University of Rhode Island

DigitalCommons@URI

Open Access Master's Theses

2018

Bayesian Model Averaging of Space Time Car Models with

Application to U.S. House Price Forecasting
Caoxin Sun
University of Rhode Island, [email protected]

Follow this and additional works at: https://digitalcommons.uri.edu/theses

Recommended Citation
Sun, Caoxin, "Bayesian Model Averaging of Space Time Car Models with Application to U.S. House Price
Forecasting" (2018). Open Access Master's Theses. Paper 1181.
https://digitalcommons.uri.edu/theses/1181

This Thesis is brought to you for free and open access by DigitalCommons@URI. It has been accepted for inclusion
in Open Access Master's Theses by an authorized administrator of DigitalCommons@URI. For more information,
please contact [email protected].
BAYESIAN MODEL AVERAGING OF SPACE TIME CAR MODELS

WITH APPLICATION TO U.S. HOUSE PRICE FORECASTING

BY

CAOXIN SUN

A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE

REQUIREMENTS FOR THE DEGREE OF

MASTER OF SCIENCE

IN

STATISTICS

UNIVERSITY OF RHODE ISLAND

2018
 MASTER OF SCIENCE THESIS

OF

CAOXIN SUN

APPROVED:
Thesis Committee:

Major Professor Gavino Puggioni

Steffen Ventz

Todd Guilfoos

Nasser H. Zawia
DEAN OF THE GRADUATE SCHOOL

UNIVERSITY OF RHODE ISLAND

2018
 ABSTRACT

The housing market has been a significant contribution to U.S. GDP.

Forecasting the house price growth rate helps to regulate risks associated with

the housing sector and further helps to stabilize the economy. However, due

to the volatility in the housing market, forecasting the house price growth

rate has been a tough task. In this thesis, we built a conditional autoregres-

sive model incorporated with bayesian model averaging (BMA-CAR) based on

quarterly observations from 1976 to 1994 and tested forecasting capability over

1995 to 2012. We extended upon the results of Bork [International Journal of

Forecasting, 31, 1 (2015)] to include the effects of spatial autocorrelation but

inhibited the allowance for the model and coefficients shifts over time. Our

model is based on a hierarchical structure that allows BMA to average out the

effects from predictors along with CAR model to account for the remaining

spatial structures in the data.

 ACKNOWLEDGMENTS

I would like to thank the Computer Science and Statistics Department

affording me the financial support studying here as a Master Student; my

thesis advisor, Dr. Gavino Puggioni, who helped to shape the framework of

my thesis and encouraged me to challenge myself during my struggle time.

After taking his Bayesian Statistics class, I decided to pursue a degree in this

department. He is the mentor that guided me through the program and lead

me into the beautiful world of Statistics; my committee members (Dr. Steffen

Ventz and Dr. Todd Guilfoos) who gave me constructive and insightful advice

on my thesis writing on model specifications and economics studies; my friends

for the sleepless nights we were working together before deadlines, and for all

the fun we have had in the last two years.

iii
 TABLE OF CONTENTS

ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii

ACKNOWLEDGMENTS . . . . . . . . . . . . . . . . . . . . . . . iii

TABLE OF CONTENTS . . . . . . . . . . . . . . . . . . . . . . . iv

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii

CHAPTER

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Studying Housing Market . . . . . . . . . . . . . . . . 1

1.2 Areal Data and Spatial CAR . . . . . . . . . . . . . . 3

1.3 Bayesian Model Averaging . . . . . . . . . . . . . . . . 4

List of References . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Bayesian Model Averaging of Space Time Conditional

Autoregressive Models(BMA-CAR) . . . . . . . . . . . . . 8

2.1 Set Up of the Hierarchical Structure . . . . . . . . . . 8

2.2 CAR for Spatial Temporal Random Effects φkt . . . . . 9

2.3 Prior Distributions . . . . . . . . . . . . . . . . . . . . 11

2.3.1 Model Space Priors . . . . . . . . . . . . . . . . 11

2.3.2 Parameter Space Priors . . . . . . . . . . . . . 12

2.4 Posterior Computation . . . . . . . . . . . . . . . . . . 14

2.5 Forecasting . . . . . . . . . . . . . . . . . . . . . . . . 15

2.5.1 Bayesian Prediction . . . . . . . . . . . . . . . . 15

iv
 Page

2.5.2 MSFE and MSFE ratio . . . . . . . . . . . . . . 15

List of References . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3.1 Data Description . . . . . . . . . . . . . . . . . . . . . 18

3.2 Preliminary Analysis of the Data . . . . . . . . . . . . 20

3.2.1 Neighbor Structure . . . . . . . . . . . . . . . . 20

3.2.2 Exploratory Data Analysis . . . . . . . . . . . . 21

3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.3.1 Model Selection Results . . . . . . . . . . . . . 26

3.3.2 Estimation Results . . . . . . . . . . . . . . . . 28

3.3.3 Forecasting Results . . . . . . . . . . . . . . . . 30

List of References . . . . . . . . . . . . . . . . . . . . . . . . . . 35

4 Discussion and Future Work . . . . . . . . . . . . . . . . . . 41

4.1 Performance Gain by Spatial Components and Limitations 41

4.1.1 Model Dimension and Variable Inclusion . . . . 41

4.1.2 Limitations of a Global Fixed Effect . . . . . . 42

4.1.3 An Empirical Model . . . . . . . . . . . . . . . 43

4.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . 44

4.2.1 Data Acquisition . . . . . . . . . . . . . . . . . 44

4.2.2 Improvement of BMA-CAR Model . . . . . . . 45

List of References . . . . . . . . . . . . . . . . . . . . . . . . . . 46

BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

v
 LIST OF FIGURES

Figure Page

1 Time Series of real house price growth rate and all the co-
variates. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2 Neighbor structure used in CAR models. . . . . . . . . . . 21

3 Eight BEA regions. . . . . . . . . . . . . . . . . . . . . . . 22

4 Time Series of real house price growth rate. 48 states are

grouped in the eight BEA regions. Dashed line separates
training data from test data. Note that the data was trans-
formed by a factor of 400 (see equation 11). . . . . . . . . 24

5 P-value from Moran’s I test of observed house price growth

rate over time. Red dashed line is the 0.05 significant level. 25

6 ACF plot of observed house price growth rate against time.

Thick black line is the mean across the 48 states. Two
dashed lines are ±0.16. . . . . . . . . . . . . . . . . . . . . 25

7 Posterior model size distribution after fitting BMA-CAR. . 27

8 Posterior model size distribution after fitting BMA-CAR

with dilution prior and power of 1. . . . . . . . . . . . . . 28

9 p-value from Moran’s I test on residuals over time. Red

dashed line is the 0.05 significant level. . . . . . . . . . . . 29

10 ACF plot of residuals against time. Thick black line is the

mean across the 48 states. Two dashed lines are ±0.16. . . 30

11 Cumulative squared forecast error difference. . . . . . . . . 32

12 Forecasting results across the 48 states. . . . . . . . . . . . 35

13 Forecast vs. observed: best predicted states (top 4) vs.

worst (bottom 4). Red dashed line represents forecast
from BMA-CAR; blue dashed line represents forecast from
MEAN benchmark. Blue line is the historic data. Vertical
line is the separation of the boom (1995:1-2006:4) and bust
period (2007:1-2012:4 ). . . . . . . . . . . . . . . . . . . . 36

vi
Figure Page

14 MSFE ratio: best predicted states (top 4) vs. worst (bottom

4). Red dashed line represents the MSFE for BMA-CAR;
black line is MSFE for MEAN benchmark. Vertical line rep-
resents the separation of the boom and bust period. Notice
the different scale on the y-axis. . . . . . . . . . . . . . . . 37

15 Forecast vs. observed: top 4 volatility states (top) vs. bot-

tom 4 (bottom). . . . . . . . . . . . . . . . . . . . . . . . 38

16 MSFE ratio: top 4 volatility states (top) vs. bottom 4

(bottom). . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

17 Forecasting results across the 48 states. . . . . . . . . . . . 44

vii
 LIST OF TABLES

Table Page

1 Predictors and Abbreviations . . . . . . . . . . . . . . . . 20

2 Average of correlation coefficients within and between re-

gions: New England(NE), Mideast(ME), Great Lakes(GL),
Plains(PL), Southwest(SW), Rocky Mountain(RM),Far
West(FW). . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3 Top 5 Most Visited Models . . . . . . . . . . . . . . . . . 27

4 Top 5 Most Visited Models with dilution prior and power

of 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

5 Forecast errors across states . . . . . . . . . . . . . . . . . 33

viii
 CHAPTER 1

Introduction
1.1 Studying Housing Market

Understanding the housing market dynamics are crucial, as the housing

sector constitutes a significant share of the GDP and it is the largest compo-

nent of household wealth in the U.S. [1]. The Bureau of Labor Statistics has

estimated in 2010 that roughly 24 percent of the total consumption of Amer-

ican homeowners goes toward housing [2]. Starting from the late 1990s, the

U.S. housing market was in a boom period until 2007 when the sub-prime crisis

occurred and was followed instantly by a downturn. Modeling and predicting

U.S. housing price has received attention from governments, real estate devel-

opers and investors. However, it has been a challenging task due to the strong

vulnerability of the housing sector to structural changes, macroeconomic poli-

cies, regime switching, and market imperfections [1].

For the past four decades, many economists have adopted time series ap-

proaches to study the relationship between the U.S. house price and the socio-

economic variables. Originally, a large portion of the research pool focused on

exploring explanatory variables in the time series regression ([3],[4],[5]) and/or

modifying the error terms([6],[7],[8]). Recently, research focused on developing

dynamic models that allow for time-varying coefficients. These models could

substantially improve prediction by accounting for subsample parameter in-

stability seen in real estate data. Most common methods include, but are not

limited to, regime switching models ([9],[10]) and AR, VAR, GARCH models

coupled with Kalman Filter techniques. ([11],[1], [12]).

As noted by [13], it is possible that contiguous states may influence each

other’s housing prices. [2] grouped real house price into 8 Bureau of Economic

1
Analysis (BEA) regions. The within region correlation is larger than the be-

tween region correlation with only one exception. Other studies, such as [14],

proved the necessity of using spatial model accounting for dependent residuals

arose from only using ordinary least squares. There had been several studies

that accounted for spatial autocorrelation, which were summarised in [2]. One

major branch of methods is the spatially adapted version of VARS (SpVAR)

[15]. SpVAR belongs to the group of Spatial-Temporal Autoregressive Moving

Average (STARMA) methods. Seemingly Unrelated Regression (SUR) and

error panel data models have also been largely used, though it favors a rel-

atively small number of regions just as STARMA models do. Other authors

attempted common correlated effects estimator (CCE)[16][13] and Spatial Dy-

namic Structural Equation models (SD-SEM)[2] to model common shocks on

housing price.

In the very recent study [12], the authors implemented Dynamic Model Se-

lection and Dynamic Model Averaging methods and demonstrated the im-

portance of allowing for both time-varying parameters and model changes.

One limitation, however, is that the natural spatial structures existing in the

data are not considered in their model. In fact, univariate analysis on each

state separately was performed. In this study, we propose to use Conditional

Autoregressive (CAR) methods incorporated with Bayesian Model Averaging

(BMA) as an alternative forecasting model on the data from [12]. We devel-

oped the model in a hierarchical structure that is composed of a fixed effect

and a random effect. Fixed effect models the influence from the predictors and

the random effect captures the spatial-temporal variations in the data after

removing the fixed effect. We hope to achieve better forecasts by taking into

consideration the spatial autocorrelation.

2
1.2 Areal Data and Spatial CAR

Unlike point process where points are all neighbors to each other on a

continuous surface, areal data have well defined boundaries and observed data

are frequently aggregations within the boundaries or the areal units themselves

constitute the units of observation [17]. CAR models are widely used to de-

scribe the spatial variation of areal data. CAR models have been extensively

used for the analysis of spatial data in diverse areas, such as demography,

economics, epidemiology and geography [18].

Modeling spatial autocorrelation for areal data requires the creation of neigh-

bor structures and corresponding weight matrix. There are multiple ways

to define ”neighbor”, such as contiguity-based, graph-based, and k nearest

neighbor, after which the neighbor object is converted into a spatial matrix

to quantify spatial dependence. Spatial weights matrix W has elements ωks

that represents the weights of the spatial link between spatial units Sk and Ss .

When little is known about the spatial process, a common approach is to take

binary representation in which one is for neighbors and zero otherwise [17].

CAR model is a method for smoothing areal data that was originally proposed

by [19]. In a CAR model, the spatial component of the center areal unit is

seen as conditionally dependent on a weighted average of the spatial compo-

nents from all the other units, and the weights come from previously created

weight matrix. If we use yk to indicate any observed value of interest, the full

conditional distribution for yk has the following format:

X ωks ys τ 2
[Yk |ys , s 6= k] ∼ N ( , ) (1)
s
ωk+ ωk+
τ2
where the ωk+ is the total number of neighbors of areal unit k, and ωk+
can

be viewed as the local variance for areal unit k. The CAR model is often used

as a prior for the random effect in the context of a hierarchical model. In a

3
hierarchical model, we normally use φk to represent the spatial random effect

that failed to be modeled by the fixed effects. Replacing the yk in Equation 1

by φk and we would have:

X ωks φs τ2
[φk |φs , s 6= k] ∼ N ( , ) (2)
s
ωk+ ωk+

This model can be denoted as CAR(W , σ −2 ) [20] with a joint distribution as

the following by applying Brook’s Lemma [21]:

 
1 T
[φ|W ] ∝ exp − 2 φ (D − W )φ (3)
2σ

where D is the diagonal matrix with elements ωk+ . The precision matrix

Σ−1
φ = (D−W ) is singular so that the above joint distribution is improper [20].

After some modifications, a proper CAR model has the following covariance

[22] :

σ 2 Σ−1
φ = ρs (D − W ) + (1 − ρs )I (4)

where 0 ≤ ρs ≤ 1. If ρs = 1, CAR model goes back to the improper case as in

Equation 1.2. The precision matrix is non-singular if Σ−1

φ 6= 1.

1.3 Bayesian Model Averaging

As for the fixed effects, a set of predictors are needed to capture covariates

effects. However, it remains a question as to which variables to include in the

model; including all covariates adds to computational burden as well as over-

fitting. This variable selecting problem has a natural Bayesian solution [23].

Advantages of using BMA include but is not limited to automatic adjustment

for multiple comparisons and efficient model space exploration as well as lower

forecasting error compared to using one single model [24].

Since the number of variables is relatively small in our case, we can exhaust

all possible models and compute their posterior probability in the model space

4
M. Suppose a set of K models M = M1 , ..., Mk are under consideration for

data Y , and θk is a vector of unknown parameters that indexes the members

of Mk . The probability that Mk in fact generated the data, conditionally on

having observed Y , is the following posterior model probability [25]:

p(Y |Mk )p(Mk )

p(Mk |Y ) = P (5)
k p(Y |Mk )p(Mk )

where p(Mk ) is the prior for model Mk and p(Y |Mk ) is the marginal

distribution of data under this model that is achieved by integrating out θk :

Z Z
p(Y |Mk ) = p(Y |θk , Mk )p(θk |Mk )dθk (6)

List of References

[1] H. S. Guirguis, C. I. Giannikos, and R. I. Anderson, “The us housing mar-

ket: asset pricing forecasts using time varying coefficients,” The Journal
of real estate finance and economics, vol. 30, no. 1, pp. 33–53, 2005.

[2] P. Valentini, L. Ippoliti, and L. Fontanella, “Modeling us housing prices

by spatial dynamic structural equation models,” The Annals of Applied
Statistics, pp. 763–798, 2013.

[3] D. DiPasquale and W. C. Wheaton, “Housing market dynamics and the

future of housing prices,” Journal of urban economics, vol. 35, no. 1, pp.
1–27, 1994.

[4] N. Pain and P. Westaway, “Modelling structural change in the uk hous-

ing market: a comparison of alternative house price models,” Economic
Modelling, vol. 14, no. 4, pp. 587–610, 1997.

[5] S. Holly and N. Jones, “House prices since the 1940s: cointegration, de-
mography and asymmetries,” Economic Modelling, vol. 14, no. 4, pp.
549–565, 1997.

[6] J. M. Quigley, “A simple hybrid model for estimating real estate price
indexes,” Journal of Housing Economics, vol. 4, no. 1, pp. 1–12, 1995.

[7] J. M. Abraham and P. H. Hendershott, “Bubbles in metropolitan housing

markets,” National Bureau of Economic Research, Tech. Rep., 1994.

[8] S. Malpezzi, “A simple error correction model of house prices,” Journal

of housing economics, vol. 8, no. 1, pp. 27–62, 1999.

5
 [9] S. Hall, Z. Psaradakis, and M. Sola, “Switching error-correction models of
house prices in the united kingdom,” Economic Modelling, vol. 14, no. 4,
pp. 517–527, 1997.

[10] G. W. Crawford and M. C. Fratantoni, “Assessing the forecasting per-

formance of regime-switching, arima and garch models of house prices,”
Real Estate Economics, vol. 31, no. 2, pp. 223–243, 2003.

[11] J. Clapp and C. Giaccotto, “Evaluating house price forecasts,” Journal

of Real Estate Research, vol. 24, no. 1, pp. 1–26, 2002.

[12] L. Bork and S. V. Møller, “Forecasting house prices in the 50 states using
dynamic model averaging and dynamic model selection,” International
Journal of Forecasting, vol. 31, no. 1, pp. 63–78, 2015.

[13] S. Holly, M. H. Pesaran, and T. Yamagata, “A spatio-temporal model of

house prices in the usa,” Journal of Econometrics, vol. 158, no. 1, pp.
160–173, 2010.

[14] J. P. LeSage, “Bayesian estimation of spatial autoregressive models,” In-

ternational Regional Science Review, vol. 20, no. 1-2, pp. 113–129, 1997.

[15] T. H. Kuethe and V. O. Pede, “Regional housing price cycles: a spatio-

temporal analysis using us state-level data,” Regional studies, vol. 45,
no. 5, pp. 563–574, 2011.

[16] B. Van Dijk, P. H. Franses, R. Paap, and D. Van Dijk, “Modelling regional
house prices,” Applied Economics, vol. 43, no. 17, pp. 2097–2110, 2011.

[17] R. S. Bivand, E. J. Pebesma, V. Gomez-Rubio, and E. J. Pebesma, “Ap-

plied spatial data analysis with r.” Springer, 2008, vol. 747248717, ch. 9,
pp. 237–268.

[18] V. De Oliveira, “Bayesian analysis of conditional autoregressive models,”

Annals of the Institute of Statistical Mathematics, vol. 64, no. 1, pp. 107–
133, 2012.

[19] J. Besag, “Spatial interaction and the statistical analysis of lattice sys-
tems,” Journal of the Royal Statistical Society. Series B (Methodological),
pp. 192–236, 1974.

[20] S. Guha and L. Ryan, “Spatio-temporal analysis of areal data and discov-
ery of neighborhood relationships in conditionally autoregressive models,”
2006.

[21] D. Brook, “On the distinction between the conditional probability and
the joint probability approaches in the specification of nearest-neighbour
systems,” Biometrika, vol. 51, no. 3/4, pp. 481–483, 1964.

6
[22] B. N. Leroux BG, Lei X, “Estimation of disease rates in small areas: a new
mixed model for spatial dependence,” in Statistical models in epidemiol-
ogy, the environment, and clinical trials. Springer, 2000, p. 179191.

[23] P. D. Hoff, A first course in Bayesian statistical methods. Springer Sci-

ence & Business Media, 2009.

[24] A. Rodriguez and G. Puggioni, “Mixed frequency models: Bayesian ap-

proaches to estimation and prediction,” International Journal of Fore-
casting, vol. 26, no. 2, pp. 293–311, 2010.

[25] H. Chipman, E. I. George, R. E. McCulloch, M. Clyde, D. P. Foster, and

R. A. Stine, “The practical implementation of bayesian model selection,”
Lecture Notes-Monograph Series, pp. 65–134, 2001.

7
 CHAPTER 2

Bayesian Model Averaging of Space Time Conditional

Autoregressive Models(BMA-CAR)

2.1 Set Up of the Hierarchical Structure

Suppose the study region covers a set of k = 1, ..., K non-overlapping

areal units S = S1 , ..., SK and data are recorded for each time unit for t =

1, ..., T consecutive time periods. A generalized linear mixed effects model is a

convenient candidate for modeling this type of data. The hierarchical structure

is as following [1]:

Ykt |µkt ∼ f (ykt |µkt , σ 2 )

g(µkt ) = Xkt β + φkt

We define here Ykt as the observed house price growth rate at time t and

location k that has in total K × T rows of observations, and Xkt as a vector

of p + 1 covariates (including the intercept). β is a vector of p + 1 that are

invariant to space and time changes. φkt has a length of K ×T . We expand the

second line of the two equations above to show the dimension of each variable

as the following:
     
µ11 x11,0 ... x11,p φ11
 ..   ..   ... 
 .   .

 β0  
 µK1   xK1,0 ... xK1,p   φK1 
     
..
= +

 µ12

  x12,0 ... x12,p

 .  φ12 

 .
 ..
  .
  ..

 βp  . 
 .. 
µKT xKT,0 ... xKT,p φKT

And we provide here an explanation of the two-stage structure in the

above hierarchical model:

(i) Ykt |µkt ∼ f (ykt |µkt , σ 2 ). At the first stage, the observed value of the

8
 house price growth rate Ykt at specific time t and location k is from a

normal distribution with true mean µkt and observational error σ 2 .

(ii) g(µkt ) = Xkt β + φkt . At the second stage, the likelihood was chosen to

be Gaussian since the house price growth rate is a continuous response.

Therefore, the g function is just the identity link. With identity link,

g(µkt ) becomes µkt and it is approximated by the fixed effects Xkt β that

captures the influence from predictors plus the spatial-temporal random

effects φkt that models the patterns seen in the data after fitting the fixed

effects. We assume here that the second stage of the model is a simple

linear mixed effects model, that both of the components are additive and

enter the model linearly without higher orders or interaction terms.

2.2 CAR for Spatial Temporal Random Effects φkt

In the context of a hierarchical model, we would like to use CAR model

as a prior for this random effect φkt . The random effect φkt can be modeled in

several methods.

(i) Linear CAR

t − t̄
φkt = β1 + φk + (α + δk )
T
PK !
ρint j=1 ωkj φj 2
τint
φk |φ−k , W ∼ N ,
ρint K
P PK
j=1 ωkj + 1 − ρint ρint j=1 ωkj + 1 − ρint
PK !
ρslo j=1 ωkj δj 2
τslo
δk |δ−k , W ∼ N ,
ρslo K
P PK
j=1 ωkj + 1 − ρslo ρint j=1 ωkj + 1 − ρslo

In this model, a linear trend with time is assumed. Each areal unit k

has its own variation of intercept φk and slope δk from the mean linear

trend (intercept of φ1 and slope α) [2]. Both of φk and δk are modeled

by CAR prior from the paper [3].

9
(ii) Time Autoregressive CAR(CARar)

CARar is a spatially autocorrelated autoregressive time series model first

introduced by [4]. The single set of random effects φ = (φ1 , φ2 , ..., φT ) is

decomposed as:
T
Y
f (φ1 , φ2 , ..., φT ) ∼ f (φ1 ) f (φj |φj−1 )
j=2

The decomposition above induces temporal autocorrelation by allowing

φj to depend on φj−1 . Then, spatial autocorrelation is introduced at φ1

by using Gaussian Markov Random Field (GMRF) prior that is constant

over time. The joint prior distribution for φ1 is given by:

φ1 ∼ N (0, τ 2 Q(W , ρS )−1 )

where the spatial component is introduced by the variance τ 2 Q(W , ρS )−1

that corresponds to proper CAR prior. The precision Q is defined in

the previous chapter in the same way as displayed in Equation 4 but

with slightly differrent representation [3]. D is the diagonal matrix with

elements ωk+ but can also be written as W 1:

Q(W , ρS ) = ρS [diag(W 1) − W ] + (1 − ρS )I (7)

where I is the K ∗ K identity matrix and 1 is a vector of ones. The

ρS controls the spatial autocorrelation structure and represents λ as in

Equation 4: ρ = 1 indicates intrinsic CAR prior [5] where the conditional

expectation is the average of the random effects from neighbor units; ρ =

0 corresponds to independent random effects. Temporal autocorrelation

is then induced by:

φt |φt−1 ∼ N (ρT φt−1 , τ 2 Q(W , ρS )−1 )

10
 where ρT is the temporal autoregressive coefficient: ρT = 1 means strong

temporal autocorrelation and in fact the temporal process turns into a

random walk; ρT = 0 leads to temporal independence. In this model, the

multivariate temporal autoregressive component is introduced via the

mean ρT φt−1 [4].

(iii) Adaptive CAR

This model assumes the autoregressive structure as the previous model

CARar. However, it allows for localized spatial structure by modeling

the non-zero elements of the neighbor matrix W as unknown parameters

[6]. The adjacent elements w+ = {wks |k ∼ s} can be estimated and

wks ∈ (0, 1).

We implemented empirical studies on our dataset using CARlinear,

CARar and CARadaptive without model averaging. CARar and CARadap-

tive gave comparable results and were superior to CARlinear. Because CARar

is the simpler method and it gave slightly better forecasting results than

CARadaptive, we only used CARar for the final model.

2.3 Prior Distributions

2.3.1 Model Space Priors

Imagine we have M models with model index m{1, 2, ..., M } and in total

p covariates. It would be convinient to index each of the 2p models by the

vector γ = (γ1 , ...γp )0 , where γj is an indicator for inclusion of variable Xj

under model Mm . Many Bayesian variable selection implementations have

used independent priors of the Bernoulli form [7]:

p(Mm ) = η qγ (1 − η)(p−qγ ) (8)

11
 P
in which case qγ = γj the number of non-zero parameters under model Mm

and the hyperparameter η is the expected probability that each variable is

included. Beta prior is assumed for η ∼ Beta(1, 1).

A refinement of the Bernoulli priors is known in the literature as dilution priors.

In order to downweight the probability of Mm for the collinearity in X m , we

imposed dilution prior as shown in Equation 9 where Rm is the correlation

matrix such that h(|Rm |) ∝ (X m )T X m . The function h is monotone and it

satisfies h(1) = 1 and h(0) = 0. Common choices of h include h(a) = a and

h(a) = a1/2 [8]. This gives:

p(Mm ) ∝ h(|Rm |)η qγ (1 − η)(p−qγ ) (9)

2.3.2 Parameter Space Priors

We used g-prior to simplify our posteriors for the model averaging process.

We would like to introduce the concept of g-prior starting from an ordinary

linear regression case.

y = Xβ + 

Assume the goal is to obtain the posterior distribution of β and σ and we

do not use g-prior, we would need to sample from the full conditionals us-

ing the Markov-Chain Monte-Carlo procedure. If we let β and 1/σ have the

priors as multivariate normal(β0 , Σ0 ) and gamma(ν0 /2, ν0 σ02 /2), we obtain the

12
following:

β ∼ multivariate normal(m, V)

m = (Σ−1 T 2 −1 −1 T 2
0 + X X/σ ) (Σ0 β0 + X y/σ )

V = (Σ−1 T 2 −1
0 + X X/σ )

σ 2 ∼ inverse gamma ν0 + n]/2, [ν0 σ02 + SSR(β)]/2

n
X
SSR(β) = (yi − β T xi )2
i=1

Under a Zellner’s g-prior [9] β0 = 0, Σ0 = gσ 2 (XT X)−1 where amount of

information from n/g observations is assumed for our prior, we can simplify

the Gibbs Sampling as the Monte Carlo:

σ 2 ∼ inverse gamma ν0 + n]/2, [ν0 σ02 + SSRg ]/2

 
T g T −1 T
SSRg = y I − X(X X) X y
g+1

β ∼ multivariate normal(m, V)
g
m= (XT X)−1 XT y
g+1
g
V= σ 2 (XT X)−1
g+1

Now we would like to apply g-prior to our model. The value of g is commonly

decided as g ∼ max(n, p2 ), where n is the sample size, p is the number of

covariates including the intercept. The other parameters have the following

assignment of priors:

σ 2 ∼ Inverse Gamma(1, 0.01)

β m ∼ N (0, gσ 2 ((X m )T X m )−1 )

τ 2 ∼ Inverse Gamma(1, 0.01)

ρS , ρT ∼ Uniform(0, 1)

13
Here the prior for σ2 has the underlying form of σ2 ∼

Inverse Gamma(ν0 /2, ν0 σ02 /2).

2.4 Posterior Computation

CARBayesST is an extension of R package CARBayes, the latter of which

models spatial autocorrelation that remains in the data after the covariates

effects are accounted for [10]. CARBayesST is the first dedicated software for

spatial-temporal areal unit modeling with conditional autoregressive priors and

is capable of capturing temporally changing spatial dynamics [1]. We adopted

part of the code from the package and added Bayesian Model Averaging for

our model. We used Gibbs Sampling to sample from the full conditional distri-

butions since the closed forms of the posterior distributions are not available.

A description of algorithm is as follows.

(i) Sample the model index m{1, 2, ..., M } according to the posterior prob-

abilities p(Mm |Y ):

p(Y |Mm )p(Mm )

p(Mm |Y ) = P
p(Y |Mm )p(Mm )

and p(Y |Mm ) is computed by:

 −(n+ν0 )/2
g T
p(Y |Mm ) ∝ (1 + g) −qγ /2
ν0 σ02 + Y T Y − (βˆm ((X ) X )βˆm )
m T m
g+1

where βˆm is the ordinary least squares estimate under model Mm . Also,

here Y = Yobs − φkt .

(ii) Sample σ 2 from the marginal:

p(σ 2 |Y ) ∼ IG((ν0 + n)/2, (ν0 σ02 + SSRg )/2)

 
T g m m T m −1 m
SSRg = Y I− X ((X ) X ) X Y
g+1

14
(iii) Sample β from the full conditional:
 
g m T m −1 m g 2 m T m −1
p(β|Y, σ) ∼ N ((X ) X ) X Y, σ ((X ) X )
g+1 g+1

(iv) Update φkt using Metropolis-Hastings

(v) Update ρT using R Metropolis-Hastings

(vi) Update τ 2 using R inverse gamma

(vii) Update ρS using Metropolis-Hastings

2.5 Forecasting
2.5.1 Bayesian Prediction

Under iteration s and model γ, we set the last time point estimation of φkt

in the training set as the initial value for φt for prediction. Then we produce

one step ahead prediction from time point t as following:

(s) (s)
Yk,t+1 ∼ N (µk,t+1 , σ 2(s) )
(s) m (s)
µk,t+1 = (Xk,t+1 )(s) (β m )(s) + φk,t+1
(s) (s) (s) (s) (s)
φk,t+1 |φk,t ∼ M V N (ρT φk,t , τ 2(s) Q(W , ρS )−1 )

Multiple steps head predictions are implemented as an extension from one

step ahead prediction.

2.5.2 MSFE and MSFE ratio

Model validation is based on mean squared forecasting error (MSFE) on

out-of-sample set 1995:1-2012:4, in order to be consistent with [11]. MSFE is

derived from squared forecast error (SFE). SFE for model j is computed as:
S
1 X (s)
SF Ek,t,j = (Ŷ − Yk,t )2
S s=1 k,t,j
(s)
where Ŷk,t is the prediction of house price growth rate Y at location k and

time t from iteration s and model j; Yk,t is the observed Y at the same time

15
and location.

The benchmark model was chosen as expanding window OLS fit with only

intercept:
Pi=t+s−1
i=1Yi
ŶM EAN,t+s =
t+s−1

SFE for benchmark mean model is computed as:

SF Ek,t,M EAN = (Ŷk,t,M EAN − Yk,t )2

When checking the prediction performance by state, we introduce the MSFE

ratio r that is the ratio between MSFE from model j relative to the MSFE

from the beancmark:

2012:4
1 X
M SF Ek,j = SF Ek,t,j
T t=1995:1

2012:4
1 X
(10)
M SF Ek,M EAN = SF Ek,t,M EAN
T t=1995:1

M SF Ek,j
rk =
M SF Ek,M EAN

List of References

[1] D. Lee, A. Rushworth, and G. Napier, “Spatio-temporal areal unit mod-

elling in r with conditional autoregressive priors using the carbayesst pack-
age.”

[2] L. Bernardinelli, D. Clayton, C. Pascutto, C. Montomoli, M. Ghislandi,

and M. Songini, “Bayesian analysis of spacetime variation in disease risk,”
Statistics in medicine, vol. 14, no. 21-22, pp. 2433–2443, 1995.

[3] B. N. Leroux BG, Lei X, “Estimation of disease rates in small areas: a new
mixed model for spatial dependence,” in Statistical models in epidemiol-
ogy, the environment, and clinical trials. Springer, 2000, p. 179191.

16
 [4] A. Rushworth, D. Lee, and R. Mitchell, “A spatio-temporal model for
estimating the long-term effects of air pollution on respiratory hospital
admissions in greater london,” Spatial and spatio-temporal epidemiology,
vol. 10, pp. 29–38, 2014.

[5] M. Besag J, York J, “A bayesian image restoration, with two applications

in spatial statistics,” Ann Inst Statist Math, vol. 43, no. 1, pp. 1–20, 1991.

[6] A. Rushworth, D. Lee, and C. Sarran, “An adaptive spatiotemporal

smoothing model for estimating trends and step changes in disease risk,”
Journal of the Royal Statistical Society: Series C (Applied Statistics),
vol. 66, no. 1, pp. 141–157, 2017.

[7] H. Chipman, E. I. George, R. E. McCulloch, M. Clyde, D. P. Foster, and

R. A. Stine, “The practical implementation of bayesian model selection,”
Lecture Notes-Monograph Series, pp. 65–134, 2001.

[8] E. I. George et al., “Dilution priors: Compensating for model space

redundancy,” in Borrowing Strength: Theory Powering Applications–A
Festschrift for Lawrence D. Brown. Institute of Mathematical Statistics,
2010, pp. 158–165.

[9] A. Zellner, “On assessing prior distributions and bayesian regression anal-
ysis with g-prior distributions,” Bayesian inference and decision tech-
niques, 1986.

[10] D. Lee, “Carbayes version 4.6: An r package for spatial areal unit mod-
elling with conditional autoregressive priors,” University of Glasgow,
Glasgow, 2017.

[11] L. Bork and S. V. Møller, “Forecasting house prices in the 50 states using
dynamic model averaging and dynamic model selection,” International
Journal of Forecasting, vol. 31, no. 1, pp. 63–78, 2015.

17
 CHAPTER 3

Application

3.1 Data Description

Data are retrieved from the provided appendix of the article [1]. The same

data were used in order to make the modeling and forecasting comparable.

House price growth rate data were converted by:

 
Pk,t
yk,t = 400ln , k = 1, ..., 48 (11)
Pk,t−1

where Pk,t denotes the level of real house prices in state k and time t. The orig-

inal dataset consisted of 50 spreadsheets of 10 variables; with each spreadsheet

containing the response variable of quarterly measurements (1975:1-2012:4) of

real house price growth rate along with the other 9 predictors for each state.

We divided the whole dataset into training and testing set. Training period

has 3552 observations 48 states multiplied by 74 quarters (1976:03-1994:4);

testing period has 3456 rows with 72 quarters of data (1995:1-2012:4). We

split our data to stay in consistency with [1]’s methodology and so our fore-

casting results will be directly comparable.

To account for strong regional difference, both state-level predictors (price-

income ratio, unemployment rate, real per capita income growth and labor

force growth) and national level ones (30-year mortgage rate, the spread be-

tween 10-year and 3-month Treasury rates, industrial production growth, real

consumption growth and housing starts) were used. Table 1 summarized the

covariates and their abbreviations that are used later in this chapter. Figure

1 included the time series plots of both the dependent variable (house price

growth rate) and the independent variables (all the 9 predictors) to show their

variations with time.

18
Figure 1: Time Series of real house price growth rate and all the
covariates.

19
 Table 1: Predictors and Abbreviations

Level Variable Abbreviation

state price income ratio (in logs) PIR
state real per capita income growth (in logs and annualized) ING
state unemployment rate UNE
state labor force growth (in logs and annualized) LFG
national 30-year mortgage rate (in first differences) MOR
national spread between 10-year and 3-month Treasury rates SPR
national housing starts (in logs) HOS
national industrial production growth (in logs and annualized) IPG
national real consumption growth (in logs and annualized) RCG

Since Alaska and Hawaii do not share borders with the other states, they

would generate vectors of zeros in the spatial neighbor matrix that can lead

to computational error. Therefore, these two states were ruled out from our

analysis. Modeling and forecasting in this paper are based on the remaining

states.

3.2 Preliminary Analysis of the Data

3.2.1 Neighbor Structure

To qualitatively check the spatial associations between the states as well as

prepare ourselves with CAR modeling, we create a neighbor structure among

the 48 states. Out of the multiple neighbor defining ways, such as contiguity-

based, graph-based, and k nearest neighbor, here we chose to use contiguity

based method to create our neighbor matrix. We have explored several other

neighbor structures while they did not exhibit significant improvement in our

final model. Hence, here we skip the discussion on the selection of neighbor

structure. Eventually, we came with Queen style contiguity neighbor structure

that allows any two states sharing at least one point on the boundary to be

neighbors [2]. Neighboring structure used in our model is shown in Figure

2. This neighbor objects described above is then transformed into a binary

20
weight matrix W , where ωij = 1 if jth area is neighbor of ith area; and ωij = 0

otherwise. This is a conservative approach for creating spatial weight matrix,

as little is known about the assumed spatial process.

Figure 2: Neighbor structure used in CAR models.

3.2.2 Exploratory Data Analysis

We first tabulated the correlations among the BEA regions and plotted

the time series plot for each region to gain the first insight into the spatial

and temporal structure in our data. Previous studies have explored the cor-

relations within and between BEA regions to check spatial interactions. The

compositions for each BEA region are shown in Figure 3). BY studying the

correlation matrix between and within the BEA regions for the duration of the

year 1984 to 2011 [3] and the year of 1975 to 2003 [4] house price, it was found

that the within region correlations are larger than the between region correla-

tions with few exceptions. A similar feature is shown in our study using data

1975-2012 (see Table 2), that the diagonal correlation coefficients are generally

larger than the off-diagonal numbers. Exceptions include Great Lakes - Far

21
West and Southwest - Far west.

Figure 3: Eight BEA regions.

Table 2: Average of correlation coefficients within and between regions:

New England(NE), Mideast(ME), Great Lakes(GL), Plains(PL),
Southwest(SW), Rocky Mountain(RM),Far West(FW).

NE ME GL PL SE SW RM FW
NE 0.417 - - - - - - -
ME 0.398 0.400 - - - - - -
GL 0.289 0.345 0.400 - - - - -
PL 0.164 0.212 0.323 0.429 - - - -
SE 0.250 0.306 0.372 0.243 0.458 - - -
SW 0.147 0.211 0.315 0.261 0.329 0.375 - -
RM 0.090 0.148 0.339 0.243 0.299 0.374 0.400 -
FW 0.272 0.372 0.449 0.262 0.367 0.383 0.352 0.375

Figure 4 shows that there are similar patterns across the country as well

as regional similarities regarding the time series of real house price growth rate

(RHPGR). Most states exhibited greater fluctuations in the first 70 quarters,

which roughly correspond to the 1975 - 1993 time period. Since then, the

housing market growth rate stabilized and possessed a slight upward trend

22
until near the 110th quarter (2003) when large uphill and downhill movements

occurred. Therefore, forecasting over the last half of the time series by studying

the first half is not an easy task, since the test set contains features that are

fairly distinguishable from the training set.

Next, we quantify the spatial autocorrelation and temporal autocorrela-

tion existing in the data by using Moran’s test and time series autocorrelation

function (ACF).

Moran’s I test is one of the commonly used tests for spatial autocorrelation:
Pn Pn
n i=1 j=1 ωij (yi − ȳ)(yj − ȳ)
I = Pn Pn P n 2
i=1 j=1 ωij i=1 (yi − ȳ)

where the observations at areal units are denoted as yi , yj and spatial structure

as ωi,j [2]. Moran’s I values for each time point was plotted and shown in Figure

5. More than 70% of the time there was significant spatial autocorrelation

present. This finding further suggests the necessity of using CAR model to

account for the spatial structure.

Autocorrelation function ρt,s shows the correlation between time points t, s as:

ρt,s = Corr(Yt , Ys )

The ACF plot that has the correlation coefficient against lags was used to check

spatial autocorrrelation embedded in the data (Figure 6). The first five lags

showed large variation with respect to the temporal autocorrelation coefficient,

with a significant proportion appearing above the significance range bounded

by ±0.16. The mean of the coefficients across states also went above the range

for the first five lags, with the first and fifth lag right at the boundary. The

ACF plot of the original data exhibits temporal autocorrelation within the

data that suggests the necessity of using time series models.

23
Figure 4: Time Series of real house price growth rate. 48 states are
grouped in the eight BEA regions. Dashed line separates training data
from test data. Note that the data was transformed by a factor of 400
(see equation 11).

24
Figure 5: P-value from Moran’s I test of observed house price growth
rate over time. Red dashed line is the 0.05 significant level.

Figure 6: ACF plot of observed house price growth rate against time.
Thick black line is the mean across the 48 states. Two dashed lines
are ±0.16.

25
3.3 Results
3.3.1 Model Selection Results

We first examined the model selection results by checking the distribution

of the selected model sizes. Figure 7 and Figure 8 reveal the distribution of

the selected sample size for BMA-CAR without and with dilution prior re-

spectively. The two dilution prior models we tested gave rather similar results

thus only one of them is listed here. The two bar charts both reflected the

centrality around model sizes of four and five and the reduction of frequency

towards the two ends. The latter model (BMA-CAR without dilution prior)

sampled models with a size of five slightly more than the latter (BMA-CAR

with dilution prior).

We also investigated the variables selected by those most frequently vis-

ited models. Table 3 and Table 4 summarize the top five selected mod-

els without and with dilution prior h(γ) = γ. Results from the two

tables are almost identical, with the most visited models having four

to five predictors that evolve around price income ratio, unemployment

rate, labor force growth, mortgage rate and housing starts. These five

models account for > 73% of all the models been visited, and in fact,

the top two models which are PIR + UNE + LFG + MOR + HOS and

UNE + LFG + MOR + HOS alone account for > 60%. The combination

of variables of UNE + LFG + MOR + HOS has a significant weight in our

model selection process.

[1] plotted out the median value along with the 16th and 84th percentiles of

the model sizes foretasted over the testing period across the 50 states. The

model sizes exhibited an increasing trend, with the median slowly climbed up

from two to four during the span of the forecasting period. The 16th and

84th percentiles band were about one variable away from the median and the

26
width stayed stable during all time. Therefore, the variations of model sizes

across states and time were high, which is not too surprising considering the

boom and bust cycle occurred in our forecasting period. Their finding of the

variation in model sizes across time and space explained why our model, CAR-

BMA that uses global covariates across states and assumes the coefficients are

constant over time, did an excellent jobs in prediction in some states and not

so well for others. A detailed discussion is left to the next chapter.

Figure 7: Posterior model size distribution after fitting BMA-CAR.

Table 3: Top 5 Most Visited Models

model relative frequency

PIR + UNE + LFG + MOR + HOS 0.343
UNE + LFG + MOR + HOS 0.256
UNE + MOR + HOS 0.06
PIR + LFG + MOR + HOS 0.045
PIR + UNE + MOR + HOS 0.044

27
 Figure 8: Posterior model size distribution after fitting BMA-CAR
with dilution prior and power of 1.

Table 4: Top 5 Most Visited Models with dilution prior and power of
1.

model relative frequency

PIR + UNE + LFG + MOR + HOS 0.341
UNE + LFG + MOR + HOS 0.286
NUE + MOR + HOS 0.074
PIR + UNE + MOR + HOS 0.063
UNE + LFG + MOR + HOS 0.042

3.3.2 Estimation Results

Moran’s I tests were also conducted on the residuals after fitting of BMA-

CAR models. The goal is to see if spatial autocorrelation is accounted for

and if so, to what extent. BMA-CAR with dilution priors gave comparable

residuals as BMA-CAR, so results from BMA-CAR with dilution prior are

not included here. As revealed in Figure 9, most of the spatial autocorrelation

were removed. In fact, only 11% of the spatial autocorrelation were significant

28
which is a drastic decline compared to 70% before model fitting. The ACF

plot of the residuals is displayed in Figure 10. Except for the line (California)

that showed significant coefficients for the first five lags, the other states all

exhibited patterns of oscillations centered around the mean 0 after the first lag.

Also, most of the coefficients were reduced with regard to the amplitude. A

much higher proportion of coefficients were within the bounds when compared

to Figure 6. Hence, the temporal autocorrelation originated from the data

were largely removed after fitting the BMA-CAR model.

Figure 9: p-value from Moran’s I test on residuals over time. Red

dashed line is the 0.05 significant level.

29
 Figure 10: ACF plot of residuals against time. Thick black line is the
mean across the 48 states. Two dashed lines are ±0.16.

3.3.3 Forecasting Results

We first would like to check the overall forecasting performance over time

by using the squared forecasting error difference between the MEAN bench-

mark and our model. We adopted the method proposed by [5]:

2012:4
X
CDSF Ek,t = (SF Ek,t,M EAN − SF Ek,t,BM A−CAR )
t=1995:1

And made a little modification in that we summed CDSFE to each time point

instead of using the whole time periods:

t
X
CDSF Ek,t = (SF Ek,t,M EAN − SF Ek,t,BM A−CAR )
t=1995:1

In Figure 11, we plotted out the sum of CDSFE for 48 states as a total:
48
X
CDSF Ek,t
k=1

[1] plotted CDSFE against time for DMA, DMS, BMA, EW and AR1. EW

is a model that used equal weighting from K OLS regression models; AR1 is

30
a lag 1 time series model. As pointed out in the paper [1], plotting CDSFE

against time reveals the when our model is superior to the MEAN benchmark

(positive slope) and those in which the MEAN benchmark predicts better

(negative slope). Compared to DMA, DMS, BMA, EW and AR1, our model

showed excellent performance for the latter bust period (after the 60th quar-

ter or 2010:Q1) and that is comparable to DMA/DMS and much better than

BMA. However, the boom period from the mid 90s to 2006 and the initial

housing market meltdown time of 2007-2008, CDSFE from our model was

merely better than CDSFE from EW and AR1.

Regardless of the scale of the forecast error, it is worth noting that similar

patterns applied to almost all of the CDSFE lines where 2008:4, 2010:1 and

2011:3 experienced significant decline, especially to the best performance mod-

els DMA and DMS. [1] did not give an explanation. However, we attribute

the observation to policy changes. These three time points are in accordance

with the time lines of the Home Affordable Refinance Program (HARP) pro-

gram. This program was initiated by the U.S. government to help those with

mortgage problems refinance their home equities. There were three main time

points that marked the initiation of this policy and the modifications, namely

HARP 1.0, HARP 2.0 and HARP 3.0, that correspond to loan to the value

(LTV) threshold of 105%, 125% and no restriction [6]. Due to the high ac-

cordance between the timelines of the HARP program and the performance

drops, we believe that the deterioration of the model forecasting performance

can be ascribed to policy change.

31
 Figure 11: Cumulative squared forecast error difference.

After checking the overall picture for 50 states, we would like to compare

the forecasting performance between states. We summarised the MSFE re-

sults across the states for four of our models (BMS-CAR, BMA-CAR, BMA-

CAR(h(γ) = γ), BMA-CAR(h(γ) = γ 1/2 )) and compared with four models

(DMA, DMS, BMA, BMS) from [1]. The columns refer to the MSFE ra-

tio (Equation 10) across the states, which are the average, standard deviation,

minimum, maximum, and number of states in which BMA-CAR outperformed

the benchmark MEAN model. Note that the results from [1] which are DMA,

DMA, BMA, BMS in Table 5 were based on 50 states rather than 48. BMS-

CAR is a model similar to BMA-CAR, except that only the model with the

highest posterior probability (PIR + UNE + LFG + MOR + HOS) was used

for modeling.

For the BMA-CAR model, an average MSFE ratio of 0.818 for BMA-CAR as

shown in Table 5 is smaller than the reported MSFE ratio for BMA(0.858)

32
and BMS(0.903), indicating that incorporating spatial dynamics did improve

the forecast accuracy compared to solely static Bayesian Model Selection and

Bayesian Model Averaging [1]. Adding dilution prior did not improve the fore-

casting results by a significant amount. BMS-CAR showed poor performance

compared to the others. Since BMA-CAR model gave the best forecasting per-

formance and there is no significant difference adding dilution prior, discussion

for the rest of this chapter uses only BMA-CAR as an example.

Table 5: Forecast errors across states

Model Average Std Minimum Maximum < 1

BMS-CAR 0.838 0.36 0.510 1.60 36/48
BMA-CAR 0.818 0.24 0.505 1.58 40/48
BMA-CAR(h(γ) = γ) 0.826 0.24 0.516 1.55 40/48
BMA-CAR(h(γ) = γ 1/2 ) 0.833 0.25 0.514 1.57 40/48
DMA 0.751 0.12 0.546 0.961 50/50
DMS 0.796 0.15 0.501 1.165 46/50
BMA 0.858 0.14 0.548 1.177 44/50
BMS 0.903 0.15 0.591 1.188 35/50

To take a closer look at which states generated better and worse forecasts,

we use Figure 12 to illustrate the state-wise distribution of the MSFE ratio

between our model and the MEAN benchmark.

Most of the worst predicted states were center around the middle of the country

while the coastal states tend to be predicted okay. We ranked the states by

MSFE ratios and the top four and bottom four states were (Idaho, Georgia,

Washington, Delaware) and (Indiana, Nebraska, Iowa, Kentucky) respectively.

To discuss the reason behind this forecasting pattern, we would need to first

introduce the concept of volatility that came from [7]. Volatility was defined

as an indicator of the magnitude of the boom-bust cycle; larger volatility

implies a larger cycle. As found by them, housing markets growth volatility

and forecasting accuracy have almost a one-to-one relationship and coastal

33
states tend to have higher volatility than the interior states. The relationship

between the volatility and the MSFE level implies better forecast in the interior

states than the coastal states.

However, when it comes to MSFE ratio, coastal states tend to have smaller

ratios that represent larger performance gain than the MEAN benchmark. As

a matter of fact, MSFE ratio is only a relative number that it is not linked

to the forecast performance directly. For those interior (stable) states, using

the MEAN benchmark is not a bad choice since the historic average may not

be too different from the current point. As a result, using very complicated

models (such as CAR-BMA) will not cause great gain the performance. In

fact, two of the four least volatility states (Iowa, Kentucky see Figure 15) are

among the worst MSFE ratio states from our model BMA-CAR (Figure 13).

We plotted Figure 15 and Figure 16 to explore the forecasting performance

difference between the most and least volatile states. We also plotted out

the best performing states and worst performing states from our BMA-CAR

model (Figure 13 and 14) to discuss the model forecast performance variation

during the boom-and-bust cycle.

In Figure 13, numbers from the forecast were plotted against the observed.

In the bottom four figures that show the worst predicted states, BMA-CAR

tends to overestimate the growth rate before the recession and underestimate

the growth rate after the recession.

34
 Figure 12: Forecasting results across the 48 states.

List of References

[1] L. Bork and S. V. Møller, “Forecasting house prices in the 50 states us-
ing dynamic model averaging and dynamic model selection,” International
Journal of Forecasting, vol. 31, no. 1, pp. 63–78, 2015.

[2] R. S. Bivand, E. J. Pebesma, V. Gomez-Rubio, and E. J. Pebesma, “Ap-

plied spatial data analysis with r.” Springer, 2008, vol. 747248717, ch. 9,
pp. 237–268.

[3] P. Valentini, L. Ippoliti, and L. Fontanella, “Modeling us housing prices

by spatial dynamic structural equation models,” The Annals of Applied
Statistics, pp. 763–798, 2013.

[4] S. Holly, M. H. Pesaran, and T. Yamagata, “A spatio-temporal model of

house prices in the usa,” Journal of Econometrics, vol. 158, no. 1, pp.
160–173, 2010.

[5] A. Goyal and I. Welch, “Predicting the equity premium with dividend
ratios,” Management Science, vol. 49, no. 5, pp. 639–654, 2003.

[6] F. H. F. Agency, “Fhfa, fannie mae and freddie mac announce

harp changes to reach more borrowers,” https:// web.archive.

35
Figure 13: Forecast vs. observed: best predicted states (top 4) vs.
worst (bottom 4). Red dashed line represents forecast from BMA-
CAR; blue dashed line represents forecast from MEAN benchmark.
Blue line is the historic data. Vertical line is the separation of the
boom (1995:1-2006:4) and bust period (2007:1-2012:4 ).

36
Figure 14: MSFE ratio: best predicted states (top 4) vs. worst (bot-
tom 4). Red dashed line represents the MSFE for BMA-CAR; black
line is MSFE for MEAN benchmark. Vertical line represents the sep-
aration of the boom and bust period. Notice the different scale on the
y-axis.

37
Figure 15: Forecast vs. observed: top 4 volatility states (top) vs.
bottom 4 (bottom).

38
Figure 16: MSFE ratio: top 4 volatility states (top) vs. bottom 4
(bottom).

39
 org/ web/ 20120111193528/ http:// www.fhfa.gov/ webfiles/ 22721/
HARP release 102411 Final.pdf , [Online; accessed 3-April-2018].

[7] D. E. Rapach and J. K. Strauss, “Differences in housing price forecastabil-

ity across us states,” International Journal of Forecasting, vol. 25, no. 2,
pp. 351–372, 2009.

40
 CHAPTER 4

Discussion and Future Work

4.1 Performance Gain by Spatial Components and Limitations

Here we present a discussion on why our model predicted better in

the coastal states and not so well for the interior states. We compare our

forecasting results with the forecasting results from [1]. Here we talk about

the results in the context of MSFE ratio that represnt the forecast gain

from using just the MEAN benchmark. Even though the forecasting results

from DMA [1] are not the true model (there is no true model anyway),

DMA is a more flexible model that allows for both state-wise and time-

wise model variations. Moreover, DMA gave smaller forecasting error than

BMA-CAR (Table 5) thereby serving as a reasonable benchmark to check with.

4.1.1 Model Dimension and Variable Inclusion

We first start the discussion from the perspective of model dimension.

Our posterior distribution for model space favored models of four and five

variables. However, discoveries from [1] suggested that many of the low

volatility states preferred parsimonious models, that typically two to four

variables were needed in their model, and larger model sizes occurred more

often around the end of the forecasting period when there was a financial

crisis. Consequently, our model potentially selected too many variables for

these low volatility states and it leads to overfitting and bad forecasting

performance.

We further studied the influence from the choice of variables on forecasting

performance. [1] pointed out that the housing market is segmented so that no

41
single variable drives the whole housing market. They plotted the inclusion

probability (median, 16th and 84th percentiles) of each predictor against time

and found that the recent economic depression causes model changes. Most

of the variables were contained between 30% − 50% across time, with the

exception of housing starts and price-income ratio that reached up to 100%

right after 2007. In addition, there was a large spread of variable inclusion.

For example, housing starts inclusion at the beginning of 2008 ranged from

below 25% to near 100%. Thus, in the bust period, housing starts was

only an important factor for some states (Arizona and Nevada). Meanwhile,

price-income ratio became an essential component in Florida. This high

degree of space and time variation in choice of variables near the financial

crisis time makes it difficult to predict with globally assumed variables such as

in our model. The important drivers, such as housing starts and price-income

ratio were selected with high probabilities in our BMA-CAR posterior

distributions, with each selected 58% and 99% of the time. High inclusions

of housing starts and price-income ratio helped with the performance for

certain states that did have these predictors as underlying factors such as

Idaho and Georgia (refer to Figure 15 and 16). Nevertheless, containing these

predictors suppressed the model performance at states like Iowa and Kentucky.

4.1.2 Limitations of a Global Fixed Effect

To conclude, the overall large model sizes chosen by our model and inclu-

sion of predictors that are unimportant for those stable states give rise to less

forecast gain in the interior states than in the coastal states. Our model had

globally fixed effects and constant coefficients for each state and time. This

restricted the ability to capture the spatial variations regarding model shift as

42
well as the temporal fluctuations seen in the housing market.

The original analysis from [1] developed house prices for each state individu-

ally, while neglecting the fact that house price of the neighboring states have

autocorrelation features. By modeling spatial component as random effects,

we were able to ”borrow strength” across the states to come up with im-

proved estimate for the house price in each [2]. As to the degree of which

forecast performance improved, the lower and upper bounds are BMA/BMS

and DMA/DMS respectively. That is to say, despite significant improvement

compared to BMA/BMS methods due to additional spatial information, CAR-

BMA model was not able to beat the model with time-variant features. One of

the challenges in our study comes from the great difference seen in the house

price growth rate data between the training period and the testing period.

House price growth rate underwent significant fluctuations in the 1970s (see

Figure reffig:TS 8 regions). The magnitudes of these variations are even larger

than those observed near the recent economic recession. Several authors have

noted a marked decline in the volatility of real activity and in the volatility

and persistence of inflation since the early 1980s [3], [4], [5]. This structural

change in the economy may have implications for the housing market.

In addition, assuming constant observation error σ 2 and revolution error τ 2

may not be feasible in our case. The states with high and low volatility showed

distinguishable patterns of MSFE ratios, thus including the variations in the

volatility in our model can help reduce this MSFE ratio difference and improve

the overall forecasts.

4.1.3 An Empirical Model

Although BMA-CAR did not generate better forecasting results than

DMA, it is worth noting that we were able to obtain comparable result to

43
what was achieved by DMA by running a space-time CAR model while just

using unemployment, mortgage and consumption as predictors. As revealed in

Figure 17, all states except for Nebraska had MSFE ratio smaller than one so

that in 47/48 states the MEAN benchmark was beaten. Contrary to forecast-

ing results from DMA or BMA-CAR, results from CAR show generally better

predictions near the coast than inland, with the exception of South Dakota

and Nebraska.

Figure 17: Forecasting results across the 48 states.

4.2 Future Work

4.2.1 Data Acquisition

Our model could be improved at the data acquisition step, in that the

data aggregated at the state level may not be reflective on the housing market

at smaller scale. By leveraging out the housing data at the state level, we

would lose information about the variations in the housing data at smaller

scales. For instance, it is not reasonable to assume similar housing market

structure in the New York City and the rest of the New York State.

We can instead use data collected at the county level, which is more sim-

44
ilar to the scale of the housing prices study in real life. Data collected at

the metropolitan statistical area (MSA) level is also available and can be a

good candidate if the primary interest is to study the hosuing market in the

population dense areas. MSA is a geographical region with a relatively high

population density at its core and close economic ties throughout the area.

Currently, the United States Office of Management and Budget (OMB) has

defined 382 Metropolitan Statistical Areas (MSAs) for the United States [6].

One of the most populous MSA is New York-Newark-Jersey City that spans

to three states (New York, New Jersey and Pennsylvania) [7]. Although this

MSA covers three states, people tend to live and commute to work within this

one area and it makes sense to study them as a whole. Real life applications

include real estate investors that use MSA data to study housing trends and

population movement [7].

4.2.2 Improvement of BMA-CAR Model

We can improve the forecasts by altering the neighbor structure. Our

model currently used geometric definition of neighbors, that the states sharing

common borders were called ”neighbors”. This may not work well for areas

such as New England, where the states tend to be very small but highly

similar. One alternative way of defining neighbor structure is to use k-nearest

neighbors, in which we define each center states having k nearest neighbors

based on the distance from their centroids. In this way, we are not limited to

the states that are adjacent to each other.

Future research includes creating state-wise variables for each predictor and

use Metropolis Hastings algorithm to select a model. Creating a variable for

each state allows for variable and coefficient shift between states and further

adds to the flexibility of BMA-CAR that assumes global fixed effects. And due

45
to the large model space(29∗48 ) to sample from, Metroplis Hastings algorithm

is more suitable in this case so that only a portion of the models will be visited.

List of References

[1] L. Bork and S. V. Møller, “Forecasting house prices in the 50 states us-
ing dynamic model averaging and dynamic model selection,” International
Journal of Forecasting, vol. 31, no. 1, pp. 63–78, 2015.

[2] S. Banerjee, B. P. Carlin, and A. E. Gelfand, Hierarchical modeling and

analysis for spatial data. Crc Press, 2014.

[3] O. Blanchard and J. Simon, “The long and large decline in us output
volatility,” Brookings papers on economic activity, vol. 2001, no. 1, pp.
135–164, 2001.

[4] M. M. McConnell and G. Perez-Quiros, “Output fluctuations in the united

states: What has changed since the early 1980’s?” American Economic
Review, vol. 90, no. 5, pp. 1464–1476, 2000.

[5] J. H. Stock and M. W. Watson, “Has the business cycle changed? evi-
dence and explanations,” Monetary policy and uncertainty: adapting to a
changing economy, pp. 9–56, 2003.

[6] U. S. O. of Management and Budget, “Omb bulletin no. 15-

01: Revised delineations of metropolitan statistical areas, mi-
cropolitan statistical areas, and combined statistical areas,
and guidance on uses of the delineations of these areas,”
https://obamawhitehouse.archives.gov/sites/default/files/omb/bulletins/2015/15-
01.pdf, 2015, [Online; accessed 3-April-2018].

[7] Investopedia, “Definition of ’metropolitan statistical area - msa’,” https:

// www.investopedia.com/ terms/ m/ msa.asp, [Online; accessed 3-April-
2018].

46
 BIBLIOGRAPHY

Abraham, J. M. and Hendershott, P. H., “Bubbles in metropolitan housing

markets,” National Bureau of Economic Research, Tech. Rep., 1994.
Agency, F. H. F., “Fhfa, fannie mae and freddie mac announce harp changes to
reach more borrowers,” https:// web.archive.org/ web/ 20120111193528/
http:// www.fhfa.gov/ webfiles/ 22721/ HARP release 102411 Final.pdf ,
[Online; accessed 3-April-2018].
Banerjee, S., Carlin, B. P., and Gelfand, A. E., Hierarchical modeling and
analysis for spatial data. Crc Press, 2014.
Bernardinelli, L., Clayton, D., Pascutto, C., Montomoli, C., Ghislandi, M., and
Songini, M., “Bayesian analysis of spacetime variation in disease risk,”
Statistics in medicine, vol. 14, no. 21-22, pp. 2433–2443, 1995.
Besag, J., “Spatial interaction and the statistical analysis of lattice systems,”
Journal of the Royal Statistical Society. Series B (Methodological), pp.
192–236, 1974.
Besag J, York J, M., “A bayesian image restoration, with two applications in
spatial statistics,” Ann Inst Statist Math, vol. 43, no. 1, pp. 1–20, 1991.
Bivand, R. S., Pebesma, E. J., Gomez-Rubio, V., and Pebesma, E. J., “Applied
spatial data analysis with r.” Springer, 2008, vol. 747248717, ch. 9, pp.
237–268.
Blanchard, O. and Simon, J., “The long and large decline in us output volatil-
ity,” Brookings papers on economic activity, vol. 2001, no. 1, pp. 135–164,
2001.
Bork, L. and Møller, S. V., “Forecasting house prices in the 50 states using
dynamic model averaging and dynamic model selection,” International
Journal of Forecasting, vol. 31, no. 1, pp. 63–78, 2015.
Brook, D., “On the distinction between the conditional probability and the
joint probability approaches in the specification of nearest-neighbour sys-
tems,” Biometrika, vol. 51, no. 3/4, pp. 481–483, 1964.
Chipman, H., George, E. I., McCulloch, R. E., Clyde, M., Foster, D. P., and
Stine, R. A., “The practical implementation of bayesian model selection,”
Lecture Notes-Monograph Series, pp. 65–134, 2001.
Clapp, J. and Giaccotto, C., “Evaluating house price forecasts,” Journal of
Real Estate Research, vol. 24, no. 1, pp. 1–26, 2002.

47
Crawford, G. W. and Fratantoni, M. C., “Assessing the forecasting perfor-
mance of regime-switching, arima and garch models of house prices,” Real
Estate Economics, vol. 31, no. 2, pp. 223–243, 2003.

De Oliveira, V., “Bayesian analysis of conditional autoregressive models,” An-

nals of the Institute of Statistical Mathematics, vol. 64, no. 1, pp. 107–133,
2012.

DiPasquale, D. and Wheaton, W. C., “Housing market dynamics and the

future of housing prices,” Journal of urban economics, vol. 35, no. 1, pp.
1–27, 1994.

George, E. I. et al., “Dilution priors: Compensating for model space re-

dundancy,” in Borrowing Strength: Theory Powering Applications–A
Festschrift for Lawrence D. Brown. Institute of Mathematical Statis-
tics, 2010, pp. 158–165.

Goyal, A. and Welch, I., “Predicting the equity premium with dividend ratios,”
Management Science, vol. 49, no. 5, pp. 639–654, 2003.

Guha, S. and Ryan, L., “Spatio-temporal analysis of areal data and discov-
ery of neighborhood relationships in conditionally autoregressive models,”
2006.

Guirguis, H. S., Giannikos, C. I., and Anderson, R. I., “The us housing market:
asset pricing forecasts using time varying coefficients,” The Journal of real
estate finance and economics, vol. 30, no. 1, pp. 33–53, 2005.

Hall, S., Psaradakis, Z., and Sola, M., “Switching error-correction models of
house prices in the united kingdom,” Economic Modelling, vol. 14, no. 4,
pp. 517–527, 1997.

Hoff, P. D., A first course in Bayesian statistical methods. Springer Science

& Business Media, 2009.

Holly, S. and Jones, N., “House prices since the 1940s: cointegration, demog-
raphy and asymmetries,” Economic Modelling, vol. 14, no. 4, pp. 549–565,
1997.

Holly, S., Pesaran, M. H., and Yamagata, T., “A spatio-temporal model of

house prices in the usa,” Journal of Econometrics, vol. 158, no. 1, pp.
160–173, 2010.

Investopedia, “Definition of ’metropolitan statistical area - msa’,” https:

// www.investopedia.com/ terms/ m/ msa.asp, [Online; accessed 3-April-
2018].

48
Kuethe, T. H. and Pede, V. O., “Regional housing price cycles: a spatio-
temporal analysis using us state-level data,” Regional studies, vol. 45,
no. 5, pp. 563–574, 2011.

Lee, D., “Carbayes version 4.6: An r package for spatial areal unit modelling
with conditional autoregressive priors,” University of Glasgow, Glasgow,
2017.

Lee, D., Rushworth, A., and Napier, G., “Spatio-temporal areal unit modelling
in r with conditional autoregressive priors using the carbayesst package.”

Leroux BG, Lei X, B. N., “Estimation of disease rates in small areas: a new
mixed model for spatial dependence,” in Statistical models in epidemiol-
ogy, the environment, and clinical trials. Springer, 2000, p. 179191.

LeSage, J. P., “Bayesian estimation of spatial autoregressive models,” Inter-

national Regional Science Review, vol. 20, no. 1-2, pp. 113–129, 1997.

Malpezzi, S., “A simple error correction model of house prices,” Journal of

housing economics, vol. 8, no. 1, pp. 27–62, 1999.

McConnell, M. M. and Perez-Quiros, G., “Output fluctuations in the united

states: What has changed since the early 1980’s?” American Economic
Review, vol. 90, no. 5, pp. 1464–1476, 2000.

of Management, U. S. O. and Budget, “Omb bulletin no. 15-

01: Revised delineations of metropolitan statistical areas, mi-
cropolitan statistical areas, and combined statistical areas,
and guidance on uses of the delineations of these areas,”
https://obamawhitehouse.archives.gov/sites/default/files/omb/bulletins/2015/15-
01.pdf, 2015, [Online; accessed 3-April-2018].

Pain, N. and Westaway, P., “Modelling structural change in the uk housing

market: a comparison of alternative house price models,” Economic Mod-
elling, vol. 14, no. 4, pp. 587–610, 1997.

Quigley, J. M., “A simple hybrid model for estimating real estate price in-
dexes,” Journal of Housing Economics, vol. 4, no. 1, pp. 1–12, 1995.

Rapach, D. E. and Strauss, J. K., “Differences in housing price forecastability

across us states,” International Journal of Forecasting, vol. 25, no. 2, pp.
351–372, 2009.

Rodriguez, A. and Puggioni, G., “Mixed frequency models: Bayesian ap-

proaches to estimation and prediction,” International Journal of Fore-
casting, vol. 26, no. 2, pp. 293–311, 2010.

49
Rushworth, A., Lee, D., and Mitchell, R., “A spatio-temporal model for esti-
mating the long-term effects of air pollution on respiratory hospital admis-
sions in greater london,” Spatial and spatio-temporal epidemiology, vol. 10,
pp. 29–38, 2014.

Rushworth, A., Lee, D., and Sarran, C., “An adaptive spatiotemporal smooth-
ing model for estimating trends and step changes in disease risk,” Journal
of the Royal Statistical Society: Series C (Applied Statistics), vol. 66,
no. 1, pp. 141–157, 2017.

Stock, J. H. and Watson, M. W., “Has the business cycle changed? evidence
and explanations,” Monetary policy and uncertainty: adapting to a chang-
ing economy, pp. 9–56, 2003.

Valentini, P., Ippoliti, L., and Fontanella, L., “Modeling us housing prices
by spatial dynamic structural equation models,” The Annals of Applied
Statistics, pp. 763–798, 2013.

Van Dijk, B., Franses, P. H., Paap, R., and Van Dijk, D., “Modelling regional
house prices,” Applied Economics, vol. 43, no. 17, pp. 2097–2110, 2011.

Zellner, A., “On assessing prior distributions and bayesian regression analysis
with g-prior distributions,” Bayesian inference and decision techniques,
1986.

50