model to estimate
deterministic components such as trading day variations, moving holidays, and
outliers that are removed from the time series and (2) SEATS that uses the
time series where the deterministic components have been removed in order to
estimate the trend-cycle and seasonality and produces a seasonally adjust series.
Transfer Function It measures the effect of a linear filter on the total variance
of the input at different frequencies, and it is generally expressed in polar
coordinates.
Trend The concept of trend is used in economics and other sciences to represent
long-term smooth variations. The causes of these variations are often associated
with structural phenomena such as population growth, technological progress,
capital accumulation, new practices of business, and economic organization. For
most economic time series, the trends evolve smoothly and gradually, whether in
a deterministic or stochastic manner.
Trend-Cycle The trend-cycle is a short-term trend that includes the impact of the
long-term trend plus that of the business cycle. This is due to the fact that most
series are relatively short and then it is impossible to estimate the long-term
trend by itself. It is important to assess the current stage of the business cycle,
particularly to forecast the coming of a turning point.
Unwanted Ripple A 10-month cycle identified by the presence of high power
at the frequency 0.10 which, due to its periodicity, often leads to the wrong
identification of a true turning point.
White Noise Process Stochastic process characterized by random variables that
have a constant expected value (usually zero), a finite variance, and are mutually
uncorrelated. If the variance is assumed constant (homoscedastic condition), the
process is referred as white noise in the strict sense. If the variance is finite but
not constant (heteroscedastic condition), the process is called white noise in the
weak sense.
X11ARIMA The X11ARIMA is an enhanced version of the Census Method II–
X11 variant that was developed mainly to produce a better current seasonal
Glossary 283

adjustment of series with seasonality that changes rapidly in a stochastic manner.

The latter characteristic is often found in main socioeconomic indicators, e.g.,
retail trade, imports and exports, unemployment, and so on. It basically consists
of modeling the original series with an ARIMA model and extending the series
with forecasts up to 3 years ahead. Then each component of the extended series
is estimated using moving averages that are symmetric for middle observations
and asymmetric for both end years. The latter are obtained via the convolution of
Census II-X11 variant weights and the ARIMA model extrapolations.
X12ARIMA The X12ARIMA is an enhanced version of the X11ARIMA sea-
sonal adjustment software. The major modifications concern: (1) extending the
automatic identification and estimation of ARIMA models for the extrapolation
option to many more than the three models available in X11ARIMA, and (2) esti-
mating trading day variations, moving holidays, and outliers in what is called
the regARIMA model. The X12ARIMA method has all the seasonal adjustment
capabilities of the X11ARIMA variant. The same seasonal and trend moving
averages are available, and the programs till offer the Census II-X11 calendar
and holiday adjustment routines incorporated in X11ARIMA. Several other
new options have been included, such as: sliding spans diagnostic procedures,
capability to produce the revision history of a given seasonal adjustment, new
options for seasonal and trend-cycle filters, several new outlier detection options
for the irregular component, and a pseudo-additive seasonal adjustment mode.
X13ARIMA-SEATS The newest in the X11 family of seasonal adjustment soft-
ware, it integrates an enhanced version of X12ARIMA with an enhanced version
of SEATS to provide both nonparametric X11-type seasonal adjustments and
ARIMA model-based SEATS-type adjustments, combined with the diagnostics
available in X12ARIMA.