TimeSeriesAnalysisLectureNotes
Uploaded by
eruygurTimeSeriesAnalysisLectureNotes
Uploaded by
eruygurSee discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/319629569
Time Series Analysis Lecture Notes
Research · January 1995
DOI: 10.13140/RG.2.2.16026.95683
CITATIONS READS
0 907
1 author:
Vijayamohanan Pillai N
175 PUBLICATIONS 675 CITATIONS
SEE PROFILE
All content following this page was uploaded by Vijayamohanan Pillai N on 11 September 2017.
The user has requested enhancement of the downloaded file.
10/15/2012
Time Series Econometrics
1
Vijayamohanan Pillai N
Vijayamohan: CDS MPhil: Time Series 1 1 15 October 2012
Essential Readings: Time series:
Enders, Walter (1995) Applied Econometric Time Series,
John Wiley & Sons.
Hamilton, James D (1994) Time Series Analysis, Princeton
University Press.
Hendry, David F. (1995) Dynamic Econometrics, Oxford
University Press.
Makridakis, S., Wheelwright, S. C. and McGee, V. E.
(1983) Forecasting – Methods and Applications, Second
edition, John Wiley & Sons.
Vijayamohan: CDS MPhil: Time Series 1 2 15 October 2012
1
10/15/2012
Essential Readings: Time series:
My CDS Working Papers:
WP 312: Electricity Demand Analysis and Forecasting – The
Tradition is Questioned!
WP 346: A Contribution to Peak Load Pricing: Theory and
Application
Vijayamohan: CDS MPhil: Time Series 1 3 15 October 2012
Essential Readings: Panel Data Analysis:
Baltagi, B. H. (2001) Econometric Analysis of Panel
Data, 2nd edition, John Wiley.
Cheng, Hsian (1986) Analysis of Panel Data,
Cambridge University Press.
Vijayamohan: CDS MPhil: Time Series 1 4 15 October 2012
2
10/15/2012
Per Capita Per Capita
A Time Series
Year NSDP (Rs.) Year NSDP (Rs.)
1960-61 4313.79 1980-81 5392.59
1961-62 4214.82 1981-82 5404.59
1962-63 4206.08 1982-83 5472.79
1963-64 4269.36 1983-84 5186.76
10000
1964-65 4286.69 1984-85 5438.95 Per Capita
1965-66 4441.45 1985-86 5566.17
1966-67 4476.00 1986-87 5367.13 9000 Net State Domestic
1967-68 4607.29 1987-88 5523.34
8000
Product of Kerala at
1968-69 4976.05 1988-89 5993.69
1969-70 5006.78 1989-90 6305.48 Constant (1993-94) Prices
1970-71 5175.36 1990-91 6683.07 7000 (in Rs.)
1971-72 5352.05 1991-92 6830.41
1972-73 5391.20 1992-93 7324.32 6000
1973-74 5272.47 1993-94 8063.36
1974-75 5235.12 1994-95 8780.75
5000
1975-76 5365.44 1995-96 9149.74
1976-77 5211.88 1996-97 9494.41
1977-78 5194.39 1997-98 9517.95 1960 1965 1970 1975 1980 1985 1990 1995 2000
1978-79 5262.51 1998-99 10024.56
1979-80 5410.00 1999-00 10502.18
Vijayamohan: CDS MPhil: Time Series 1 5 15 October 2012
The modern era in time series started in
1927 with George Udny Yule
(Scottish statistician: 1871-1951)
"On a Method of Investigating Periodicities in
Disturbed Series, with Special Reference to Wolfer's
Sunspot Numbers", Philosophical Transactions of the
Royal Society of London, Ser. A, Vol. 226 (1927), pp.
267–298.
Vijayamohan: CDS MPhil: Time Series 1 6 15 October 2012
3
10/15/2012
Yule showed such a series can
be better described as a function of
its past values.
Thus he introduced the concept of
autoregression, even though he restricted
himself to an order of four or fewer terms.
Vijayamohan: CDS MPhil: Time Series 1 7 15 October 2012
Yule's approach was extended by
Sir Gilbert Thomas Walker
(British physicist and statistician: 1868-
1868-1958) :
defined (1931)
the general autoregressive scheme.
Gilbert Walker On Periodicity in Series of Related
Terms, Proceedings of the Royal Society of
London, Ser. A, Vol. 131, (1931) 518--
518--532.
--532.
Evgeny Evgenievich (or Eugen) Slutsky
(Russian Statistician, Economist: 1880-
1880-1948)
presented (1937)
the moving-average scheme.
Vijayamohan: CDS MPhil: Time Series 1 8 15 October 2012
4
10/15/2012
Herman Ole Andreas Wold
(1908- 1992)
(Swedish Statistician)
provided (1938; 1954)
the theoretical foundation
of combined ARMA processes.
Vijayamohan: CDS MPhil: Time Series 1 9 15 October 2012
George Edward Pelham Box and
Gwilym Meirion Jenkins (1970; 1976)
codified the applied univariate time
series ARIMA modelling
Box, George and Jenkins, Gwilym
(1970) Time series analysis:
Forecasting and control, San Francisco:
Holden-Day.
Vijayamohan: CDS MPhil: Time Series 1 10 15 October 2012
5
10/15/2012
Stationarity
A time series is stationary
if it fluctuates
around a constant
mean.
A non-
non-stationary series
includes a longer-term
secular trend.
Vijayamohan: CDS MPhil: Time Series 1 11 15 October 2012
Stationarity
But,
a large number of actual time series
are not stationary;
however, not an insolvable problem,
several methods to transform
a non-stationary series into a stationary one.
Vijayamohan: CDS MPhil: Time Series 1 12 15 October 2012
6
10/15/2012
YEAR MONTHS WPI All
Commodities
1971-72 april 101.7 Wholesale Price Index
may 102.3
june 104.7
july 106 (All Commodities)
august 107.2
september 107.5
october 106.6
november 105.1
december 103.9
january 107
february 107
march 108.1
1972-73 april 108.7
may 109.7
….. ……
….. ……
1987-88 april 381.2
may 390.3
june 394
july 400.6
august 409.6
september 408.9
october 409.5
november 411.1
december 410.4
january 416
february 415.8
march 417.6
Vijayamohan: CDS MPhil: Time Series 1 13 15 October 2012
Notation for the time series
A time series variable: yt.
Yt = the random variable Y at period t.
A time series: (yt, yt−1, yt−2, . . . , y1, y0).
T = standard number of
observations in a time series.
Vijayamohan: CDS MPhil: Time Series 1 14 15 October 2012
7
10/15/2012
Time series process
a stochastic process
A sequence of random variables
ordered in time.
Random (stochastic) variable:
takes values in a certain range
with probabilities,
specified by a pdf.
Vijayamohan: CDS MPhil: Time Series 1 15 15 October 2012
Time series: Stochastic Difference
Equation.
Difference equations:
Mathematics of Economic Dynamics.
deal with time as a discrete variable –
changing only from period to period –
DE expresses the value of a variable as a
function of its own lagged values, time and
other variables:
yt = a +b yt – 1:
a linear first order difference equation.
Vijayamohan: CDS MPhil: Time Series 1 16 15 October 2012
8
10/15/2012
If a random variable, ut , added,
yt = a + b yt–1 + ut :
Stochastic DE : Time Series.
First order Autoregressive, AR(1), process.
ut : equilibrium error, disturbance,
: information shock → ‘innovation’ →
the only new information to yt.
: assumed to be a ‘white noise process’.
Noise, white noise: from signal theory, physics
Vijayamohan: CDS MPhil: Time Series 1 17 15 October 2012
Noise? Disturbance : Error
Filter
Signal Output
(Input)
Noise
Noise
Filtering =
Decomposing output into input (signal) and noise.
Vijayamohan: CDS MPhil: Time Series 1 18 15 October 2012
9
10/15/2012
Filter
Signal Output
(Input)
Noise
Noise
Filtering = Decomposing output into input (signal
(signal)
signal) and noise.
noise.
Time series = Signal (input) + Noise;
= Regression function + Noise.
Filters: Regression; Auto Regression; Moving Average; ….
Vijayamohan: CDS MPhil: Time Series 1 19 15 October 2012
White noise
analogous to white light which contains all
frequencies.
a random signal (or process) with a flat
power spectral density.
The signal's power spectral density has equal
power in any band.
Vijayamohan: CDS MPhil: Time Series 1 20 15 October 2012
10
10/15/2012
White Noise
Calculated spectrum
of a generated
approximation of
Intensity (db)
white noise
Frequency (Hz)
An example
realization of a
white noise process.
Vijayamohan: CDS MPhil: Time Series 1 21 15 October 2012
A flat power spectral density =
Zero mean
Constant variance
Zero autocovariance = No autocorrelation.
A white noise process
Vijayamohan: CDS MPhil: Time Series 1 22 15 October 2012
11
10/15/2012
So, in yt = a + b yt–1 + ut :
ut : White noise
(a) E(ut) = E(ut - 1) = …. = 0
(b) Var(ut) = Var(ut - 1) = …. = σ2
(c) Cov(ut; ut–k ) = Cov(ut–s; ut–s–k) = …. = 0,
for all k and s; k ≠ 0:
‘No memory’.
(d) ut is normally distributed. (This assumption
not essential.)
Vijayamohan: CDS MPhil: Time Series 1 23 15 October 2012
White noise process = stationary process
Not vice versa!
For a stationary process,
Mean, variance, and covariance : all constant;
independent of t:
Termed covariance stationarity.
(or second order,
weak,
wide sense stationarity).
Vijayamohan: CDS MPhil: Time Series 1 24 15 October 2012
12
10/15/2012
Stationarity in two senses:
1. First-order (Strict) Stationarity
2. Second-order (Weak) Stationarity
Vijayamohan: CDS MPhil: Time Series 1 25 15 October 2012
1. First-order (Strict) Stationarity
The time series process Xt is completely (strictly)
stationary, if its joint pdf is not affected by a time
translation.
i.e., at whatever point in time it starts, a sample of
T successive observations will have the same pdf
for all t :
Time invariant probabilistic structure.
The process is in ‘stochastic equilibrium’.
Vijayamohan: CDS MPhil: Time Series 1 26 15 October 2012
13
10/15/2012
1. First-order (Strict) Stationarity
Formally:
The process {X(t)} is strictly stationary
if, for any admissible t1, t2, …, tn, and
any interval k,
the joint pdf of {X(t1), X(t2), …, X(tn)} is
identical with the joint pdf of
{X(t1+k), X(t2+k), …, X(tn+k)}.
Vijayamohan: CDS MPhil: Time Series 1 27 15 October 2012
2. Second-order (Weak) Stationarity
A process Yt is weakly stationary if :
1. E(Yt) = µ<∞ (constant mean)
2. Cov(Yt, Yt-k) = γk < ∞ (depends only on k not on t)
For k = 0, the second condition implies
Var (Yt ) = σ 2 a constant.
i.e., the first and second order moment structure of Yt is
constant over time.
Vijayamohan: CDS MPhil: Time Series 1 28 15 October 2012
14
10/15/2012
2. Second-order (Weak) Stationarity
i.e., the first and second order moment structure of Yt
is constant over time.
‘Wide-sense stationarity’ or ‘covariance stationarity’:
No constant mean condition is too restrictive for
most economic time series (having clear trends).
We restrict our attention to weak stationarity and
use `stationarity' to mean weak stationarity.
Vijayamohan: CDS MPhil: Time Series 1 29 15 October 2012
In
yt = a + b yt–1 + ut ;
ut : White noise
But yt may NOT be stationary.
Depends on the magnitude of
b, the root of DE.
Vijayamohan: CDS MPhil: Time Series 1 30 15 October 2012
15
10/15/2012
Consider a DE: yt = b yt–1 ;
Its solution, time path:
yt = y0 bt ;
Nature of the time path yt,
whether it converges or not as t → ∞,
depends on the sign and magnitude of b.
Consider the following cases, with y0 = 1:
yt = bt .
Vijayamohan: CDS MPhil: Time Series 1 31 15 October 2012
Value Time path of yt
of b bt t=0 t=1 t=2 t=3 t=4 t=5 t=6
b>1 2t 1 2 4 8 16 32 64
| b |>1
Vijayamohan: CDS MPhil: Time Series 1 32 15 October 2012
16
10/15/2012
Value Time path of yt
of b bt t=0 t=1 t=2 t=3 t=4 t=5 t=6
b=1 1t 1 1 1 1 1 1 1
|b|=1
Vijayamohan: CDS MPhil: Time Series 1 33 15 October 2012
Value Time path of yt
of b bt t=0 t=1 t=2 t=3 t=4 t=5 t=6
b<1 (1/2)t 1 1/2 1/4 1/8 1/16 1/32 1/64
| b |<1
Vijayamohan: CDS MPhil: Time Series 1 34 15 October 2012
17
10/15/2012
Value Time path of yt
of b bt t=0 t=1 t=2 t=3 t=4 t=5 t=6
-1< b < 0 (-1/2)t 1 -1/2 1/4 -1/8 1/16 -1/32 1/64
|b|<1
Vijayamohan: CDS MPhil: Time Series 1 35 15 October 2012
Value Time path of yt
of b bt t=0 t=1 t=2 t=3 t=4 t=5 t=6
b = -1 -1t 1 -1 1 -1 1 -1 1
|b|=1
Vijayamohan: CDS MPhil: Time Series 1 36 15 October 2012
18
10/15/2012
Value Time path of yt
of b bt t=0 t=1 t=2 t=3 t=4 t=5 t=6
b < -1 -2t 1 -2 4 -8 16 -32 64
| b |>1
Vijayamohan: CDS MPhil: Time Series 1 37 15 October 2012
Thus given a first order DE, yt – byt–1 = 0,
Its time path yt may be
ocillatory,
or non-oscillatory,
converging(damped)
or diverging (explosive)
What are the conditions?
Vijayamohan: CDS MPhil: Time Series 1 38 15 October 2012
19
10/15/2012
Time path yt will be
ocillatory, if b < 0, (negative root) and
non-oscillatory, if b > 0 (positive root).
Vijayamohan: CDS MPhil: Time Series 1 39 15 October 2012
Time path yt will be
converging, if 0 < b < 1, or −1 < b < 0,
that is, −1 < b < 1, or | b | < 1,
and
Non-converging, if b ≥ 1, or b ≤ −1,
that is, −1 ≥ b ≥ 1, or | b | ≥ 1.
Vijayamohan: CDS MPhil: Time Series 1 40 15 October 2012
20
10/15/2012
To recap,
Negative root ⇒ Oscillatory
Positive root ⇒ Non-oscillatory
Absolute value of the root < 1 ⇒
Convergence.
Absolute value of the root ≥ 1 ⇒
Non-convergence.
So, with unit root, |b| = 1,
No convergence: yt unstable; non-stationary.
Vijayamohan: CDS MPhil: Time Series 1 41 15 October 2012
Non-stationarity :
With unit root, b = 1, yt = yt–1 + ut ;
Its time path yt = ut + ut−1 + …. = Σui
for i = 1, 2, …, t;
cumulation (from past to the present) of all
random shocks: stochastic trend
That is, the shock persists;
the process is non-stationary:
unit root problem = non-stationarity problem
Vijayamohan: CDS MPhil: Time Series 1 42 15 October 2012
21
10/15/2012
t −1
yt = byt–1 + ut = ∑ b i ut − i
i =0
Thus with unit root , b = 1 :
Σ ui) = 0, but
E(yt) = E(Σ
Σui) = tσu2, for i = 1, 2, .., t,
var(yt) = var(Σ
Σui Σui + k ) = tσu2,
cov(yt yt + k) = E{Σ
both functions of time.
Non-stationarity
Vijayamohan: CDS MPhil: Time Series 1 43 15 October 2012
Stationarity
Time path yt will be
converging if −1 < b < 1, or | b | < 1.
yt stationary
Vijayamohan: CDS MPhil: Time Series 1 44 15 October 2012
22
10/15/2012
Stationarity :
With | b | < 1 in yt = byt–1 + ut ;
1. E(Yt) = 0 ( = a, with an intercept)
2. Var(Yt) = σu2 /(1– b2)
3. Cov(Yt; Yt-k) = bk σu2 /(1– b2):
all constant;
independent of t:
yt : stationary process.
Vijayamohan: CDS MPhil: Time Series 1 45 15 October 2012
Stationary Processes: Some Examples:
A Stationary
Time Series
White noise ut
Vijayamohan: CDS MPhil: Time Series 1 46 15 October 2012
23
10/15/2012
Stationary Processes: Some Examples:
1. White Noise (Purely random process):
Simplest form of a time series.
The white noise process is a zero mean, constant
variance collection of random variables which are
uncorrelated over time.
Vijayamohan: CDS MPhil: Time Series 1 47 15 October 2012
1. White Noise (Purely random process):
Yt is a white noise process if Yt = εt, where:
(a) E(εt) = 0 (zero mean assumption)
(b) Var(εt) = σ2 (constant variance assumption)
±k) = 0 for all k; t; k ≠ 0:
(c) Cov(εt; εt± (independence of
errors assumption): ‘No memory’.
(d) εt, are normally distributed. (This assumption not essential.)
2
-1
-2
-3
0 50 100 150 200 250 300
Vijayamohan: CDS MPhil: Time Series 1 48 15 October 2012
24
10/15/2012
A Stationary
Time Series
AR(1)
yt = 0.6yt–1 +ut
A Stationary
Time Series
MA(1)
yt = ut + 0.9ut–1
Vijayamohan: CDS MPhil: Time Series 1 49 15 October 2012
Stationary processes
Time series: approximated to
Autoregressive (AR) process
Moving Average (MA) process
Combination of AR and MA process – ARMA.
AR(1) process: yt = a + b yt–1 + ut .
MA(1) process: yt = ut + d ut – 1 .
ARMA(1, 1) process:
yt = a + b yt–1 + ut + d ut – 1 .
Vijayamohan: CDS MPhil: Time Series 1 50 15 October 2012
25
10/15/2012
15 October 2012 Vijayamohan: CDS MPhil: 51
Time Series 1
26
10/19/2012
Time Series Econometrics
2
Vijayamohanan Pillai N
Vijayamohan: CDS MPhil: Time Series 2 1 19 October 2012
Autocovariance function (ACV)
Autocorrelation function (ACF)
Partial Autocorrelation function (PACF)
Vijayamohan: CDS MPhil: Time Series 2 2 19 October 2012
1
10/19/2012
Autocovariance function (ACV) :
Cov(Yt; Yt±k) =
γk = E[(Yt - µ)(Yt±k - µ)]; t = 0, 1, 2, ….;
k = 0, 1, 2, ….;
Then γ0 = Var(Yt); and
ρk = γ k / γ 0 :
Autocorrelation function (ACF)
or Correlogram
– correlation coefficient between pairs of
values of {Yt}, separated by an interval of
length k.
Vijayamohan: CDS MPhil: Time Series 2 3 19 October 2012
In general, theoretical ACF is unknown;
hence we estimate
Sample ACF = sample ACVk / sample ACV0 =
T −k
∑ (Yt − Y )(Yt + k − Y )
t =1
T
∑ (Yt − Y )2
t =1
As k ↑, T – k ↓, so the numerator sum ↓;
the denominator remaining the same,
sample ACF ↓, as k ↑.
Vijayamohan: CDS MPhil: Time Series 2 4 19 October 2012
2
10/19/2012
Properties of Autocovariance function:
1. γ0 = Var(Yt) = σ2 ;
2. | γk| ≤ γ0 , ∀ k; (since | ρk| ≤ 1, ∀ k;
3. When Yt is a real-valued process, γk is
an even function:
γ- k = γk, ∀ k;
Vijayamohan: CDS MPhil: Time Series 2 5 19 October 2012
Properties of Autocorrelation Function:
1. ρ0 = 1; (follows from the definition);
2. | ρk| ≤ 1, ∀ k;
3. ρ-k = ρk, ∀ k; when Yt is a real-valued
process.
Thus ACF is symmetric about the origin
(k = 0), and attains max value of unity
at k = 0.
Vijayamohan: CDS MPhil: Time Series 2 6 19 October 2012
3
10/19/2012
For a white noise (stationary) process:
Cov(εt; εt±±k) = 0 for all k; t; k ≠ 0;
(Serially Uncorrelated)
= σ2 , k = 0. (Constant variance)
So the ACF
ρk = 0, for all k; k ≠ 0;
= 1, k = 0.
- This is the basis for test on white noise
process.
Vijayamohan: CDS MPhil: Time Series 2 7 19 October 2012
Tests of Significance on ACF
(i.e., White Noise):
1. Individual autocorrelation coefficient:
H0: sample ρk = 0, k > 0.
If the series is generated by a white noise
process, i.e., Yt ∼ N(0, 1),
then sample ρk ∼ approximately N(0, 1/T);
T = sample size. (Bartlett 1946)
So a 95% confidence interval is:
1
± 1.96
T
Vijayamohan: CDS MPhil: Time Series 2 8 19 October 2012
4
10/19/2012
Tests of Significance on ACF
(i.e., White Noise):
1. Individual autocorrelation coefficient:
H0: sample ρk = 0, k > 0.
95% confidence interval is: 1
± 1.96
T
H0: sample ρk = 0 is rejected, if sample ρk falls
outside this region for any k.
(Bartlett, MS, 1946: ‘On the Theoretical
Specification of Sampling Properties of
Autocorrelated Time Series’, Journal of the
Royal Statistical Society, ser B8, Vol. 27.)
Vijayamohan: CDS MPhil: Time Series 2 9 19 October 2012
Example: (with T = 100)
Lag 1 2 3 4 5 6
r 0.191 0.148 -0.1 0.065 0.021 -0.01
95% confidence interval =
±1.96/10 = (- 0.196, +0.196).
We cannot reject the null of zero
autocorrelation;
all the coefficients are statistically zero.
The series is a white noise process
(statistically).
Vijayamohan: CDS MPhil: Time Series 2 10 19 October 2012
5
10/19/2012
Autocorrelation function (ACF)
or Correlogram
1.96 0.191
0.148
0.065
0.021
-0.01
0
1 2 3 4 5 6
-0.1
-1.96
Vijayamohan: CDS MPhil: Time Series 2 11 19 October 2012
2. Joint Hypothesis:
H0: sample ρ1 = 0, ρ2 = 0, ρ3 = 0, …….
m
(a)Box-Pierce Q statistic (1970): Q=T ∑ ρρ̂ k2
k =1
T = sample size; m = maximum lag.
Q ∼ χ2m under the null
(all m coefficients are zero).
Small sample problem.
Vijayamohan: CDS MPhil: Time Series 2 12 19 October 2012
6
10/19/2012
2. Joint Hypothesis:
H0: sample ρ1 = 0, ρ2 = 0, ρ3 = 0, …….
(b) Ljung-Box Q statistic (1978):
m
ˆk2
ρ
Q* = T (T + 2) ∑
k =1 T −k
Q* ∼ χ2m under the null.
Asymptotically,
Q* is equivalent to Q,
as (T+2) and (T – k) cancel out.
Vijayamohan: CDS MPhil: Time Series 2 13 19 October 2012
Example: (with T = 100)
Lag 1 2 3 4 5 6
r 0.191 0.148 -0.1 0.065 0.021 -0.01
Q = 100 x (0.1912 + 0.1482 + - 0.1012 +
0.06542 + 0.02112 + -0.00912) = 7.339.
Q* = 100 x 102 x (0.1912/99 + 0.1482/98 +
- 0.1012/97 + 0.06542/96 + 0.02112/95 +
-0.00912/94) = 7.622.
χ26 = 12.592.
We cannot reject the joint null also.
The series is a white noise process
14
Vijayamohan: CDS MPhil: Time Series 2
(statistically). 19 October 2012
7
10/19/2012
Sample Correlogram
Autocorrelation coefficients
Lags
White noise process: ACF all zero
Vijayamohan: CDS MPhil: Time Series 2 15 19 October 2012
Sample Correlogram
Autocorrelation coefficients
Lags
Stationary series: AR(1) : ACF fast decreasing
Vijayamohan: CDS MPhil: Time Series 2 16 19 October 2012
8
10/19/2012
Sample Correlogram
Autocorrelation coefficients
Lags
Nonstationary time series: ACF is slowly decreasing
Vijayamohan: CDS MPhil: Time Series 2 17 19 October 2012
Partial Autocorrelation Function (PACF)
Correlation between the current value and
the value k periods ago,
after controlling for observations at
intermediate lags (i.e., all lags < k).
At lag 1, autocorrelation coefficient =
partial autocorrelation coefficient.
(With Yt and Yt-1,
no intermediate lag effects to eliminate.)
Vijayamohan: CDS MPhil: Time Series 2 18 19 October 2012
9
10/19/2012
Partial Autocorrelation Function (PACF)
Analogous to partial correlation coefficient.
Given X1, X2, and X3,
r12 − r13 r23
r12 ⋅3 =
2 2
(1 − r13 )(1 − r23 )
Obtained after removing the linear relation
of X3 with X1 and with X2 (i.e., r13r23) from
r12.
Vijayamohan: CDS MPhil: Time Series 2 19 19 October 2012
Partial Autocorrelation Function (PACF):
||ly, given Yt, Yt-1, and Yt-2,
PAC of Yt with Yt-2, is:
ρ2⋅1 = (ρ2 − ρ12 ) /(1 − ρ12 )
where
ρ1 = γ1 / γ0 and
ρ2 = γ2 / γ0 are AC coefficients.
Vijayamohan: CDS MPhil: Time Series 2 20 19 October 2012
10
10/19/2012
Tests of Significance on PACF (i.e., Stationarity):
Same as for ACF:
Individual partial autocorrelation coefficient:
1
95% confidence interval is: ± 1.96
T
H0: sample PACC = 0 is rejected,
if it falls outside this region for any k.
Vijayamohan: M Phil: Time Series 21 10/19/2012 1:38:00 PM
Partial Autocorrelation Function (PACF):
In a regression, the coefficients are partial
regression coefficients.
So in an AR model, the autoregression
coefficients are the PAC coefficients.
In an AR(1), Yt = Φ1 Yt-1 + εt ,
Φ1 is both AC and PAC coefficients:
PACC1 = Φ1 = ρ1.
Vijayamohan: CDS MPhil: Time Series 2 22 19 October 2012
11
10/19/2012
Partial Autocorrelation Function (PACF):
In an AR(2), Y = Φ1 Yt-1 + Φ2 Yt-2 + ε ,
t t
Φ1 is the first PAC coefficient:
PACC1 = Φ1 = ρ1(1–Φ2); .
Φ2 is the PAC coefficient at lag 2:
PACC2 = Φ2 = (ρ2 − ρ12 ) /(1 − ρ12 )
In an AR(p), number of PAC coefficients = p:
This property is used for identifying (order
of) an AR model.
Vijayamohan: CDS MPhil: Time Series 2 23 19 October 2012
Partial Autocorrelation Function (PACF)
of a white noise process
Partial Autocorrelation coefficients
All PACF zero
Lags
Vijayamohan: CDS MPhil: Time Series 2 24 19 October 2012
12
10/19/2012
Partial Autocorrelation Function (PACF)
of a stationary process: AR(1)
Partial Autocorrelation coefficients
One significant spike
Lags
Vijayamohan: CDS MPhil: Time Series 2 25 19 October 2012
Partial Autocorrelation Function (PACF)
of a stationary process: MA(1)
Partial Autocorrelation coefficients
PACF fast decreasing
Lags
Vijayamohan: CDS MPhil: Time Series 2 26 19 October 2012
13
10/19/2012
Partial Autocorrelation Function (PACF)
of a non-stationary process
Partial Autocorrelation coefficients
One significant spike
First differencing needed
Lags
Vijayamohan: CDS MPhil: Time Series 2 27 19 October 2012
Stationary processes
Time series: approximated to
Autoregressive (AR) process
Moving Average (MA) process
Combination of AR and MA process – ARMA.
AR(1) process: yt = + yt–1 + t .
MA(1) process: yt = t +d t–1 .
ARMA(1, 1) process:
yt = + yt–1 + t +d t–1 .
Vijayamohan: CDS MPhil: Time Series 2 28 19 October 2012
14
10/19/2012
First order
Autorregresive process
AR(1):
Vijayamohan: CDS MPhil: Time Series 2 29 19 October 2012
Consider a (linear) filter –
Yt = ∑Φ
k
k Xt −k
The system may be taken as
(1) a regression in Yt and Xt;
(2) An AR in Yt , if Xt = Yt;
(3) An MA in Yt , if Xt is white noise.
Vijayamohan: CDS MPhil: Time Series 2 30 19 October 2012
15
10/19/2012
In the (linear) filter – Yt = ∑Φ
k
k Xt −k
= response in the output at time t to a unit
k
pulse in the input at time t – k or
= response in the output at time t+k to a unit
k
pulse in the input at time t: ∂Y t + k ∂Y t
=
∂X t ∂X t −k
one-period multiplier or transient response or
impulse response: ∂Yt + k
∂X t
The impact multiplier is: ∂ t
Y
∂X t
Vijayamohan: CDS MPhil: Time Series 2 31 19 October 2012
The first order Autorregresive process AR(1):
Yt = ΦYt-1 + εt ; εt white noise process.
Yt is a first order stochastic difference equation.
‘autoregressive process’ : Yt has a regression
on its own past values.
In AR(1), only one period dependence:
one period memory: Markov chain.
By repeated substitutions, we get
t −1
Yt = ΦYt-1 + εt = ∑ Φ k ε t − k
k =0
Vijayamohan: CDS MPhil: Time Series 3 32 19 October 2012
16
10/19/2012
The first order Autorregresive process AR(1):
t −1
Yt = ΦYt-1 + εt = ∑ Φ k ε t − k
k =0
∂Yt + k ∂Yt
The impulse responses are: =
∂ε t ∂ε t − k
= Φk, k = 0, 1, 2, ….. : infinite.
Vijayamohan: CDS MPhil: Time Series 3 33 19 October 2012
The first order Autorregresive process AR(1):
t −1
Yt = ΦYt-1 + εt = ∑ Φ k ε t − k
k =0
This process is stationary if and only if |Φ| < 1.
Then,
1. E(Yt) = 0 (zero mean)
2. Var(Yt) = σ2 /(1–Φ2) (constant variance)
3. Cov(Yt; Yt-k) = γk = Φk σ2 /(1–Φ2)
(depends only on k, not on t)
Vijayamohan: CDS MPhil: Time Series 3 34 19 October 2012
17
10/19/2012
The first order Autorregresive process AR(1):
3. Cov(Yt; Yt+k) = γk = Φk σ2 /(1–Φ2) (depends
only on k, not on t)
4. ACF = ρk = γk / γ0 = Φk, k = 0, 1, 2, …..
Impulse responses
ACF ↓, as k ↑, since |Φ| < 1: Sign of stationarity.
+Φ: direct convergence;
negative Φ: oscillatory convergence.
5. PACF = ρ1 = Φ; only one coefficient:
Sign of an AR(1) process.
Vijayamohan: CDS MPhil: Time Series 3 35 19 October 2012
The general AR(1) process is specified as
Yt = + ΦYt-1 + εt =
t −1 t −1
α ∑Φ k + ∑Φ k
ε t −k
k =0 k =0
For stationary series, if |Φ| < 1,
E (Yt) = / (1 – Φ), constant;
Variance and covariances are constant,
exactly the same as before.
Vijayamohan: CDS MPhil: Time Series 3 36 19 October 2012
18
10/19/2012
A Stationary
Time Series
AR(1)
yt = 0.6yt–1 +ut
A Stationary
Time Series
AR(1)
yt = 1.3 + 0.6yt–1 +ut
Vijayamohan: CDS MPhil: Time Series 3 37 19 October 2012
Autocorrelation Function (ACF) of a
stationary process: AR(1)
yt = 0.6yt–1 +ut
Autocorrelation coefficients
Lags
Stationary series: AR(1) : ACF fast decreasing
Vijayamohan: CDS MPhil: Time Series 3 38 19 October 2012
19
10/19/2012
Autocorrelation Function (ACF) of a
stationary process: AR(1)
Autocorrelation coefficients yt = 1.3 +0.6yt–1 +ut
Lags
Stationary series: AR(1) : ACF fast decreasing
Vijayamohan: CDS MPhil: Time Series 3 39 19 October 2012
Partial Autocorrelation Function (PACF)
of a stationary process: AR(1)
yt = 0.6yt–1 +ut
Partial Autocorrelation coefficients
One significant spike
Lags
Vijayamohan: CDS MPhil: Time Series 3 40 19 October 2012
20
10/19/2012
Partial Autocorrelation Function (PACF)
of a stationary process: AR(1)
Partial Autocorrelation coefficients
yt = 1.3 +0.6yt–1 +ut
One significant spike
Lags
Vijayamohan: CDS MPhil: Time Series 3 41 19 October 2012
The first order Autorregresive process AR(1):
If Φ = 1, unit root case, random walk:
t −1
Yt = Yt-1 + εt = ∑ ε t − k
k =0
An accumulation of past shocks.
Shocks persist.
Yt is non-stationary. Then …….
Vijayamohan: CDS MPhil: Time Series 3 42 19 October 2012
21
10/19/2012
The first order Autorregresive process AR(1):
If Φ = 1, unit root case, random walk:
t −1
Yt = Yt-1 + εt = ∑ ε t − k
k =0
Then
1. E(Yt) = 0.
2. Var(Yt) = σ2t (function of t).
3. Cov(Yt; Yt+k) = γk = σ2t (function of t).
Vijayamohan: CDS MPhil: Time Series 3 43 19 October 2012
t −1
−
Random walk: Yt = Yt-1 + εt = ∑ εt −k
k =0
3. Cov(Yt; Yt+k) = γk = σ2t (function of t).
4. ACF: ρk = γk / γ0 = 1: Non-decaying
(slow decay in sample data):
Sign of non-stationarity.
5. PACF: Φ = 1. One coefficient:
Sign of order of integration
(order of differecning to make it stationary).
First difference: ∆ Yt = Yt – Yt-1 = εt :
Stationary.
Vijayamohan: CDS MPhil: Time Series 3 44 19 October 2012
22
10/19/2012
Eight random walks starting at 0:
yt = yt–1 + ut
Vijayamohan: CDS MPhil: Time Series 3 45 19 October 2012
The general AR(1) process is specified as
t −1 t −1
Yt = + ΦYt-1 + εt = α ∑ Φ k + ∑Φ k
ε t −k
k =0 k =0
If Φ = 1, unit root case, t −1
αt + ∑ε t−k
Yt = + Yt-1 + εt = k =0
E (Yt) = t, function of time;
Variance and covariances are functions of
time, exactly the same as before.
Vijayamohan: CDS MPhil: Time Series 3 46 19 October 2012
23
10/19/2012
A Random walk
(without drift)
yt = yt–1 +ut
A Random walk
with drift
yt = 0.3 +yt–1 +ut
Vijayamohan: CDS MPhil: Time Series 3 47 19 October 2012
Autocorrelation Function (ACF) of a
random walk
yt = yt–1 +ut
Vijayamohan: CDS MPhil: Time Series 3 48 19 October 2012
24
10/19/2012
Partial Autocorrelation Function (PACF)
of a random walk
yt = yt–1 +ut
Vijayamohan: CDS MPhil: Time Series 3 49 19 October 2012
Autocorrelation Function (ACF) of a
random walk with drift
yt = 0.3 +yt–1 +ut
Vijayamohan: CDS MPhil: Time Series 3 50 19 October 2012
25
10/19/2012
Partial Autocorrelation Function (PACF)
of a random walk with drift
yt = 0.3 + 0.6yt–1 +ut
Vijayamohan: CDS MPhil: Time Series 3 51 19 October 2012
Second order
Autorregresive process
AR(2)
Vijayamohan: CDS MPhil: Time Series 3 52 19 October 2012
26
10/19/2012
The second order autorregresive process
AR(2):
Yt = Φ1Yt-1 + Φ2Yt-2 + εt
where εt is a white noise process.
Yt is a second order stochastic diff equation.
Vijayamohan: CDS MPhil: Time Series 3 53 19 October 2012
The second order autorregresive process
AR(2):
Yt = Φ1Yt-1 + Φ2Yt-2 + εt
E(εt) = 0
E(εtεt-k) = E(εt Yt-k) = 0, k>0
= E(εt 2) = σ2, k = 0.
Vijayamohan: CDS MPhil: Time Series 3 54 19 October 2012
27
10/19/2012
The second order autorregresive process AR(2):
Yt = Φ1Yt-1 + Φ2Yt-2 + εt
With the lag operator L,
(1 - Φ1L - Φ2L2) Yt = εt ,
where Lyt = Yt-1; ….., Lkyt = Yt-k.
Or Φ(L)Yt = εt ,
Where Φ(L) = (1 - Φ1L - Φ2L2)
a lag polynomial of order 2.
Vijayamohan: CDS MPhil: Time Series 3 55 19 October 2012
The second order autorregresive process AR(2)
Stationarity conditions are obtained
from the expressions of moments.
1. Mean of AR(2):
E(Yt) = E{(1 - Φ1L - Φ2L2)Yt} =
(1 - Φ1 - Φ2)E(Yt) = E(εt) = 0.
With a constant ,
E(Yt) = /(1 - Φ1- Φ2); (Φ1+Φ2) < 1.
One stationarity condition: (Φ1+Φ2) < 1.
Vijayamohan: CDS MPhil: Time Series 3 56 19 October 2012
28
10/19/2012
The second order autorregresive process
AR(2):
2. Second moments:
Yt = Φ1Yt-1 + Φ2Yt-2 + εt ,
multiply by Yt-k and take expectations to get the
Yule-Walker equations:
E(Yt Yt-k) = E(Φ1Yt-1 Yt-k) + E(Φ2Yt-2 Yt-k) +
E(εt Yt-k) ….(1)
Vijayamohan: CDS MPhil: Time Series 3 57 19 October 2012
The second order autorregresive process
AR(2):
2. Second moments:
E(Yt Yt-k) = E(Φ1Yt-1 Yt-k) + E(Φ2Yt-2 Yt-k) +
E(εt Yt-k) ….(1)
E(εt Yt-k) = 0, k ≥ 1;
= E(εt Yt) = E(εt 2) = σ2, k = 0.
Now from (1):
γk = Φ1 γk-1 + Φ2 γk-2 ; k = 1, 2,…
Vijayamohan: CDS MPhil: Time Series 3 58 19 October 2012
29
10/19/2012
The second order autorregresive process AR(2):
(1)Cov(Yt Yt-k) = γk
= Φ1 γk-1 + Φ2 γk-2 ; k = 1,2,..
∴ (2) ACF: ρk
= Φ1 ρk-1 + Φ2 ρk-2 ; k = 1, 2, ….
A difference equation of order 2.
Thus all ρk calculated recursively:
Vijayamohan: CDS MPhil: Time Series 3 59 19 October 2012
The second order autorregresive process AR(2):
ACF: ρk = Φ1 ρk-1 + Φ2 ρk-2 ; k = 1, 2, ….
Thus all ρk calculated recursively:
k = 1, ρ1 = Φ1ρ0 + Φ2ρ-1 = Φ1ρ0 + Φ2ρ1
= Φ1/(1–Φ2); |Φ2| < 1.
k = 2, ρ2 = Φ1 ρ1 + Φ2 ρ0
= Φ12/(1–Φ2) + Φ2; |Φ2| < 1.
||ly for other AC coefficients, ρk , that show
geometric decay.
Thus
Vijayamohan: CDS MPhil: Time Series 3 another 60
stationarity condition: |Φ 2| < 1.
19 October 2012
30
10/19/2012
The second order autorregresive process AR(2):
(3) Var(Yt) = γ0 = Φ1 γ-1 + Φ2 γ-2 + σ2 =
Φ1 γ1 + Φ2 γ2 + σ2
Multiply through by γ0 /γ0 and rewrite:
γ0(1 – Φ1 ρ 1 – Φ2 ρ 2) = σ2;
Or γ0 = σ2/ (1 – Φ1 ρ 1 – Φ2 ρ 2);
and substitute:
ρ1 = Φ1/(1–Φ2); and ρ2 = Φ12/(1–Φ2) + Φ2.
Thus Var(Yt) = γ0 = 1 − Φ2 σ2
1 + Φ2 (Φ1 + Φ2 − 1)(Φ2 − Φ1 − 1)
Vijayamohan: CDS MPhil: Time Series 3 61 19 October 2012
The second order autorregresive process
AR(2):
1 − Φ2 σ2
Var(Yt) = γ0 =
1 + Φ2 (Φ1 + Φ2 − 1)(Φ2 − Φ1 − 1)
This gives the stationarity conditions :
1. |Φ2| < 1;
2. |Φ1+Φ2| <1;
3. |Φ2 – Φ1| < 1.
Vijayamohan: CDS MPhil: Time Series 3 62 19 October 2012
31
10/19/2012
The second order autorregresive process AR(2):
(4) PACF: Given
Yt = Φ1 Yt-1 + Φ2 Yt-2 + εt ,
(1) Φ1 is the first PAC (and AC) coefficient:
PACC1 =Φ1= ρ1;
(2) Φ2 is the PAC coefficient at lag 2:
PACC2 = Φ2 =
(ρ2 − ρ12 ) /(1 − ρ12 )
where ρ1 = Φ1/(1–Φ2); ρ2 = Φ12/(1–Φ2) + Φ2.
Vijayamohan: CDS MPhil: Time Series 3 63 19 October 2012
A Stationary Time Series
AR(2)
yt = 0.4yt–1 – 0.7yt–2+ut
Vijayamohan: CDS MPhil: Time Series 3 64 19 October 2012
32
10/19/2012
A Stationary Time Series
AR(2)
yt = 3 + 0.4yt–1 – 0.7yt–2+ut
Vijayamohan: CDS MPhil: Time Series 3 65 19 October 2012
Autocorrelation Function (ACF) of a
stationary process: AR(2)
yt = 0.4yt–1 – 0.7yt–2+ut
Vijayamohan: CDS MPhil: Time Series 3 66 19 October 2012
33
10/19/2012
Partial Autocorrelation Function (PACF) of
a stationary process: AR(2)
yt = 0.4yt–1 – 0.7yt–2+ut
Vijayamohan: CDS MPhil: Time Series 3 67 19 October 2012
Φ1, Φ2 > 0 Φ1 < 0; Φ2 > 0
ACF and PACF patterns for real roots AR(2)
Vijayamohan: CDS MPhil: Time Series 3 68 19 October 2012
34
10/19/2012
Is the following process stationary?
Yt – 2Yt-1 + Yt-2 = εt..
To find the roots of the AR(2) process:
(1)Rewrite using the complementary solution:
Yt = Art:
Art – 2 Art-1 + Art-2 = 0.
Dividing thru by Art-2,
the ‘characteristic equation’
r2 – 2r + 1 = 0,
gives the two roots: r1, r2 = 1: unit root.
But differences in interpretation.
Vijayamohan: CDS MPhil: Time Series 3 69 19 October 2012
Is the following process stationary?
Yt – 2Yt-1 + Yt-2 = εt..
To find the roots of the AR(2) process:
(2) Also try with the lag polynomial:
(1 - 2L + L2) Yt = εt .
The ‘characteristic equation’
1 - 2L + L2 = 0,
gives the two roots: L1, L2 = 1.
But differences in interpretation.
Vijayamohan: CDS MPhil: Time Series 3 70 19 October 2012
35
10/19/2012
Yt = 2Yt-1 - Yt-2 + εt..
Vijayamohan: CDS MPhil: Time Series 3 71 19 October 2012
ACF of
Yt = 2Yt-1 - Yt-2 + εt..
Vijayamohan: CDS MPhil: Time Series 3 72 19 October 2012
36
10/19/2012
PACF of
Yt = 2Yt-1 - Yt-2 + εt..
Vijayamohan: CDS MPhil: Time Series 3 73 19 October 2012
First difference of Yt = 2Yt-1 - Yt-2 + εt..
Vijayamohan: CDS MPhil: Time Series 3 74 19 October 2012
37
10/19/2012
ACF and PACF of
First difference of Yt = 2Yt-1 - Yt-2 + εt..
Vijayamohan: CDS MPhil: Time Series 3 75 19 October 2012
Second difference of Yt = 2Yt-1 - Yt-2 + εt..
Vijayamohan: CDS MPhil: Time Series 3 76 19 October 2012
38
10/19/2012
ACF and PACF of
Second difference of Yt = 2Yt-1 - Yt-2 + εt..
Vijayamohan: CDS MPhil: Time Series 3 77 19 October 2012
Differences in interpretation will be clear with
this example:
Try Yt + 0.3Yt-1 – 0.4Yt-2 = εt..
(1) Using the complementary solution:
r2 + 0.3r – 0.4 = 0,
giving the two roots:
r1 = – 0.8; r2 = 0.5.
The DE roots lie within the unit circle:
stationary process.
Vijayamohan: CDS MPhil: Time Series 3 78 19 October 2012
39
10/19/2012
Differences in interpretation will be clear with
this example:
Try Yt + 0.3Yt-1 – 0.4Yt-2 = εt..
(2) With the lag polynomial:
(1 + 0.3L – 0.4 L2) Yt = εt .
1 + 0.3L – 0.4 L2 = 0,
giving the two roots:
L1 = -1/0.8 = –1.25; L2 = 1/0.5 = 2.
The lag polynomial roots lie outside the unit
circle: stationary process.
Vijayamohan: CDS MPhil: Time Series 3 79 19 October 2012
Yt + 0.3Yt-1 – 0.4Yt-2 = εt..
Vijayamohan: CDS MPhil: Time Series 3 80 19 October 2012
40
10/19/2012
ACF and PACF of
Yt + 0.3Yt-1 – 0.4Yt-2 = εt..
Vijayamohan: CDS MPhil: Time Series 3 81 19 October 2012
Remember
the Stationarity conditions for an AR(2):
(1) |Φ2| < 1;
(2) |Φ1+Φ2| < 1;
(3) |Φ2 – Φ1| < 1.
Now try these with the above example:
That is, for Yt + 0.3Yt-1 – 0.4Yt-2 = εt..
Vijayamohan: CDS MPhil: Time Series 3 82 19 October 2012
41
10/19/2012
The Moving Average
process
MA(q)
Vijayamohan: CDS MPhil: Time Series 3 83 19 October 2012
19 October 2012 Vijayamohan: CDS MPhil: Time 84
Series 2
42
10/15/2012
Time Series Econometrics
3
Vijayamohanan Pillai N
Vijayamohan: CDS MPhil: Time Series 4 1 15 October 2012
The Moving Average
process
MA(q)
Vijayamohan: CDS MPhil: Time Series 4 2 15 October 2012
1
10/15/2012
The (linear) filter –
Yt = ∑Φ
k
k Xt −k
may be taken as
(1) a regression in Yt and Xt;
(2) An AR in Yt , if Xt = Yt;
(3) An MA in Yt , if Xt is white noise.
Vijayamohan: CDS MPhil: Time Series 2 3 15 October 2012
The Moving Average process MA(q):
q
Yt = ∑ θkεt − k
k =0
the output of a filter with
a white noise input and
transient response θk.
In general, θ0 = 1.
Vijayamohan: CDS MPhil: Time Series 4 4 15 October 2012
2
10/15/2012
The first order Moving Average process MA(1):
Yt = εt – θ1εt-1; or with lag operator:
Yt = (1 – θ1L)εt .
See: there is
only one period impulse response,
unlike in AR(1);
MA process is a
finite impulse response filter.
Vijayamohan: CDS MPhil: Time Series 4 5 15 October 2012
The first order Moving Average process MA(1):
Yt = εt – θ1εt-1; or with lag operator:
Yt = (1 – θ1L)εt .
1. E(Yt) = E(εt ) = 0;
With a constant , E(Yt) = .
2. Var (Yt) = Var (εt – θ1εt-1) =
σ2(1 + θ12), constant.
Vijayamohan: CDS MPhil: Time Series 4 6 15 October 2012
3
10/15/2012
The first order Moving Average process MA(1):
3. Cov(Yt; Yt-k) = γk :
with k = 1:
γ1 = E(Yt Yt-1) =
E{(εt – θ1εt-1)(εt-1 – θ1εt-2)} =
–θ1 σ2.
Vijayamohan: CDS MPhil: Time Series 4 7 15 October 2012
The first order Moving Average process MA(1):
3. Cov(Yt; Yt-k) = γk :
γ1 = –θ1 σ2.
With k = 2:
γ2 = E(Yt Yt-2) =
E{(εt – θ1εt-1)(εt-2 – θ1εt-3)} = 0.
Thus Cov(Yt; Yt-k) = γk = 0, ∀ k > 1.
Vijayamohan: CDS MPhil: Time Series 4 8 15 October 2012
4
10/15/2012
The first order Moving Average process MA(1):
Yt has zero mean and
constant variance; but
Cov(Yt; Yt-k) = –θ1 σ2 ≠ 0, for k = 1.
Note: Though εt is white noise,
Yt is not, if the θks differ from zero.
Vijayamohan: CDS MPhil: Time Series 4 9 15 October 2012
The first order Moving Average process MA(1):
4. ACF = ρk = γk / γ0
= –θ1 / (1 + θ12), k = 1;
= 0, k > 1.
The process has
only one period impulse response (memory).
Vijayamohan: CDS MPhil: Time Series 4 10 15 October 2012
5
10/15/2012
The first order Moving Average process MA(1):
5. PACF: Applicable with Yt and its lagged
values.
So MA process to be expressed as an AR one.
As long as MA process is invertible,
∞) and vice versa.
it can be expressed as an AR(∞
So we come to Invertibility Condition:
from Wold’s Decomposition Theorem:
Vijayamohan: CDS MPhil: Time Series 4 11 15 October 2012
Wold’s Decomposition Theorem:
“Every weakly stationary,
purely non-deterministic,
stochastic process (Yt – µ)
can be written as
a linear filter (linear combination)
of a sequence of uncorrelated random
variables.”
(Wold 1938).
Vijayamohan: CDS MPhil: Time Series 4 12 15 October 2012
6
10/15/2012
Wold’s Decomposition Theorem:
“Every weakly stationary, purely non-
deterministic, stochastic process (Yt – µ) can be
written as a linear filter (linear combination) of
a sequence of uncorrelated random variables.”
(Wold 1938).
‘Purely non-deterministic’
= purely non-predictable
= all linearly deterministic components,
say, a (constant) mean, if any,
have been subtracted from Yt:
(Yt – µ)
Vijayamohan: CDS MPhil: Time Series 4 13 15 October 2012
Wold’s Decomposition Theorem:
Such a linear filter is:
∞
Yt – µ = ∑ψ j ε t − j ,
j =0
ψ0 = 1.
εt is white noise, and
ψj s are psi-weights.
Vijayamohan: CDS MPhil: Time Series 4 14 15 October 2012
7
10/15/2012
Wold’s Decomposition Theorem:
∞
Yt – µ = ∑ψ j ε t − j , ψ0 = 1.
j =0
Let µ = 0, without loss of generality, and
(1) Let ψ1 = –θ1, and ψj = 0, for j ≥ 2:
then we get a MA(1) process:
Yt = εt – θ1εt-1.
Vijayamohan: CDS MPhil: Time Series 4 15 15 October 2012
Wold’s Decomposition Theorem:
∞
Yt – µ = ∑ψ j ε t − j , ψ0 = 1.
j =0
Let µ = 0, without loss of generality, and
(2) Let ψj = Φ j, then
Yt = εt + Φεt-1+ Φ2εt-2 + …..
= εt + Φ Yt-1 , an AR(1) process.
⇒ There is a relationship between the two!
Vijayamohan: CDS MPhil: Time Series 4 16 15 October 2012
8
10/15/2012
Relationship between the AR and MA processes:
(1) By inverting AR(1) process,
we get an infinite MA process:
Yt = Φ Yt-1+ εt or (1 – ΦL) Yt = εt,
so that Yt = (1– ΦL)–1εt.
Expanding by Binomial theorem:
Yt = (1+ΦL+ Φ2L2 + ……) εt,
= εt +Φεt-1+ Φ2εt-2 + …..
Vijayamohan: CDS MPhil: Time Series 4 17 15 October 2012
Relationship between the AR and MA processes:
(1)By inverting AR(1) process,
we get an infinite MA process:
Yt = εt +Φεt-1+ Φ2εt-2 + …..
∞) series will converge asymptotically
This MA(∞
if |Φ|< 1:
stationarity condition:
bounds of stationarity.
Moments of AR(1) can be found based on this
∞) series also.
MA(∞
Vijayamohan: CDS MPhil: Time Series 4 18 15 October 2012
9
10/15/2012
Relationship between the AR and MA processes:
(2) By inverting MA(1) process,
we get an infinite AR process:
Yt = εt – θ1εt-1 = (1 – θ1L)εt
so that Yt (1 – θ1L) –1 = εt .
Expanding by Binomial theorem:
Yt (1+θ1L+ θ12L2 + ……) = εt, that is
Yt + θ1 Yt-1 + θ12 Yt-2 + …… = εt.
This AR(∞ ∞) series will converge asymptotically
if |θ1 | < 1: Invertibility condition:
Vijayamohan: CDS MPhil: Time Series 4 bounds
19 of invertibility.
15 October 2012
The first order Moving Average process MA(1):
∞) series,
Given the AR(∞
obtained by inverting the MA(1) series:
Yt + θ1 Yt-1 + θ12 Yt-2 + …… = εt,
We can find the PACF of MA(1):
Yt = εt – θ1εt-1 ⇒
Given Yt = Φ1Yt-1 + Φ2Yt-2 + …….. + εt ,
The first PAC coefficient = Φ1= ρ1 =
–θ1 / (1 + θ12),
Vijayamohan: CDS MPhil: Time Series 4 20 15 October 2012
10
10/15/2012
The first order Moving Average process MA(1):
Given Yt = Φ1Yt-1 + Φ2Yt-2 + …….. + εt ,
The first PAC coefficient = Φ1= ρ1 =
–θ1 / (1 + θ12),
The second PACC = Φ2 = (ρ2 − ρ12 ) /(1 − ρ12 )
= –θ12 / (1 + θ12 + θ14),
where ρ2 = 0 and ρ1 is as above.
Vijayamohan: CDS MPhil: Time Series 4 21 15 October 2012
The first order Moving Average process MA(1):
||ly for other higher order PACF.
If MA(1) shock is positive, ACF, PACF: negative;
If MA(1) shock is negative, ACF, PACF: positive;
In both cases, PACF exponentially declines.
Note: For MA(1), ACF has value zero after the
first ACC, while PACF declines geometrically.
Note: For AR(1), ACF declines geometrically,
while PACF has value zero after the first PACC.
Vijayamohan: CDS MPhil: Time Series 4 22 15 October 2012
11
10/15/2012
MA(1)
yt = ut + 0.6ut–1
Vijayamohan: CDS MPhil: Time Series 4 23 15 October 2012
Theoretical ACF and PACF of MA(1) process
θ1 positive θ1 negative
Vijayamohan: CDS MPhil: Time Series 4 24 15 October 2012
12
10/15/2012
The second order Moving Average process MA(2)
Yt = εt – θ1εt-1 – θ2εt-2;
With lag operator,
Yt = θ(L)εt , where
θ(L) = (1 – θ1L – θ2L2).
Vijayamohan: CDS MPhil: Time Series 4 25 15 October 2012
The second order Moving Average process MA(2):
Yt = εt – θ1εt-1 – θ2εt-2;
1. E(Yt) = E(εt ) = 0;
With a constant , E(Yt) = , as in MA(1).
2. Var (Yt) = Var (εt – θ1εt-1 – θ2εt-2) =
σ2(1 + θ12 + θ22), constant.
Vijayamohan: CDS MPhil: Time Series 4 26 15 October 2012
13
10/15/2012
The second order Moving Average process MA(2):
3. Cov(Yt; Yt-k) = γk :
with k = 1:
γ1 = E(Yt Yt-1) =
E{(εt – θ1εt-1 – θ2εt-2)(εt-1 – θ1εt-2 – θ2εt-3)}
= –θ1 (1 – θ2) σ2.
Vijayamohan: CDS MPhil: Time Series 4 27 15 October 2012
The second order Moving Average process MA(2):
3. Cov(Yt; Yt-k) = γk :
γ1 = –θ1 (1 – θ2) σ2.
With k = 2:
γ2 = E(Yt Yt-2) =
E{(εt – θ1εt-1 – θ2εt-2)(εt-2 – θ1εt-3 – θ2εt-4)}
= – θ2 σ2.
Cov(Yt; Yt-k) = γk = 0, ∀ k > 2.
Vijayamohan: CDS MPhil: Time Series 4 28 15 October 2012
14
10/15/2012
The second order Moving Average process MA(2):
4. ACF = ρk = γk / γ0 =
–θ1 (1 – θ2) / (1 + θ12 + θ22), k = 1;
= – θ2 / (1 + θ12 + θ22), k = 2;
= 0, k > 2.
The process has
two period- impulse response (memory) only.
Vijayamohan: CDS MPhil: Time Series 4 29 15 October 2012
The second order Moving Average process
MA(2):
5. PACF:
The first PAC coefficient = Φ1;
The second PACC = Φ2
|ly for other higher order PACF.
Vijayamohan: CDS MPhil: Time Series 4 30 15 October 2012
15
10/15/2012
The second order Moving Average process MA(2):
If MA(2) shocks are positive, ACF, PACF: negative;
If MA(2) shocks are negative, ACF, PACF: positive;
In both cases, PACF exponentially declines.
Note: For MA(2), ACF has value zero after the
second ACC, while PACF declines geometrically.
Note: For AR(2), ACF declines geometrically,
while PACF has value zero after the second PACC.
Vijayamohan: CDS MPhil: Time Series 4 31 15 October 2012
Yt = εt + 1.1εt-1 + 0.2εt-2
Vijayamohan: CDS MPhil: Time Series 4 32 15 October 2012
16
10/15/2012
Theoretical ACF and PACF of MA(2) process
θ1, θ2 positive θ1 < 0; θ2 > 0
Vijayamohan: CDS MPhil: Time Series 4 33 15 October 2012
The general Moving Average process MA(q):
q
Yt = ∑ θ k ε t − k, θ0 = 1.
k =0
Or Yt = εt – θ1εt-1 – θ2εt-2 – ………. – θqεt-q ;
With lag operator, Yt = θ(L)εt ,
where
θ(L) = (1 – θ1L – θ2L2 – …. – θqLq).
ACF of MA(q) cuts off after lag q:
only q period impulse response.
Vijayamohan: CDS MPhil: Time Series 4 34 15 October 2012
17
10/15/2012
Moving Average process:
MA processes arise in econometrics thru
trend elimination or over-differencing:
(1): Trend elimination thru differencing :
e.g., Yt = a + bt +εt; then
Yt-1 = a + b(t–1) +εt-1;
⇒ ∆Yt = Yt –Yt-1 = b + εt – εt-1.
∆Yt has a MA unit root!
Vijayamohan: CDS MPhil: Time Series 4 35 15 October 2012
Moving Average process:
⇒ ∆Yt = Yt –Yt-1 = b + εt – εt-1.
∆Yt has a MA unit root!
Differencing detrends, but generates MA
process that can show a cycle when there is
none in the original series:
‘spurious cycle’:
Slutzky effect (Slutzky 1937).
Vijayamohan: CDS MPhil: Time Series 4 36 15 October 2012
18
10/15/2012
Yt = 2 + 5t +εt;
Vijayamohan: CDS MPhil: Time Series 4 37 15 October 2012
Differenced series: ∆Yt = Yt –Yt-1 = 5 + εt – εt-1
Detrended series: Yt – (2 + 0.05t) = εt
Vijayamohan: CDS MPhil: Time Series 4 38 15 October 2012
19
10/15/2012
Moving Average process:
(2) Over-differencing:
Differencing a stationary (white noise) series,
say, Yt = εt , generates MA process:
∆Yt = Yt –Yt-1 = εt – εt-1.
⇒ ∆Yt has a MA unit root!
If | θ | = 1, as above,
MA process is non-invertible.
Vijayamohan: CDS MPhil: Time Series 4 39 15 October 2012
Autoregressive
Moving Average
Process:
ARMA(p, q)
Vijayamohan: CDS MPhil: Time Series 5 40 15 October 2012
20
10/15/2012
15 October 2012 Vijayamohan: CDS MPhil: Time 41
Series 4
21
10/15/2012
Time Series Econometrics
4
Vijayamohanan Pillai N
Vijayamohan: CDS MPhil: Time Series 5 1 15 October 2012
Autoregressive
Moving Average
Process:
ARMA(p, q)
Vijayamohan: CDS MPhil: Time Series 5 2 15 October 2012
1
10/15/2012
Autoregressive Moving Average Process:
ARMA(p, q)
The general ARMA(p, q):
AR of order p and
MA of order q :
Yt = Φ1Yt-1 + Φ2Yt-2 + …. + ΦpYt-p
+ εt – θ1εt-1 – θ2εt-2 – …. – θqεt-q ,
where εt is a white noise process.
Vijayamohan: CDS MPhil: Time Series 5 3 15 October 2012
Autoregressive Moving Average Process: ARMA(p, q)
Yt = Φ1Yt-1 + Φ2Yt-2 + …. + ΦpYt-p +
εt – θ1εt-1 – θ2εt-2 – …. – θqεt-q ,
With the lag operator:
Φ(L)Yt = θ (L)εt , where
Φ(L) = (1 – Φ1L – Φ2L2 –…. – ΦpLp) ; and
θ(L) = (1 – θ1L – θ2L2 –…. – θqLq ).
Properties of ARMA process:
a mixture of those of AR and MA processes.
Vijayamohan: CDS MPhil: Time Series 5 4 15 October 2012
2
10/15/2012
The simplest ARMA process: ARMA(1, 1):
Yt = Φ1Yt-1 + εt – θ1εt-1.
(1)E(Yt) = 0;
With a constant ,
E(Yt) = /(1 - Φ1); |Φ1| < 1.
Vijayamohan: CDS MPhil: Time Series 5 5 15 October 2012
ARMA(1, 1)
(2) Var(Yt) = E(Φ1Yt-1 + εt – θ1εt-1)2:
= Φ12Var(Yt) –2Φ1 θ1E(Yt-1 εt-1) + Var(εt) +
θ12Var(εt-1):
Since Var(εt) = Var(εt-1) = σ2;
and E(Yt-1 εt-1) = E(Yt εt) = σ2,
Var(Yt) = σ2(1 + θ12 –2Φ1 θ1)/ (1 – Φ12),
|Φ1| < 1.
Vijayamohan: CDS MPhil: Time Series 5 6 15 October 2012
3
10/15/2012
ARMA(1, 1)
(3) Cov(Yt; Yt-k) = γk :
The covariances are recursively estimated:
with k = 1:
γ1 = E(Yt Yt-1)
= E{(Φ1Yt-1 + εt – θ1εt-1) Yt-1}
= Φ1Var(Yt) –θ1 σ2 = Φ1 γ0 –θ1 σ2
= σ2(1 – Φ1 θ1)(Φ1 –θ1)/ (1 – Φ12),
|Φ1| < 1.
Vijayamohan: CDS MPhil: Time Series 5 7 15 October 2012
ARMA(1, 1)
(3) Cov(Yt; Yt-k) = γk :
with k = 2:
γ2 = E(Yt Yt-2)
= E{(Φ1Yt-1 + εt – θ1εt-1) Yt-2}
= Φ1 E(Yt-1 Yt-2) = Φ1 E(Yt Yt-1)
= Φ1 γ1.
Thus γk = Φ1 γk-1, k ≥ 2.
Vijayamohan: CDS MPhil: Time Series 5 8 15 October 2012
4
10/15/2012
ARMA(1, 1)
(4) ACF:
ρk = γ k / γ 0
= (1 – Φ1 θ1)(Φ1 –θ1)/(1 + θ12 –2Φ1 θ1), k = 1;
= Φ1 ρk-1, k ≥ 2.
Thus ACF decays geometrically
from the starting value, ρ1,
a function of both Φ1 and θ1.
Vijayamohan: CDS MPhil: Time Series 5 9 15 October 2012
ARMA(1, 1)
(4) ACF: = ρk = γk / γ0
= Φ1 ρk-1, k ≥ 2.
Thus ACF decays geometrically from the starting
value, ρ1, a function of both Φ1 and θ1.
Note: The decay is determined by Φ1 only:
the MA part has only one period memory.
Note: When Φ1 = θ1, ρk = 0, as Yt collapses to a
white noise process: Yt = εt +θ1 (Yt-1 – εt-1).
Vijayamohan: CDS MPhil: Time Series 5 10 15 October 2012
5
10/15/2012
ARMA(1, 1)
5. PACF:
The first PAC coefficient = Φ1;
The second PACC = Φ2
||ly for other higher order PACF.
PACF generally declines as the lag increases.
Vijayamohan: CDS MPhil: Time Series 5 11 15 October 2012
Yt = 0.4Yt-1 + εt – 0.2εt-1.
Vijayamohan: CDS MPhil: Time Series 5 12 15 October 2012
6
10/15/2012
Yt = 0.4Yt-1 + εt – 0.2εt-1.
Vijayamohan: CDS MPhil: Time Series 5 13 15 October 2012
Yt = – 0.4Yt-1 – 0.5Yt-2+ εt – 0.2εt-1 –0.1 εt-2.
Vijayamohan: CDS MPhil: Time Series 5 14 15 October 2012
7
10/15/2012
Yt = – 0.4Yt-1 – 0.5Yt-2+ εt – 0.2εt-1 –0.1 εt-2.
Vijayamohan: CDS MPhil: Time Series 5 15 15 October 2012
Theoretical ACF and PACF of ARMA(1, 1) process
Φ1 positive Φ1 negative
Vijayamohan: CDS MPhil: Time Series 5 16 15 October 2012
8
10/15/2012
Non-stationarity
Non-stationarity due to
(1)Integrated process: random walk:
Yt = Yt-1 + εt
Difference stationary process
Stochastic trend
(2) Trend: Trend stationary process
Both stochastic and deterministic trend
Vijayamohan: CDS MPhil: Time Series 5 17 15 October 2012
Consequences of Non-stationarity
(1) With stationary series, possible to model the
process via a fixed-coefficients equation
estimated from past data.
Not possible if the structural relationship
changes over time (if non-stationary).
(2) For non-stationary series, Var and Covs are
functions of time:
so the conventional asymptotic theory cannot
be applied to these series.
Vijayamohan: CDS MPhil: Time Series 5 18 15 October 2012
9
10/15/2012
Consequences of Non-stationarity
(2) For non-stationary series, Var and Covs are
functions of time: so the conventional
asymptotic theory cannot be applied to these
series.
For example: In a regression of Yt on Xt:
βˆ = Cov (Yt , X t ) Var ( X t )
If Xt is non-stationary, Var(Xt) ↑ infinitely, and
dominates the Cov(Yt, Xt):
Then the OLS estimator does not have an
asymptotic distribution.
Vijayamohan: CDS MPhil: Time Series 5 19 15 October 2012
Consequence of unit root (non-stationarity):
Consider two unrelated RW processes:
Yt = Yt–1 + Ut; Ut ∼ IIN(0, σU2)
Xt = Xt–1 + Vt; Vt ∼ IIN(0, σV2);
Cov(Ut , Vt) = 0.
Vijayamohan: CDS MPhil: Time Series 5 20 15 October 2012
10
10/15/2012
Consequence of unit root (non-stationarity):
y x
20
15
10
-5
-10
0 50 100 150 200 250 300
Yt = Yt–1 + Ut; Ut ∼ IIN(0, σU2)
Xt = Xt–1 + Vt; Vt ∼ IIN(0, σV2);
Cov(Ut , Vt) = 0.
Vijayamohan: CDS MPhil: Time Series 5 21 15 October 2012
Consequence of unit root (non-stationarity):
3 eps ep
-1
-2
-3
0 50 100 150 200 250 300
Ut ∼ IIN(0, σU2)
Vt ∼ IIN(0, σV2)
Vijayamohan: CDS MPhil: Time Series 5 22 15 October 2012
11
10/15/2012
Consequence of unit root (non-stationarity):
Cross Correlation function
1.00
CCF-eps x ep CCF-ep x eps
0.75
0.50
0.25
0.00
-0.25
-0.50
-0.75
0 5 10
Cov(Ut , Vt) = 0.
Vijayamohan: CDS MPhil: Time Series 5 23 15 October 2012
Consequence of unit root (non-stationarity):
Now consider the regression:
Yt =β 0 +β 1 Xt + ε t.
We expect R2 from this regression would tend
to zero; BUT….
Vijayamohan: CDS MPhil: Time Series 5 24 15 October 2012
12
10/15/2012
Consequence of unit root (non-stationarity):
Vijayamohan: CDS MPhil: Time Series 5 25 15 October 2012
Consequence of unit root (non-stationarity):
Vijayamohan: CDS MPhil: Time Series 5 26 15 October 2012
13
10/15/2012
Consequence of unit root (non-stationarity):
Granger and Newbold (1974):
High R2 and highly significant t,
but a low DW statistic.
When the regression was run
in first differences……
?
Vijayamohan: CDS MPhil: Time Series 5 27 15 October 2012
Consequence of unit root (non-stationarity):
Vijayamohan: CDS MPhil: Time Series 5 28 15 October 2012
14
10/15/2012
Consequence of unit root (non-stationarity):
When the regression was run
in first differences,
R2 close to zero; DW close to 2: ⇒
No relationship between Yt and Xt ; and
the high R2 obtained was ‘spurious’.
R2 > DW ⇒ ‘Spurious Regression’.
Vijayamohan: CDS MPhil: Time Series 5 29 15 October 2012
Spurious or Nonsense Regression?
Long-standing Puzzle over high correlations
between what ought to be unrelated time series
variables; e.g.,
High positive correlation between the murder rate
and membership of the Church of England:
Yule(1926).
Yule(1926) classifies:
(1): Nonsense regression;
(2): Spurious regression.
Vijayamohan: CDS MPhil: Time Series 5 30 15 October 2012
15
10/15/2012
Spurious or Nonsense Regression?
(1): Nonsense regression:
integrated, but mutually independent, time series;
(high serial correlation in each series
→ high correlation);
(2): Spurious regression:
Variables depending on common third factors
(e.g., having a linear trend).
Vijayamohan: CDS MPhil: Time Series 5 31 15 October 2012
So, check for stationarity of time series:
Classical: Autocorrelation function (ACF):
Correlogram
Fast-decreasing ACF ⇒ Stationarity
Modern: Unit root tests:
(Augmented) Dickey-Fuller test;
Phillips-Perron non-parametric test; etc.
Vijayamohan: CDS MPhil: Time Series 5 32 15 October 2012
16
10/15/2012
If non-stationary series,
run regression in (first) differences as in
Classical (ARIMA) modelling
or check for Cointegration among the series
Vijayamohan: CDS MPhil: Time Series 5 33 15 October 2012
Non-stationarity and differencing:
Consider the random walk:
Yt = Yt-1 + εt where εt is a white noise
First difference: ∆Yt = Yt – Yt-1 = εt :
Stationary.
If a non-stationary series is transformed into a
stationary one
by differencing once,
the series in level =
integrated of order one:
I(1): one unit root;
Vijayamohan: CDS MPhil: Time Series 5 34 15 October 2012
17
10/15/2012
Non-stationarity and differencing:
Consider the random walk:
Yt = Yt-1 + εt where εt is a white noise
First difference:
∆Yt = Yt – Yt-1 = εt :
Stationary.
The series in level =
integrated of order one:
I(1): one unit root;
Vijayamohan: CDS MPhil: Time Series 5 35 15 October 2012
Non-stationarity and differencing:
Process of inverse of differencing =
integration.
With one unit root, first differencing, to
make yt stationary.
Then yt = integrated of order one:
I(1): one unit root;
The differenced, ∆yt , stationary series =
integrated of order zero: I(0); no unit root.
Thus Yt ∼ I(1); and ∆Yt∼ I(0).
Vijayamohan: CDS MPhil: Time Series 5 36 15 October 2012
18
10/15/2012
I(1) versus I(0) processes
(1) I(1) series wanders widely;
I(0) series = mean reverting:
direct/oscillatory convergence to mean.
(2) I(1) series has infinitely long memory of its
past behaviour :
shocks permanently affect the process:
evident from slow decay of ACF.
I(0) series has limited memory:
shock effects are only transitory:
evident from fast decay of ACF.
Vijayamohan: CDS MPhil: Time Series 5 37 15 October 2012
I(1): the wanderer and
I(0): the mean-reverter
I(1): No tendency
to return to zero
time
I(0): Rarely drifts from zero
Vijayamohan: CDS MPhil: Time Series 5 38 15 October 2012
19
10/15/2012
Non-stationarity and differencing:
If a non-stationary series, Yt →
into a stationary one
by differencing d times, ∆d,
the series in level, Yt = integrated of order d:
I(d): d unit roots.
Then ∆d Yt = stationary: I(0).
Vijayamohan: CDS MPhil: Time Series 5 39 15 October 2012
Non-stationarity and differencing:
If ∆d Yt = stationary process
that can be represented
by an ARMA(p, q) model,
then
Yt = an integrated autoregressive moving
average process of order: d, p, and q:
ARIMA(p, d, q),
d = number of differences = order of integration.
Vijayamohan: CDS MPhil: Time Series 5 40 15 October 2012
20
10/15/2012
Non-stationarity
Non-stationarity due to
(1)Integrated process: random walk:
Yt = Yt-1 + εt
Difference stationary process
Stochastic trend
(2) Trend: Trend stationary process
Both stochastic and deterministic trend
Vijayamohan: CDS MPhil: Time Series 5 41 15 October 2012
Non-stationarity :
Trend stationary process: TSP:
The determistic trend :
Yt = a +bt+ ut ,
ut : white noise process.
Stationary fluctuations around a linear trend.
Non-stationary process?
Vijayamohan: CDS MPhil: Time Series 5 42 15 October 2012
21
10/15/2012
Non-stationarity :
Trend stationary process: TSP:
Yt = a +bt+ ut , ut : white noise process.
mean changes with time:
E(Yt) = a +bt.
But constant variance : Var(Yt) =
Var(ut) = σu2.
Cov(Yt , Yt-k) = Cov(ut , ut-k) = 0.
Vijayamohan: CDS MPhil: Time Series 5 43 15 October 2012
Non-stationarity :
To make a TSP stationary, detrend it:
Yt – E(Yt) =
= Yt – (a +bt)
= ut , stationary.
A TSP:
Yt = 1.5 + 0.5t+ ut Detrended Series
Vijayamohan: CDS MPhil: Time Series 5 44 15 October 2012
22
10/15/2012
EQ( 1) Modelling y by OLS
The estimation sample is: 1 to 100
A TSP:
Yt = 1.5 + 0.5t+ ut
Coefficient Std.Error t-value t-prob
Constant 1.39136 0.2023 6.88 0.000
Trend 0.502021 0.003478 144. 0.000
sigma .00398 RSS 98.7822601 R2 0.995318 F(1,98) = 2.083e+004 [0.000]**
log-likelihood -141.281 DW 1.71 no. of observations 100 no. of parameters 2
Fitted series: Ytest = 1.39 + 0.5t Detrended series: Yt – Ytest
Vijayamohan: CDS MPhil: Time Series 5 45 15 October 2012
Differencing a TSP:
Yt = a + bt +εt; then
Yt-1 = a + b(t–1) +εt-1;
⇒ ∆Yt = Yt –Yt-1 = b + εt – εt-1.
∆Yt has a MA unit root!
Vijayamohan: CDS MPhil: Time Series 4 46 15 October 2012
23
10/15/2012
Differencing a TSP:
Differencing detrends,
but generates MA process
that can show a cycle
when there is none in the original series:
‘spurious cycle’: Slutzky effect (Slutzky 1937).
Differenced series:
∆Yt = Yt –Yt-1 = 5 + εt – εt-1
Detrended series:
Yt – (2 + 0.05t) = εt
Vijayamohan: CDS MPhil: Time Series 4 47 15 October 2012
Detrending :
Time series regression on time:
The classic result of Ragnar Frisch
and F. V. Waugh (1933):
including a time trend in a regression =
detrending the variables by regressing them
individually on time.
(1) Residuals interpreted as cyclical
components in business cycle theory;
(2) Estimating trend growth rates
(Crafts, et al., 1989)
Vijayamohan: CDS MPhil: Time Series 5 48 15 October 2012
24
10/15/2012
15 October 2012 Vijayamohan: CDS MPhil: Time 49
Series 5
25
10/15/2012
Time Series Econometrics
5
Vijayamohanan Pillai N
Vijayamohan: CDS MPhil: Time Series 6 1 15 October 2012
Box-
Box-Jenkins
Methodology:
ARIMA Modelling
Vijayamohan: CDS MPhil: Time Series 6 2 15 October 2012
1
10/15/2012
ARIMA Modelling: Box – Jenkins Methodology:
AR models: first introduced by Yule (1926)
and later generalized by Walker (1931);
MA models first used by Slutzky (1937).
Wold (1938): theoretical foundation of
combined ARMA processes.
George Udny Yule (1871 - 1951) Scottish Statistician
Evgeny Evgenievich Slutzky (1880 – 1948) Russian Statistician
Herman Ole Andreas Wold (1908 – 1992) Swedish Statistician
Vijayamohan: CDS MPhil: Time Series 6 3 15 October 2012
ARIMA Modelling: Box – Jenkins Methodology:
George Box and Gwilym Jenkins (1970; 1976):
Comprehensively put together all the threads of
ARIMA modelling.
The term ‘time series/ARIMA modelling’
= ‘Box-Jenkins methodology’.
Vijayamohan: CDS MPhil: Time Series 6 4 15 October 2012
2
10/15/2012
ARIMA Modelling: Box – Jenkins Methodology:
Objective of Box – Jenkins methodology:
obtain a parsimonious model:
one that describes all the features of the data of
interest using as few parameters (or as simple a
model) as possible.
Ockham’s razor: Lex parsimoniae:
‘Entia non sunt multiplicanda praeter
necessitatem’:
(Entities are not to be multiplied beyond
necessity).
Vijayamohan: CDS MPhil: Time Series 6 5 15 October 2012
ARIMA Modelling: Box – Jenkins Methodology:
Ockham’s razor: Lex parsimoniae:
Entities are not to be multiplied beyond
necessity.
William of Ockham (1285 – 1347/49):
English Franciscan Philosopher
Vijayamohan: CDS MPhil: Time Series 6 6 15 October 2012
3
10/15/2012
ARIMA Modelling:
Box – Jenkins Methodology:
Variance of estimators is inversely proportional
to number of degrees of freedom:
Var (βˆ) = σˆu2 ( X ' X )−1
where
σ̂ u2 =
σ ∑ uˆt2 T−k
t
Larger k ⇒ higher SE ⇒ smaller t-value
not rejecting a false null:
Type 2 error.
Vijayamohan: CDS MPhil: Time Series 6 7 15 October 2012
ARIMA Modelling:
Box – Jenkins Methodology:
Basic steps:
1. Identification of a tentative model;
Successive differencing to achieve stationarity;
2. Estimation of the model; and
3. Diagnostic checking.
Involves more of a judgmental procedure than
the use of any clear-cut rules: Trial and error.
Vijayamohan: CDS MPhil: Time Series 6 8 15 October 2012
4
10/15/2012
BOX-
BOX-JENKINS METHODOLOGY
Plot the series
Does the series appear No Apply transformation
stationary? / Differencing
Yes
Obtain ACF and PACF
No
Does the Correlogram (ACF) decay to zero?
Yes
Identification (Model Selection :
Check ACF and PACF)
Vijayamohan: CDS MPhil: Time Series 6 9 15 October 2012
Identification (Model Selection –
Check ACF and PACF)
Is there a sharp cut-off in ACF?
Yes No
MA Is there a sharp cut-off in PACF?
Yes No
AR ARMA
Estimate parameter values
No Modify
Diagnosis:
model
Are the residuals white noise? Yes
Check ACF and PACF Forecast
Vijayamohan: CDS MPhil: Time Series 6 10 15 October 2012
5
10/15/2012
Identification of ARMA models:
Process ACF PACF
White noise No significant No significant spikes
ARIMA(0,0,0) spikes
DSP ARIMA(0,1,0) Slow decay One significant spike
Autoregressive processes ARIMA(p,0,0)
AR(1) φ1 > 0 Exponential decay: 1 +ve spike at lag 1
+ve spike
AR(1) φ1 < 0 Oscillatory decay,
starts with –ve spike
1 –ve spike at lag 1.
AR(2) Exponential decay, 2 +ve spikes at lags 1
φ1, φ2 > 0 +ve spikes. and 2.
AR(2) φ1 <0, Oscillating 1 negative spike at lag
exponential decay. 1; and 1 +ve spike at
φ2 > 0 lag 2.
AR(p) Decays toward zero; Spikes up to lag p.
coefficients may
Vijayamohan: CDS MPhil: Time Series oscillate.
6 11 15 October 2012
Identification of ARMA models:
Process ACF PACF
Moving Average processes ARIMA(0,0,q)
MA(1) θ1 >0 1 –ve spike at lag 1. Exponential decay of
–ve spikes
MA(1) θ1 <0 1 +ve spike at lag 1 Oscillatory
exponential decay of
+ve and –ve spikes
MA(2) 2 –ve spikes at lags Exponential decay of
θ 1, θ 2 > 0 1 and 2. –ve spikes.
MA(2) 2 +ve spikes at lags Oscillating
θ1, θ2 <0 1 and 2. exponential decay of
+ve and –ve spikes.
MA(q) Spikes up to lag q. Exponential/oscillati
ng decay.
Vijayamohan: CDS MPhil: Time Series 6 12 15 October 2012
6
10/15/2012
Identification of ARMA models:
Process ACF PACF
Hybrid processes ARIMA(p,0,q)
ARIMA(1,0,1) Exponential decay of Exponential decay of
φ1 > 0, θ1 > 0 +ve spikes +ve spikes
ARIMA(1,0,1) Exponential decay of Oscillatory
φ1 > 0, θ1 < 0 +ve spikes exponential decay of
+ve and –ve spikes
ARIMA(1,0,1) Oscillatory Exponential decay of
φ1 < 0, θ1 > 0 exponential decay –ve spikes.
ARIMA(1,0,1) Oscillatory Oscillating
φ1 < 0, θ1 <0
exponential decay of
–ve and +ve spikes.
exponential decay of
–ve and+ve spikes.
ARIMA(p,d,q) Decay (either direct Decay (either direct
or oscillatory) or oscillatory)
beginning at lag q. beginning at lag p.
Vijayamohan: CDS MPhil: Time Series 6 13 15 October 2012
Identification of ARMA models:
Using ACF and PACF:
Apply Tests of Significance on ACF and PACF
(i.e., for Stationarity):
For Individual ACF/PACF:
1
95% confidence interval is: ± 1.96
T
H0: sample ACC/PACC = 0 is rejected,
if it falls outside this region for any k.
Vijayamohan: CDS MPhil: Time Series 6 14 15 October 2012
7
10/15/2012
Identification of ARMA models:
ACF and PACF for a sample of 100 observations:
Lag 1 2 3 4 5 6 7 8 9
ACF 0.321 0.301 0.215 0.142 -0.02 -0.01 0.005 0.001 0.011
PACF 0.321 0.155 0.131 0.102 0.085 0.051 0.008 0.014 0.005
95% confidence interval
= ±1.96/10 = (- 0.196, +0.196).
ACF fast declining and PACF one spike at lag 1
AR(1)
Vijayamohan: CDS MPhil: Time Series 6 15 15 October 2012
Identification of ARMA models:
ACF and PACF for a sample of 100 observations:
Lag 1 2 3 4 5 6 7 8 9
ACF 0.321 0.155 0.131 0.102 0.085 0.051 0.008 0.014 0.005
PACF 0.321 0.301 0.215 0.142 -0.02 -0.01 0.005 0.001 0.011
95% confidence interval
= ±1.96/10 = (- 0.196, +0.196).
ACF one spike at lag 1 and PACF fast declining
MA(1)
Vijayamohan: CDS MPhil: Time Series 6 16 15 October 2012
8
10/15/2012
ARIMA Modelling: Box – Jenkins Methodology:
2. Estimation:
AR Model: straightforward:
OLS method;
MA Model:
ML method;
ARMA Model:
ML method for the MA component.
Vijayamohan: CDS MPhil: Time Series 6 17 15 October 2012
ARIMA Modelling: Box – Jenkins Methodology:
3. Diagnostic checking: ‘Model adequacy’ tests:
(a) ‘Residual analysis’ and
(b) ‘Trial overfitting’ (Box and Jenkins)
(i) ‘Trial Overfitting’:
If an ARMA(p, q) model is chosen, also
estimate an ARMA(p+1, q) model and an
ARMA(p, q+1) model and test for
significance of the additional parameters.
Vijayamohan: CDS MPhil: Time Series 6 18 15 October 2012
9
10/15/2012
ARIMA Modelling: Box – Jenkins Methodology:
(ii) ‘Residual analysis’ (Refer: My working paper on
Forecasting’)
‘Electricity Demand Analysis and Forecasting’
⇒ Residuals of an adequate model = white noise;
purely
random;
⇒ Plot of residuals; check for outliers;
⇒ No serial correlation; examine ACF; PACF;
Ljung-Box (1978) portmanteau statistic
Q* ∼ χ2(k – p – q) for an ARMA(p, q),
if correctly specified.
Small Q* value or large p-value: ⇒ model adequacy.
Vijayamohan: CDS MPhil: Time Series 6 19 15 October 2012
ARIMA Modelling: Box – Jenkins Methodology:
Residual analysis
Problem with Q*: Not reliable: (same as with DW
with lagged endogenous variable; so Durbin’s h):
Residual autocorrelations are biased towards zero,
when lagged dependent variable is included as
regressors in the model; e.g., DW → 2, even when
residuals are serially correlated.
Lagrange Multiplier F-test (Harvey 1981):
Small F-value or large p-vale ⇒
No autocorrelation.
Harvey 1981, The Econometric Analysis of Time
Series, Philip Allen, Deddington.
Vijayamohan: CDS MPhil: Time Series 6 20 15 October 2012
10
10/15/2012
ARIMA Modelling: Box – Jenkins Methodology:
Model Selection Criteria:
If a MA(2) model provides the same fit as an AR(10)
model, select the first.
Adding more lags (p, q) necessarily reduces residual
sum of squares ⇒ R2 , goodness of fit ↑; but degrees of
freedom ↓:
Trade-off between ‘Goodness of fit’ and ‘Parsimony’:
Various ‘Information Criteria’ that trade off the two:
Vijayamohan: CDS MPhil: Time Series 6 21 15 October 2012
ARIMA Modelling: Box – Jenkins Methodology:
Theory of estimation:
Information content (of a parameter)
in a random sample is represented by
variance of its (unbiased) estimator:
small variance ⇒ large information.
Vijayamohan: CDS MPhil: Time Series 6 22 15 October 2012
11
10/15/2012
ARIMA Modelling: Box – Jenkins Methodology:
‘Information Criteria’: In general
IC(k) = T ln σˆ 2 + k { f (T )}, for k = 1, …, p+q
where σ̂ 2
σ is the estimated error variance:
(RSS/(T – k); and
k{f(T)} = ‘Penalty’ function for increasing the order of
model, for the loss of degrees of freedom.
We choose the ARMA model with the lowest IC.
Vijayamohan: CDS MPhil: Time Series 6 23 15 October 2012
ARIMA Modelling: Box – Jenkins Methodology:
‘Information Criteria’: In general
IC(k) = T ln σˆ 2 + k { f (T )}, for k = 1, …, p+q
(1) f(T) = 2 → Akaike IC = T ln σˆ 2 + 2k
Hirotugu Akaike (赤池 弘次)
弘次)
(1927 –2009)
2009) Japanese statistician.
Akaike, Hirotugu (December 1974).
"A new look at the statistical model identification".
IEEE Transactions on Automatic Control 19 (6):716–723.
Vijayamohan: CDS MPhil: Time Series 6 24 15 October 2012
12
10/15/2012
ARIMA Modelling: Box – Jenkins Methodology:
‘Information Criteria’: In general
IC(k) = T ln σˆ 2 + k { f (T )}, for k = 1, …, p+q
2. f(T) = lnT → Schwartz Bayesian IC:
T ln σˆ 2 + k ln T
Schwarz,
Schwarz, Gideon E. (1978).
"Estimating the dimension of a model".
Annals of Statistics 6 (2): 461–
461–464.
Vijayamohan: CDS MPhil: Time Series 6 25 15 October 2012
ARIMA Modelling: Box – Jenkins Methodology:
‘Information Criteria’: In general
IC(k) = T ln σˆ 2 + k { f (T )}, for k = 1, …, p+q
(3) f(T) = 2 ln(lnT) → Hannan-Quinn IC:
T ln σˆ 2 + 2k ln(ln T )
Hannan,
Hannan, E. J., and B. G. Quinn (1979)
The Determination of the Order of an Autoregression,
Autoregression,
Journal of the Royal Statistical Society, B, 41, 190–
190–195.
Vijayamohan: CDS MPhil: Time Series 6 26 15 October 2012
13
10/15/2012
ARIMA Modelling: Box – Jenkins Methodology:
‘Information Criteria’:
(1) f(T) = 2 → Akaike (1974) IC = T ln σˆ 2 + 2k
(2) f(T) = lnT → Schwartz (1978) Bayesian IC:
T ln σˆ 2 + k ln T
BIC gives more weight to k than AIC if T > 7:
an ↑ in k requires a larger ↓ in σ̂ 2
σ
under BIC than under AIC.
As lnT > 2, (T > 7), BIC always select a
more parsimonious model than AIC.
Vijayamohan: CDS MPhil: Time Series 6 27 15 October 2012
ARIMA Modelling: Box – Jenkins Methodology:
‘Information Criteria’:
(3) f(T) = 2 ln(lnT) → Hannan-Quinn (1979) IC:
T ln σˆ 2 + 2k ln(ln T )
For HQIC, weight on k is greater than 2, if T > 15.
Vijayamohan: CDS MPhil: Time Series 6 28 15 October 2012
14
10/15/2012
Model selection for the Growth of the Annual Index of Output:
1955 – 1980
SE ACF 0.339 –0.24 –0.524 –0.294
0.196 PACF 0.339 –0.402 –0.368 –0.081
Model RSS k LM FAC AIC BIC
AR(1) 0.7951 1 4.226 –88.67 –87.41
AR(2) 0.6715 2 3.003 –91.06 –88.55
MA(1) 0.7495 1 0.057 –90.21 –88.95
MA(2) 0.7377 2 1.732 –88.62 –86.1
ARMA(1,1) 0.7501 2 2.301 –88.19 –85.67
ARMA(2,1) 0.5117 3 4.398** –96.03 –92.36
ARMA(1,2) 0.6050 3 7.888** –91.78 –88
ARMA(2,2) 0.5153 4 3.719 –93.95 –88.92
Note: ** = Significant at 5 % level. Source: Franses, Philip Hans, 1998,
Time Series Models for Business and Economic Forecasting, CUP: Table 3.2.
Vijayamohan: CDS MPhil: Time Series 6 29 15 October 2012
Box-
Box-Jenkins
Methodology:
ARIMA Modelling
Examples
Vijayamohan: CDS MPhil: Time Series 6 30 15 October 2012
15
10/15/2012
Yt
ACF PACF
Vijayamohan: CDS MPhil: Time Series 6 31 15 October 2012
∆Yt
ACF PACF
Vijayamohan: CDS MPhil: Time Series 6 32 15 October 2012
16
10/15/2012
EQ( 1) Modelling ∆yt by OLS (using Data1)
The estimation sample is: 2 to 200
Coefficient Std.Error t-value t-prob
∆ yt – 1 0.605620 0.05675 10.7 0.000
Constant 0.0120266 0.06819 0.176 0.860
sigma 0.961773 RSS 182.226355
R^2 0.36634 F(1,197) = 113.9 [0.000]**
log-likelihood -273.607 DW 1.93
no. of observations 199 no. of parameters 2
AR 1-2 test: F(2,195) = 1.4581 [0.2352]
ARCH 1-1 test: F(1,195) = 0.70516 [0.4021]
Normality test: Chi^2(2) = 0.70709 [0.7022]
hetero test: F(2,194) = 0.59821 [0.5508]
hetero-X test: F(2,194) = 0.59821 [0.5508]
RESET test: F(1,196) = 2.2481 [0.1354]
Vijayamohan: CDS MPhil: Time Series 6 33 15 October 2012
Residuals
ACF PACF
Vijayamohan: CDS MPhil: Time Series 6 34 15 October 2012
17
10/15/2012
EQ( 2) Modelling ∆yt by OLS (using Data1)
The estimation sample is: 3 to 200
Coefficient Std.Error t-value t-prob
∆ yt – 1 0.639012 0.07156 8.93 0.000
∆ yt – 2 -0.0551546 0.07156 -0.771 0.442
Constant 0.0125556 0.06861 0.183 0.855
sigma 0.965223 RSS 181.672753
R^2 0.368264 F(2,195) = 56.84 [0.000]**
log-likelihood -272.43 DW 2.01
no. of observations 198 no. of parameters 3
AR 1-2 test: F(2,193) = 1.1629 [0.3148]
ARCH 1-1 test: F(1,193) = 0.46776 [0.4948]
Normality test: Chi^2(2) = 1.2041 [0.5477]
hetero test: F(4,190) = 2.3548 [0.0553]
hetero-X test: F(5,189) = 2.0145 [0.0784]
RESET test: F(1,194) = 2.5648 [0.1109]
Vijayamohan: CDS MPhil: Time Series 6 35 15 October 2012
Xt
ACF PACF
Vijayamohan: CDS MPhil: Time Series 6 36 15 October 2012
18
10/15/2012
∆ Xt
ACF PACF
Vijayamohan: CDS MPhil: Time Series 6 37 15 October 2012
EQ(13) Modelling ∆Xt by RALS (using Data1)
The estimation sample is: 3 to 200
Coefficient Std.Error t-value t-prob
Constant 0.00838088 0.05187 0.162 0.872
Uhat_1 -0.326969 0.06741 -4.85 0.000
sigma 0.96844 RSS 183.823765
no. of observations 198 no. of parameters 2
mean(∆ ∆ Xt ) 0.00750231 ∆ Xt )
var(∆ 1.03984
Roots of error polynomial: real imag modulus
0.32697 0.00000 0.32697
ARCH 1-1 test: F(1,194) = 0.98408 [0.3224]
Normality test: Chi^2(2) = 0.57691 [0.7494]
Vijayamohan: CDS MPhil: Time Series 6 38 15 October 2012
19
10/15/2012
Residuals
ACF PACF
Vijayamohan: CDS MPhil: Time Series 6 39 15 October 2012
Yt
ACF PACF
Vijayamohan: CDS MPhil: Time Series 6 40 15 October 2012
20
10/15/2012
EQ(10) Modelling yt by RALS (using Data1)
The estimation sample is: 3 to 200
Coefficient Std.Error t-value t-prob
yt – 1 0.544175 0.08824 6.17 0.000
Constant 0.0103062 0.05609 0.184 0.854
Uhat_1 -0.218511 0.1026 -2.13 0.034
sigma 0.961309 RSS 180.202435
no. of observations 198 no. of parameters 3
Roots of error polynomial: real imag modulus
0.21851 0.00000 0.21851
ARCH 1-1 test: F(1,193) = 0.84109 [0.3602]
Normality test: Chi^2(2) = 0.90110 [0.6373]
hetero test: F(2,193) = 1.7595 [0.1749]
hetero-X test: F(2,193) = 1.7595 [0.1749]
Vijayamohan: CDS MPhil: Time Series 6 41 15 October 2012
Residuals
ACF
PACF
Vijayamohan: CDS MPhil: Time Series 6 42 15 October 2012
21
10/15/2012
15 October 2012 Vijayamohan: CDS MPhil: Time 43
Series 6
22
11/5/2013
Unit Root
Time Series Econometrics Tests
7
Vijayamohanan Pillai N
Vijayamohan: CDS M Phil: Time Series 7 1 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 2 5 November 2013
Non-stationarity
Non-stationarity due to
Unit Root (1)Integrated process: random walk:
Tests Yt = Yt-1 + εt
Difference stationary process
‘Jack and Jill went up the hill, to look at the stars. Stochastic trend
“Do you see any unit roots there?”
Jack asks Jill who was using a telescope.
“The stars all seem to be cointegrated”, replies Jill.’ (2) Trend: Trend stationary process
Both stochastic and deterministic trend
G.S. Maddala (1998) ‘Recent Developments in Dynamic
Econometric Modelling: A Personal Viewpoint’,
Vijayamohan: CDS M Phil: Time Series 7 3
Political Analysis5; November
7: 59-87 2013 Vijayamohan: CDS MPhil: Time Series 5 4 5 November 2013
Consequences of Non-stationarity Consequences of Non-stationarity
(2) For non-stationary series, Var and Covs are
(1) With stationary series, possible to model the functions of time: so the conventional
process via a fixed-coefficients equation asymptotic theory cannot be applied to these
estimated from past data. series.
Not possible if the structural relationship
changes over time (if non-stationary).
For example: In a regression of Yt on Xt:
(2) For non-stationary series, Var and Covs are βˆ = Cov (Yt , X t ) Var ( X t )
functions of time:
so the conventional asymptotic theory cannot If Xt is non-stationary, Var(Xt) ↑ infinitely, and
be applied to these series. dominates the Cov(Yt, Xt):
Then the OLS estimator does not have an
asymptotic distribution.
Vijayamohan: CDS MPhil: Time Series 5 5 5 November 2013 Vijayamohan: CDS MPhil: Time Series 5 6 5 November 2013
1
11/5/2013
Consequence of unit root (non-stationarity): Consequence of unit root (non-stationarity):
y x
20
15
Consider two unrelated RW processes: 10
5
Yt = Yt–1 + Ut; Ut ∼ IIN(0, σU2)
0
Xt = Xt–1 + Vt; Vt ∼ IIN(0, σV2); -5
-10
Cov(Ut , Vt) = 0. 0 50 100 150 200 250 300
Yt = Yt–1 + Ut; Ut ∼ IIN(0, σU2)
Xt = Xt–1 + Vt; Vt ∼ IIN(0, σV2);
Cov(Ut , Vt) = 0.
Vijayamohan: CDS MPhil: Time Series 5 7 5 November 2013 Vijayamohan: CDS MPhil: Time Series 5 8 5 November 2013
Consequence of unit root (non-stationarity): Consequence of unit root (non-stationarity):
3 eps ep
Cross Correlation function
1.00
2 CCF-eps x ep CCF-ep x eps
0.75
1
0.50
0
0.25
-1
0.00
-2 -0.25
-3 -0.50
0 50 100 150 200 250 300 -0.75
Ut ∼ IIN(0, σU2) 0 5 10
Cov(Ut , Vt) = 0.
Vt ∼ IIN(0, σV2)
Vijayamohan: CDS MPhil: Time Series 5 9 5 November 2013 Vijayamohan: CDS MPhil: Time Series 5 10 5 November 2013
Consequence of unit root (non-stationarity): Consequence of unit root (non-stationarity):
Now consider the regression:
Yt =β 0 + β 1 Xt + εt.
We expect R2 from this regression would tend
to zero; BUT….
Vijayamohan: CDS MPhil: Time Series 5 11 5 November 2013 Vijayamohan: CDS MPhil: Time Series 5 12 5 November 2013
2
11/5/2013
Consequence of unit root (non-stationarity): Consequence of unit root (non-stationarity):
Granger and Newbold (1974):
High R2 and highly significant t,
but a low DW statistic.
When the regression was run
in first differences……
?
Vijayamohan: CDS MPhil: Time Series 5 13 5 November 2013 Vijayamohan: CDS MPhil: Time Series 5 14 5 November 2013
Consequence of unit root (non-stationarity): Consequence of unit root (non-stationarity):
When the regression was run
in first differences,
R2 close to zero; DW close to 2: ⇒
No relationship between Yt and Xt ; and
the high R2 obtained was ‘spurious’.
R2 > DW ⇒ ‘Spurious Regression’.
Vijayamohan: CDS MPhil: Time Series 5 15 5 November 2013 Vijayamohan: CDS MPhil: Time Series 5 16 5 November 2013
So, check for stationarity of time series:
If non-stationary series,
Classical: Autocorrelation function (ACF):
Correlogram run regression in (first) differences as in
Fast-decreasing ACF ⇒ Stationarity Classical (ARIMA) modelling
Modern: Unit root tests: or check for Cointegration among the series
(Augmented) Dickey-Fuller test;
Phillips-Perron non-parametric test; etc.
Vijayamohan: CDS MPhil: Time Series 5 17 5 November 2013 Vijayamohan: CDS MPhil: Time Series 5 18 5 November 2013
3
11/5/2013
Question of choice between Main Differences between TSP and DSP
differencing and detrending →
Deterministic Random Walk
Trend with /without Drift
Recognizing the differentiation between TSP and DSP.
Transformations to Detrend Difference
TSP: stationary around a deterministic trend: achieve stationarity
xt = α + βt + ut;
DSP: Only stochastic trend: Effects of shocks Wears out Permanent
yt = yt−−1 + ut: as ∆yt is stationary.
Variance Bounded Unbounded
Vijayamohan: CDS M Phil: Time Series 7 19 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 20 5 November 2013
Dickey-Fuller (DF) Unit Root Test
Consider the following model:
yt = α + β t +ρ yt-1 + ut ,
Discrimination between TSP and DSP models: essential;
Ut : white noise.
⇒ whether the root of the series ρ = 1 or | ρ | < 1.
We have the following possibilities:
Hence the significance of unit root tests.
1. When β ≠ 0, | ρ | < 1,
yt has a linear trend and hence is a TSP.
2. When β = 0, then yt = α +ρ yt-1 + ut .
3.When α = β = 0, then yt = ρ yt-1 + ut .
Vijayamohan: CDS M Phil: Time Series 7 21 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 22 5 November 2013
Dickey-Fuller (DF) Unit Root Test Dickey-Fuller (DF) Unit Root Test
Given yt = α + βt +ρ yt-1 + ut , Given yt = α + βt +ρ yt-1 + ut ,
When β = 0, yt = α +ρ yt-1 + ut . When α = β = 0, yt = ρ yt-1 + ut ,
We have two cases:
(i) | ρ | < 1, yt is a stationary series; Two cases here are:
(ii) ρ = 1, yt is a DSP with a drift. (i) | ρ | < 1, yt is stationary;
(ii) ρ = 1, yt is a DSP without drift.
Vijayamohan: CDS M Phil: Time Series 7 23 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 24 5 November 2013
4
11/5/2013
Dickey-Fuller (DF) Unit Root Test Dickey-Fuller (DF) Unit Root Test
Rewrite (1) yt = α + β t +ρ yt-1 + ut as ||ly rewrite (2) yt = α +ρ yt-1 + ut
as ∆ yt = α +γ yt-1 + ut .
∆ yt = α + β t +γ yt-1 + ut ,
where γ = (ρ −1).
and (3) yt = ρ yt-1 + ut ,
Now, testing the null hypothesis
as ∆ yt = γ yt-1 + ut .
Ho: γ = 0,
in the usual way is equivalent to testing
Then test for Ho: γ = 0,
Ho: ρ = 1.
vs. the one-sided alternative | ρ | < 1 .
Vijayamohan: CDS M Phil: Time Series 7 25 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 26 5 November 2013
Dickey-Fuller (DF) Unit Root Dickey-Fuller (DF) Unit Root Test
Test: 3 DF Type Formulations But we cannot use the usual t-test to test
Note: In (1), ∆ yt = α + β t +γ yt-1 + ut , Ho: ρ = 1,
where γ = (ρ −1), because under the null, yt is I(1),
the model has both a constant and a trend; and hence the t-statistic does not have an
asymptotic normal distribution (Dickey and
Note: In (2), ∆ yt = α +γ yt-1 + ut , Fuller 1979).
the model has a constant; and
Note: In (3), ∆ yt = γ yt-1 + ut , Its asymptotic distribution,
the model is without constant. based on Wiener processes, is called
Dickey-Fuller distribution,
In a DF test all 3 formulations are considered. and the statistic, Dickey-Fuller τ (tau) statistic.
Vijayamohan: CDS M Phil: Time Series 7 27 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 28 5 November 2013
Wiener process:
a continuous-time stochastic process
Also called Brownian Motion Brownian motion:
1827
Norbert Wiener (1894 –1964,)
American mathematician;
originator of cybernetics:
cybernetics:
interdisciplinary study of the structure of Robert Brown (1773 – 1858)
regulatory systems Scottish botanist
Vijayamohan: CDS M Phil: Time Series 7 29 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 30 5 November 2013
5
11/5/2013
Graphs of five sampled Dickey-Fuller (DF) Unit Root Test
Brownian motions:
motions:
Critical values tabulated by Dickey and Fuller
(1979) and MacKinnon (1990) for a wider
range of sample.
Most of the statistic outcomes are negative;
Note: Ho: γ = (ρ −1) = 0.
If the estimated τ–value is more negative
(i.e., less) than the critical value
at the chosen significance level, reject Ho.
Robert Brown (1773 – 1858)
Scottish botanist
Vijayamohan: CDS M Phil: Time Series 7 31 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 32 5 November 2013
Augmented Dickey-Fuller (ADF) Unit Root Test Augmented Dickey-Fuller (ADF) Unit Root Test
Dickey and Fuller (1979) and
In deriving the asymptotic distributions,
Dickey and Fuller (1979, 1981) assumed: Said and Dickey (1984)
ut ∼ iid(0,σ2 ). modified the DF test
by means of AR correction:
But, if the errors are non-orthogonal Adding lags to ‘whiten’ the residuals
(i.e., serially correlated), gives ADF test.
the limiting distributions cease to be Augmented Dickey-Fuller test (ADF)
appropriate.
Vijayamohan: CDS M Phil: Time Series 7 33 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 34 5 November 2013
Augmented Dickey-Fuller (ADF) Unit Root Test
Double Unit Roots Testing
Augmented Dickey-Fuller test (ADF):
Test for a unit root in Yt ;
by estimating an autoregression of ∆yt on its
own lags and yt-1 using OLS: If the unit root null is not rejected
p (if Yt appears I(1)),
∆y t = γy t −1 + ∑ β i∆y t−i + ut . test for a second unit root (see if Yt is I(2)):
i =1
Testing: (1):
When γ = 0, ρ = 1. Estimate the regression of ∆2Yt on a constant,
The (t -) test statistic follows the same ∆Yt–1, and the lagged values of ∆2Yt, and
DF distribution (τ–statistic)
compare the ‘t-ratio’ of the coefficient of ∆Yt–1
with the DF critical values.
Vijayamohan: CDS M Phil: Time Series 7 35 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 36 5 November 2013
6
11/5/2013
Testing for Unit Root:
Double Unit Roots Testing (Augmented) Dickey - Fuller test:
Testing: (2):
Consumption
Estimate the regression of ∆2Yt on Yt–1, ∆Yt–1,
and the lagged values of ∆2Yt, and
compute the usual F-statistic for testing the
joint significance of Yt–1 and ∆Yt–1,
using the critical values given as
Φ1(2) by Hasza and Fuller (1979).
∆Consumption
Vijayamohan: CDS M Phil: Time Series 7 37 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 38 5 November 2013
Testing for Unit Root: Testing for Unit Root:
(Augmented) Dickey - Fuller test: (Augmented) Dickey - Fuller test:
EQ( 1) Modelling ∆Ct by OLS (using Data.in7) EQ( 1) Modelling ∆2Ct by OLS (using Data.in7)
The estimation sample is: 1953 (2) to 1992 (3) The estimation sample is: 1953 (3) to 1992 (3)
Coefficient Std.Error t-value t-prob Coefficient Std.Error t-value t-prob
Constant 9.09221 11.46 0.793 0.429 Constant -0.148864 0.1736 -0.857 0.393
Ct-1 -0.0106221 0.01308 -0.812 0.418 ∆Ct -1 -0.813544 0.07824 -10.4 0.000
sigma 2.21251 RSS 763.648214 sigma 2.16501 RSS 726.524426
R^2 0.00420671 F(1,156) = 0.659 [0.418] R^2 0.410943 F(1,155) = 108.1 [0.000]**
log-likelihood -348.658 DW 1.6 log-likelihood -343.037 DW 2.04
no. of observations 158 no. of parameters 2 no. of observations 157 no. of parameters 2
∆Ct = 9.092 – 0.0106 Ct – 1 ∆2Ct = –0.149 – 0.814 ∆Ct – 1
t = (–0.857) (–10.399)
t = (0.793) (–0.812)
Critical values used in ADF test: Critical values used in ADF test:
5%= -2.88, 1%= -3.473 5%=-2.88, 1%=-3.473
Vijayamohan: CDS M Phil: Time Series 7 39 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 40 5 November 2013
Testing for Unit Root: Testing for Unit Root:
(Augmented) Dickey - Fuller test: (Augmented) Dickey - Fuller test:
Variable Ct in level Variable Ct in First Difference
CONS: ADF tests (T=152, Constant; 5% = -2.88 1% = -3.47) DCONS: ADF tests (T=152, Constant; 5% = -2.88 1% = -3.47)
D-lag t-adf beta Y_1 sigma t-DY_lag t-prob AIC F-prob D-lag t-adf beta Y_1 sigma t-DY_lag t-prob AIC F-prob
5 -1.094 0.98551 2.123 -2.767 0.0064 1.551 5 -4.870** 0.28208 2.132 0.06925 0.9449 1.559
4 -1.464 0.98038 2.171 1.253 0.2123 1.589 0.0064 4 -5.305** 0.28618 2.125 2.948 0.0037 1.546 0.9449
3 -4.448** 0.42488 2.180 -1.053 0.2941 1.591 0.0151
3 -1.298 0.98274 2.176 1.448 0.1496 1.587 0.0110
2 -5.313** 0.37033 2.181 -1.293 0.1981 1.585 0.0232
2 -1.120 0.98516 2.184 1.687 0.0937 1.588 0.0112
1 -6.801** 0.29538 2.185 -1.570 0.1185 1.583 0.0246
1 -0.9317 0.98766 2.197 2.481 0.0142 1.594 0.0075
0 -10.08** 0.19164 2.196 1.586 0.0182
0 -0.6361 0.99149 2.234 1.621 0.0013
Vijayamohan: CDS M Phil: Time Series 7 41 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 42 5 November 2013
7
11/5/2013
Results of Nelson and Plosser (1982, Table 5, p. 151):
Unit Root Test for TSP vs. DSP: Critical value at 5% level: −3.45
Nelson and Plosser (1982) Series Sample size Lag (k) τ-value
(T)
The first ever attempt: Nelson and Plosser Real GDP 62 2 −2.99
(1982), using the (augmented) Dickey-Fuller Nominal GDP 62 2 −2.32
unit root tests: Real per capita GNP 62 2 −3.04
Industrial production 111 6 −2.53
Employment 81 3 −2.66
H0: a time series belongs to DSP class against Unemployment rate 81 4 −3.55*
Ha: it belongs to TSP class. GNP deflator 82 2 −2.52
Consumer prices 111 4 −1.97
Wages 71 3 −2.09
Nelson and Plosser found that 13 out of 14 US Real wages 71 2 −3.04
macroeconomic time series that they analyzed Money stock 82 2 −3.08
belonged to the DSP class (the exception being Velocity 102 1 −1.66
the unemployment rate). Interest rate 71 3 0.686
Common stock
5 November prices
2013 Vijayamohan: 100
CDS M Phil: Time Series 7 3 −2.05
Vijayamohan: CDS M Phil: Time Series 7 43 5 November 2013 44
Other Unit Root Tests Other Unit Root Tests
Mushrooming of Unit root tests ever since N and P
(1982) feat: 8. Choi (1992): Pseudo-IV estimator test;
1. Sargan and Bhargava (1983): based on DW statistic; 9. Yap and Reinsel (1995): Likelihood ratio test in
ARMA models;
2. Phillips and Perron (1988): non-parametric test;
10. Leybourne (1995): based on forward and reverse
3. Cochrane (1988): Variance ratio test; DF regressions;
4. Sims (1988): Bayesian approach to unit root testing; 11. Park and Fuller (1995): Weighted symmetric
5. Perron (1989): unit root test under structural break; estimator test;
6. Pantula and Hall (1991): IV test in ARMA models; 12. Elliott, Rothenberg and Stock (1996): Dickey-
Fuller GLS test;
7. Schmidt and Phillips (1992): LM test;
Vijayamohan: CDS M Phil: Time Series 7 45 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 46 5 November 2013
Other Unit Root Tests Other Unit Root Tests
Tests with stationarity as null:
1. Park (1990)’s J-test;
Panel data unit root tests:
2. Kwiatkowski, Phillips, Schmidt and Shin (1992):
(KPSS) test; • Levin and Lin (1993);
3. Bierens and Guo (1993): based on Cauchy • Breitung and Meyer (1994)
distribution; • Quah (1994);
4. Leybourne and McCabe (1994): Modified KPSS • Pesaran and Shin (1996);
test;
5. Choi (1994): based on testing for a MA unit root;
6. Arellano and Pantula (1995): based on testing for
a MA unit root.
Vijayamohan: CDS M Phil: Time Series 7 47 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 48 5 November 2013
8
11/5/2013
Empirical Studies by Model Unit Root Stationary Time series in STATA
Unit Root Tests with the Nelson and
(with (with
possible possible Tsset : Declare data to be time-series data
Plosser’s Data (1982) Set
breaks) breaks)
Copy-paste data in data editor
Nelson and Plosser ADF test with no 13 1
Generate a time variable by typing the command:
(1982) break
Perron (1989)** Exogenous with 3 11
one break For monthly data starting with 1995 July:
Zivot and Andrews Endogenous with 10 3
(1992)* one break . generate time = m(1995m7) + _n -1
. format t %tm
Lumsdaine and Endogenous with 8 5
Papell (1997)* two breaks You can now tsset your data set
. tsset time
Lee and Strazicich Endogenous with 10 4
(2003)** two breaks
time variable: time, 1995m7 to 2004m6
delta: 1 month
* Assume no break(s) under the H0 of unit root.
** Assume break(s) under both the null and the HA
Vijayamohan: CDS M Phil: Time Series 7 49 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 50 5 November 2013
If the first observation is for the first quarter of 1990, For weekly data starting at 1994w1 type:
generate time = q(1990q1) + _n-1 . generate time = w(1994w1) + _n-1
format time %tq . format time %tw
tsset time . tsset time
For yearly data starting at 1942 type: For daily data starting at 1jan1999 type:
generate time = y(1942) + _n-1
format time %ty . generate time = d(1jan1999) + _n-1
tsset time . format time %td
. tsset time
For half yearly data starting at 1921h2 type:
generate time = h(1921h2) + _n-1
format time %th
tsset time
Vijayamohan: CDS M Phil: Time Series 7 51 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 52 5 November 2013
Unit Root Tests in STATA:
Statistics →
Time series →
. dfuller D.cons, lags(5)
Tests
Augmented Dickey-Fuller test for unit root Number of obs = 152
. dfuller cons, lags(5)
---------- Interpolated Dickey-Fuller ---------
Augmented Dickey-Fuller test for unit root Number of obs = 153 Test 1% Critical 5% Critical 10% Critical
Statistic Value Value Value
---------- Interpolated Dickey-Fuller --------- ------------------------------------------------------------------------------
Test 1% Critical 5% Critical 10% Critical Z(t) -4.870 -3.493 -2.887 -2.577
Statistic Value Value Value ------------------------------------------------------------------------------
------------------------------------------------------------------------------ MacKinnon approximate p-value for Z(t) = 0.0000
Z(t) -1.117 -3.492 -2.886 -2.576
------------------------------------------------------------------------------
MacKinnon approximate p-value for Z(t) = 0.7081
Vijayamohan: CDS M Phil: Time Series 7 53 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 54 5 November 2013
9
11/5/2013
. gen lcons=ln(cons)
. pperron cons, lags(5)
. pperron lcons, lags(5)
Phillips-Perron test for unit root Number of obs = 158
Newey-West lags = 5 Phillips-Perron test for unit root Number of obs = 158
Newey-West lags = 5
---------- Interpolated Dickey-Fuller ---------
Test 1% Critical 5% Critical 10% Critical ---------- Interpolated Dickey-Fuller ---------
Statistic Value Value Value Test 1% Critical 5% Critical 10% Critical
------------------------------------------------------------------------- Statistic Value Value Value
----- ------------------------------------------------------------------------------
Z(rho) -3.253 -19.993 -13.816 -11.077 Z(rho) -3.242 -19.993 -13.816 -11.077
Z(t) -1.194 -3.491 -2.886 -2.576 Z(t) -1.190 -3.491 -2.886 -2.576
------------------------------------------------------------------------- ------------------------------------------------------------------------------
----- MacKinnon approximate p-value for Z(t) = 0.6778
MacKinnon approximate p-value for Z(t) = 0.6760
Vijayamohan: CDS M Phil: Time Series 7 55 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 56 5 November 2013
. pperron D.lcons, lags(5)
. pperron D.cons, lags(5)
Phillips-Perron test for unit root Number of obs = 157
Phillips-Perron test for unit root Number of obs = 157
Newey-West lags = 5
Newey-West lags = 5
---------- Interpolated Dickey-Fuller ---------
---------- Interpolated Dickey-Fuller ---------
Test 1% Critical 5% Critical 10% Critical
Test 1% Critical 5% Critical 10% Critical
Statistic Value Value Value
Statistic Value Value Value
------------------------------------------------------------------------------
Z(rho) -146.064 -19.990 -13.814 -11.076
Z(rho) -146.107 -19.990 -13.814 -11.076
Z(t) -10.656 -3.491 -2.886 -2.576
Z(t) -10.662 -3.491 -2.886 -2.576
------------------------------------------------------------------------------
MacKinnon approximate p-value for Z(t) = 0.0000
MacKinnon approximate p-value for Z(t) = 0.0000
Vijayamohan: CDS M Phil: Time Series 7 57 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 58 5 November 2013
Unit Roots Test: Summing up Unit Roots Test: Summing up
Why so many unit root tests?
“In summary,
There is no uniformly powerful test for the
unit root hypothesis (Stock, 1994). it is high time we asked the question:
Why are we interested in testing for unit
Why all this unit root testing
roots?
We need unit root tests as a prelude to rather than keep suggesting more and more
cointegration analysis.
unit root tests
If so, unit root tests are pre-tests;
and use the Nelson-Plosser data as a guinea
then shouldn’t the significance level be much
higher (say 25% or more)? pig for every unit root test suggested.”
Isn't it Data-mining? Maddala and Kim (1998: 146).
Vijayamohan: CDS M Phil: Time Series 7 59 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 60 5 November 2013
10
11/5/2013
Regression of Integrated Variables
Yt
Deterministic Stochastic
Deterministic Regression Spurious
Xt valid regression
Stochastic Spurious Spurious
regression regression
unless Xt and
Yt are
cointegrated
Vijayamohan: CDS MPhil: Time Series 8 61 05 November 2013 Vijayamohan: CDS MPhil: Time Series 8 62 05 November 2013
Cointegration Cointegration
Spurious (Nonsense) regression with ∵ differencing → ‘valuable long-run information
integrated variables being lost’’.
→ using differenced variables in regression. ∆Yt on ∆Xt):
Regression in differences (∆
BUT….. short run only;
In the long run (in equilibrium):
Solving non-stationarity problem via
∆Yt = 0 and ∆Xt = 0.
differencing =
Most economic relationships are stated in
“Throwing the baby out with the bath water” theory as long run relationships between
∵ differencing → variables in their levels, not in their
“valuable long-run information being lost’’. differences.
Vijayamohan: CDS MPhil: Time Series 8 63 05 November 2013 Vijayamohan: CDS MPhil: Time Series 8 64 05 November 2013
Cointegration
Two seemingly irreconcilable objectives : If two series yt and xt are both I(1),
then in general, any linear combination
1. Avoid spurious regression of I(d) variables
of them will also be I(1);
and
e.g. Yt – Ct = St: All show upward trend.
2. Conserve the long-term relationship.
Hence Cointegration Income
(Granger 1981; Engle and Granger 1987).
Saving
Consumption
Vijayamohan CDS Time Series: Introduction 65 5 November 2013 Vijayamohan: CDS MPhil: Time Series 8 66 05 November 2013
11
11/5/2013
Cointegration Cointegration
an important property of I(1) variables :
there can be some linear combinations of Thus,
them that are in fact I(0), i.e., stationary.
a set of integrated time series is cointegrated,
e.g. Yt – Ct = St ∼ I(0) if some linear combination of
those (non-stationary) series is stationary.
Income
Savng
Consumption
Vijayamohan: CDS MPhil: Time Series 8 67 05 November 2013 Vijayamohan: CDS MPhil: Time Series 8 68 05 November 2013
Cointegration Cointegration
The old man and the dog are joined by a
The old woman and the boy are unrelated leash.
to one another,
Individually, the dog and the man are each on a
except that they are both on a random random walk.
walk in the park.
BUT they cannot wander too far from one
Information about the boy's location tells another because of the leash.
us nothing about the old woman's
location. We say that the random processes describing
their paths are cointegrated.
Vijayamohan: CDS MPhil: Time Series 8 69 05 November 2013 Vijayamohan: CDS MPhil: Time Series 8 70 05 November 2013
Cointegration
Cointegration = Stationary linear combination of
integrated variables.
Clive WJ Granger 1981 :
Note: Variables of the same degree of integration.
combines short-run and long-run perspectives.
The Sveriges Riksbank Prize in Economic
Granger Representation Theorem: Sciences in Memory of Alfred Nobel 2003
If there is an equilibrium relationship between "for methods of analyzing economic time
two economic variables, series with common trends (cointegration)“
they may deviate from the equilibrium in the United Kingdom
short run, but will adjust towards the University of California
San Diego, CA, USA
equilibrium in the longer run. b. 1934
Vijayamohan: Time Series 71 5 November 2013 Vijayamohan: Time Series 72 5 November 2013
12
11/5/2013
Cointegration Cointegration
If there exists a relationship between two non
The system is in long-run equilibrium when
stationary I(1) series, Yt and Xt , such that the
E(Yt – βXt) = 0.
residuals of the regression:
∴ ut = equilibrium error (deviation from long-
Yt = βXt + ut , that is, ut = Yt – βXt run equilibrium).
are stationary, i.e., I(0), then the variables, For the equilibrium to be meaningful,
equilibrium error process must be stationary.
Yt and Xt , are said to be cointegrated.
The system is in long-run equilibrium when
β is called the constant of cointegration, and
E(Yt – βXt) = 0. the equation is called cointegrating regression
∴ ut = equilibrium error (deviation from long- or Cointegrating vector (CV).
run equilibrium).
Vijayamohan: CDS MPhil: Time Series 8 73 05 November 2013 Vijayamohan: CDS MPhil: Time Series 8 74 05 November 2013
Cointegration
Equilibrium Errors: (i.e. ut = Yt - βXt)
Two series Yt and Xt are said to be
cointegrated if:
1. Both are I(d), d ≠ 0 and the same for both
No tendency
to return to series (having the same ‘wave length’); and
zero
ut 2. There is a linear combination of them that
is I(0),
i.e., there exists ut = Yt – βXt that is I(0).
0 time In this light, the regression of these two
Error rarely drifts
from zero
variables, yt = β xt + ut makes sense
(is not spurious).
Vijayamohan: CDS MPhil: Time Series 8 75 05 November 2013 Vijayamohan: CDS MPhil: Time Series 8 76 05 November 2013
Cointegration Testing for Cointegration: Residual based tests
How can we distinguish between a genuine long- (1) After estimating the model, save residuals from
run relationship and a spurious regression? static regression. Consider whether residuals are
We need a test. stationary.
1.0
residuals
0.5
Two categories:
0.0
1. Single equation method: -0.5 0 90
Residual-based test:
10 20 30 40 50 60 70 80 100
1.0
ACF
using (A)DF -statistic: 0.5
Engle and Granger (1987): (A)EG test; 0.0
1. System (Multiple equation) method: -0.5
Johansen and Juselius (1990): JJ test. 0 1 2 3 4 5 6 7 8 9 10 11 12
Vijayamohan: CDS MPhil: Time Series 8 77 05 November 2013 Vijayamohan: CDS MPhil: Time Series 8 78 05 November 2013
13
11/5/2013
Testing for Cointegration: Residual based tests Testing for Cointegration: Residual based tests:
(2): Cointegrating Regression (A)DF test: (Augmented) Engle-Granger test
(Augmented) Engle-Granger test (A)EG test is a test of no-cointegration:
We proceed in three steps: H0: No Cointegration (= unit root in ut)
Step 1: Test that both variables have the same order
of integration, say, that they are both I(1). This can be Acceptance of a unit root in the residuals suggests that
performed with the unit-root tests described before. the residual term is non-stationary, which implies
NO cointegration:
Step 2: Estimate a `long-run relationship by OLS:
α + β xt + u t
yt =α = Not rejecting H0. That is,
Step 3: Extract the residuals of this regression (ut )
If the estimated (A)DF τ–value is more negative
and test for a unit-root in this series, using (A)DF test
(i.e., less) than the critical value at the chosen
statistic: significance level, reject Ho
ut has a unit root ⇒ NO cointegration. = There IS cointegration.
Vijayamohan: CDS MPhil: Time Series 8 79 05 November 2013 Vijayamohan: CDS MPhil: Time Series 8 80 05 November 2013
Testing for Cointegration: Residual based tests: Testing for Cointegration: Residual based tests:
(Augmented) Engle-Granger test (Augmented) Engle-Granger test
Step 1: Test for the order of integration of Xt and Yt: Step 1: Test for the order of integration of Xt and Yt:
Unit-root tests (using Data1) Unit-root tests (using Data1)
The sample is 4 - 300 The sample is 5 - 300
X: ADF tests (T=297, Constant; 5% = -2.87; 1% = -3.45) DX: ADF tests (T=296, Constant; 5% = -2.87; 1% = -3.45)
D-lag t-adf beta Y_1 sigma t-DY_lag t-prob AIC F-prob D-lag t-adf beta Y_1 sigma t-DY_lag t-prob AIC F-prob
2 1.941 1.0006 0.9939 -0.1358 0.8920 0.001096 2 -9.604** 0.057433 1.002 -0.1413 0.8878 0.01713
1 1.942 1.0006 0.9922 0.5127 0.6085 -0.005575 0.8920 1 -11.77** 0.049579 1.000 -0.08180 0.9349 0.01045 0.8878
0 2.028 1.0007 0.9910 -0.01142 0.8692 0 -16.39** 0.045016 0.9985 0.003713 0.9868
Y: ADF tests (T=297, Constant; 5%=-2.87 1%=-3.45) DY: ADF tests (T=296, Constant; 5% = -2.87; 1% = -3.45)
D-lag t-adf beta Y_1 sigma t-DY_lag t-prob AIC F-prob D-lag t-adf beta Y_1 sigma t-DY_lag t-prob AIC F-prob
2 -0.9251 0.99935 0.9635 1.163 0.2459 -0.06091 2 -8.860** 0.099869 0.9527 -1.232 0.2191 -0.08349
1 -0.9554 0.99932 0.9641 -0.9526 0.3416 -0.06304 0.2459 1 -11.60** 0.028890 0.9535 -1.094 0.2749 -0.08507 0.2191
0 -0.9305 0.99934 0.9640 -0.06669 0.3244 0 -18.02** -0.037637 0.9539 -0.08775 0.2589
Vijayamohan: CDS MPhil: Time Series 8 81 05 November 2013 Vijayamohan: CDS MPhil: Time Series 8 82 05 November 2013
Testing for Cointegration: Residual based tests: Testing for Cointegration: Residual based tests:
(Augmented) Engle-Granger test (Augmented) Engle-Granger test
Step 3: Check for unit root in the residuals:
Step 2: Estimate cointegrating regression
Yt = β0 + β1Xt + ut EQ( 6) Modelling Dresiduals by OLS (using Data1.in7)
The estimation sample is: 4 to 300
EQ( 1) Modelling Y by OLS (using Data1)
Coefficient Std.Error t-value t-prob Part.R^2
The estimation sample is: 1 to 300
residuals_1 -1.16140 0.1024 -11.3 0.000 0.5805
Coefficient Std.Error t-value t-prob Part.R^2
sigma 0.545133 RSS 27.6367834
Constant 4.85755 0.1375 35.3 0.000 0.9279
log-likelihood -75.8453 DW 1.95
X 1.00792 0.005081 198. 0.000 0.9975
sigma 0.564679 RSS 30.9296673 Which means ∆ut = -1.161 ut-1 + et
R^2 0.997541 F(1,97) = 3.935e+004 [0.000]** (-11.3)
log-likelihood -82.8864 DW 2.28
no. of observations 99 no. of parameters 2
CRDF test statistic = -11.3 << -3.39 = 1% Critical Value from
MacKinnon.
And save residuals. Hence we reject null of no cointegration between X and Y.
Vijayamohan: CDS MPhil: Time Series 8 83 05 November 2013 Vijayamohan: CDS MPhil: Time Series 8 84 05 November 2013
14
11/5/2013
Testing for Cointegration: Residual based tests: Testing for Cointegration: Residual based tests:
(Augmented) Engle-Granger test (Augmented) Engle-Granger test
Step 3: Check for unit root in the residuals: Step 3: Check for stationarity (unit root) of the residuals:
Residuals
2.5
Unit-root tests (using Data1)
0.0
The sample is 4 - 300 -2.5
0 50 100 150 200 250 300
1
Residuals: ADF tests (T=297, Constant; 5% = -2.87; 1% = -3.39) ACF
0
D-lag t-adf beta Y_1 sigma t-DY_lag t-prob AIC F-prob
2 -9.867** 0.016854 0.9869 0.4090 0.6828 -0.01308 0 5 10
1
1 -10.72** 0.040004 0.9855 -0.2931 0.7696 -0.01924 0.6828 PACF
0
0 -11.33** 1.16140 0.5451 -0.02568 0.8812
0 5 10
Vijayamohan: CDS MPhil: Time Series 8 85 05 November 2013 Vijayamohan: CDS MPhil: Time Series 8 86 05 November 2013
Testing for Cointegration: System method: Testing for Cointegration: System method:
Johansen – Juselius test Johansen – Juselius test
Most popular system method: JJ -test (Johansen
1988; Johansen and Juselius 1990), JJ-test is in the framework of VAR (Sims 1980)
Provides two likelihood ratio (LR) tests. In a VAR, all the variables are endogenous:
(1)The trace test: tests the hypothesis that there With two variables in a model, X and Y: two
are at most r cointegrating vectors, and endogenous variables in VAR;
(2) The maximum eigenvalue test: tests the null That is, two equations:
hypothesis that there are r cointegrating (1): Yt = a + b Xt; and (2): Xt = c + d Yt;
vectors against the hypothesis that there are
r+1 cointegrating vectors. Thus two possible cointegrating vectors
(CVs):
Johansen and Juselius (1990) recommend the
second test as better. We need to identify the CVs.
Vijayamohan: CDS MPhil: Time Series 8 87 05 November 2013 Vijayamohan: CDS MPhil: Time Series 8 88 05 November 2013
Testing for Cointegration: System method: Testing for Cointegration: System method:
Johansen – Juselius test Johansen – Juselius test
Trace test: Maximum eigenvalue test:
This tests Ho: r CVs against H1: r + 1 CVs. Thus
This tests the null hypothesis that there are at
the first row tests Ho: r = 0 against H1: r = 1; if
most r (i.e., 0 ≤ r ≤ n) cointegrating vectors
(CVs) this is significant, Ho is rejected and the next
row is considered.
Ho: r CVs against H1: > r CVs.
A potential problem with the size of these test
Thus the first row tests Ho: r = 0 against statistics in small samples: that is, the JJ
H1: r > 0; if this is significant, Ho is rejected and procedure tends to over-reject the null when it is
the next row is considered. true (Reimers 1992). Hence a small-sample
correction is applied to these statistics, replacing
Thus the rank (r = number of CVs) is chosen as T by T− − np, where T is the number of
the last significant statistic, or as zero if the first observations, n is the number of variables and p
is the lag length of the VAR.
is not CDS
Vijayamohan: significant.
MPhil: Time Series 8 89 05 November 2013 Vijayamohan: CDS MPhil: Time Series 8 90 05 November 2013
15
11/5/2013
Testing for Cointegration: System method: Testing for Cointegration: System method:
Johansen – Juselius test Johansen – Juselius test
Level First-
variables Differenced
variables
Vijayamohan: CDS M Phil: Time Series 7 91 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 92 5 November 2013
Unit-root tests (using Data.in7) Unit-root tests (using Data.in7)
The sample is 1954 (1) - 1992 (3) The sample is 1954 (2) - 1992 (3)
First-Differenced Variables
CONS: ADF tests (T=155; 5%=-1.94 1%=-2.58) DCONS: ADF tests (T=154; 5%=-1.94 1%=-2.58)
D-lag t-adf beta Y_1 sigma t-DY_lag t-prob AIC F-prob D-lag t-adf beta Y_1 sigma t-DY_lag t-prob AIC F-prob
Unit Root Tests on the
Unit Root Tests on the
3 -0.6661 0.99987 2.161 1.241 0.2167 1.567 3 -4.453** 0.43565 2.161 -1.124 0.2628 1.566
2 -0.7434 0.99985 2.165 1.604 0.1109 1.564 0.2167 2 -5.316** 0.37979 2.163 -1.337 0.1831 1.562 0.2628
1 -0.8371 0.99983 2.176 2.392 0.0180 1.568 0.1311 1 -6.812** 0.30387 2.168 -1.648 0.1014 1.561 0.2204
Level Variables
0 -1.033 0.99979 2.209 1.592 0.0219 0 -10.13** 0.19674 2.180 1.565 0.1272
INC: ADF tests (T=155; 5%=-1.94 1%=-2.58) DINC: ADF tests (T=154; 5%=-1.94 1%=-2.58)
D-lag t-adf beta Y_1 sigma t-DY_lag t-prob AIC F-prob D-lag t-adf beta Y_1 sigma t-DY_lag t-prob AIC F-prob
3 -0.5555 0.99983 3.348 0.6839 0.4951 2.442 3 -5.886** -0.0099325 3.355 -0.4643 0.6431 2.446
2 -0.5895 0.99982 3.342 -0.1570 0.8754 2.432 0.4951 2 -6.953** -0.047705 3.346 -0.5646 0.5732 2.435 0.6431
1 -0.5864 0.99982 3.332 -1.052 0.2945 2.420 0.7821 1 -9.144** -0.098829 3.338 0.1576 0.8750 2.424 0.7665
0 -0.5397 0.99984 3.333 2.414 0.6628 0 -13.49** -0.084793 3.328 2.411 0.9059
PRICE INDEX: ADF tests (T=155; 5%=-1.94 1%=-2.58) DPRICE INDEX: ADF tests (T=154; 5%=-1.94 1%=-2.58)
D-lag t-adf beta Y_1 sigma t-DY_lag t-prob AIC F-prob D-lag t-adf beta Y_1 sigma t-DY_lag t-prob AIC F-prob
3 -0.4096 0.99990 3.511 -0.4221 0.6736 2.537 3 -6.972** 0.010731 3.488 1.764 0.0798 2.524
2 -0.3979 0.99991 3.501 -0.5968 0.5516 2.525 0.6736 2 -6.971** 0.13464 3.512 0.3980 0.6912 2.532 0.0798
1 -0.3806 0.99991 3.494 2.468 0.0147 2.515 0.7667 1 -8.126** 0.16195 3.502 0.5392 0.5905 2.520 0.1982
0 -0.4436 0.99989 3.551 2.541 0.0918 0 -10.11** 0.19731 3.494 2.509 0.3162
Vijayamohan: CDS M Phil: Time Series 7 93 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 94 5 November 2013
Testing for Cointegration: System method: Testing for Cointegration: System method:
Johansen – Juselius test in PcGive 10 Johansen – Juselius test in Stata
Testing for cointegration among 3 variables: 3 possible CVs:
. vecrank cons inc Priceindex, trend(none)
I(1) cointegration analysis, 1953 (2) to 1992 (3) Johansen tests for cointegration
eigenvalue loglik for rank Trend: none Number of obs = 157
-1113.071 0 Sample: 1953q3 - 1992q3 Lags = 2
0.27034 -1088.172 1 ------------------------------------------------------------------------------
0.23740 -1066.761 2 5%
0.0049281 -1066.371 3 maximum trace critical
rank parms LL eigenvalue statistic value
rank Trace test [ Prob] Max test [ Prob] Trace test (T-nm) Max test 0 9 -1083.6944 . 41.2984 24.31
(T-nm) 1 14 -1066.35 0.19824 6.6095* 12.53
0 93.40 [0.000]** 49.80 [0.000]** 91.63 [0.000]** 48.85 [0.000]** 2 17 -1063.3777 0.03716 0.6649 3.84
1 43.60 [0.000]** 42.82 [0.000]** 42.77 [0.000]** 42.01 [0.000]** 3 18 -1063.0452 0.00423
2 0.78 [0.377] 0.78 [0.377] 0.77 [0.382] 0.77 [0.382] ------------------------------------------------------------------------------
-
Vijayamohan: CDS M Phil: Time Series 7 95 5 November 2013 Vijayamohan: CDS M Phil: Time Series 7 96 5 November 2013
16
11/5/2013
Testing for Cointegration: System method: Summing Up:
In Stata Johansen – Juselius test in Stata ‘Much Ado about Nothing’?
Statistics → Multivariate Time Series → Cointegrating rank of a
VECM If ut = Yt – βXt is I(0),
yt and xt are cointegrated
(Both I(d), d ≠ 0 )
(input all the
variables in the and the regression yt = β xt + ut
column for
dependent is not spurious.
variables)
Vijayamohan: CDS M Phil: Time Series 7 97 5 November 2013 Vijayamohan: CDS MPhil: Time Series 8 98 05 November 2013
Summing Up:
‘Full of Sound and Fury; Signifying Nothing’?
Most macroeconomic variables : non stationary.
ut ~ I(0) ⇒ ut is white noise Non-stationarity problem:
Consequence: Spurious Regression
⇒ OLS assumptions satisfied!
Approaches: Classical Modern
So just see if OLS assumptions are satisfied!
Diagnosis: ACF tests Unit root tests
No need for the pre-tests
Solution: Differencing Differencing/
Cointegration
(of unit root and cointegration)!
Estimation: ARIMA (Box-Jenkins)/ VAR (Sims 1980)/
MARIMA (V)ECM (Sargan 1964;
(See my book ‘Econometrics of Electricity Demand: (Harvey 1997) Davidson et al. 1978)/
Questioning the Tradition’ 2010, Lambert Academic GETS (Hendry 1987)
Publishing, Germany)
Vijayamohan: CDS MPhil: Time Series 8 99 05 November 2013 Vijayamohan CDS Time Series: Introduction 100 5 November 2013
5 November 2013 Vijayamohan: CDS M Phil: Time 101
Series 7
17
11/7/2013
Time Series Econometrics
Time Series
10
Methodology
Vijayamohanan Pillai N
7 November 2013 Vijayamohan: CDS MPhil: Time 1 7 November 2013 Vijayamohan: CDS MPhil: Time 2
Series 1 Series 1
Time Series: Methodology GETS Approach
Three alternative approaches: Suppose, the theory implies that there is a
(1): general to specific model (GETS): relationship between consumption (Ct) and
LSE Approach income (Yt ) i.e., say
Hendry, Pagan and Sargan (1984) and Ct = α0 + α1Yt (1)
Hendry (1987). Since this is an equilibrium relationship,
(2): vector autoregressions model (VAR): a dynamic adjustment equation can be searched by
Sims (1980): starting first with a very general and elaborate
dominant approach in the USA; and specification.
(3): vector error correction model (VECM): This initial general specification is termed
Follows the Granger Representation the general unrestricted model (GUM).
theorem (Engle and Granger 1987)
Thursday, November Vijayamohan: CDS M Phil 103: Time 3 11/7/2013 10:19:14 AM Vijayamohan: M Phil: Time Series 4
07, 2013 Series
GETS Approach GETS Approach
A good GUM for the consumption equation in general:
A good GUM for the consumption equation:
n m
∆C t = ∑ β ci ∆C t − i + ∑ β yi ∆Yt − i + λ (C t −1 − kYt −1 ) + ε t
Ct = αCt–1 + β0Yt + β 1Yt–1 + εt ; i i
where Yt’= (Y1t, Y2t,…., Ykt), Include enough lagged variables so that there is no
serial correlation in the residuals of the GUM.
∆Ct = (α−1)Ct–1 + β0 ∆Yt + (β0 +β1)Yt–1 + εt ; or Finally, a parsimonious version of equation (2) is
developed,, by deleting the insignificant variables and
developed
∆ C t = β 0 ∆ Y t + λ [ C t – 1 − k Y t – 1] + ε t ;
imposing constraints on the estimated coefficients.
coefficients.
where k = (β0 +β1)/ (1− α) and λ = (α−1)
GETS is thus a highly empirical approach.
Thursday, November CDS MPhil FQ 5 Thursday, November CDS MPhil FQ 6
07, 2013 Vijayamohan 07, 2013 Vijayamohan
1
11/7/2013
Consumption Income Consumption
Inflation
Income
Thursday, November Vijayamohan: CDS M Phil 103: Time 7 Thursday, November Vijayamohan: CDS M Phil 103: Time 8
07, 2013 Series 07, 2013 Series
Thursday, November Vijayamohan: CDS M Phil 103: Time 9 Thursday, November Vijayamohan: CDS M Phil 103: Time 10
07, 2013 Series 07, 2013 Series
Residual
Analysis
ACF PACF
Thursday, November Vijayamohan: CDS M Phil 103: Time 11 Thursday, November Vijayamohan: CDS M Phil 103: Time 12
07, 2013 Series 07, 2013 Series
2
11/7/2013
Consumption Income
Inflation
Thursday, November Vijayamohan: CDS M Phil 103: Time 13 11/7/2013 10:19:14 AM Vijayamohan: M Phil: Time Series 14
07, 2013 Series
Thursday, November 15 Thursday, November Vijayamohan: CDS M Phil 103: Time 16
Vijayamohan: CDS M Phil 103: Time
07, 2013 07, 2013 Series
Series
PcGive: Test: Dynamic Analysis PcGive: Test: Dynamic Analysis
Thursday, November CDS MPhil FQ 17 Thursday, November Vijayamohan: CDS M Phil 103: Time 18
07, 2013 Vijayamohan 07, 2013 Series
3
11/7/2013
Residual GETS Approach
Analysis Note: the equilibrium theoretical consumption
relationship can be recovered from equation(2)
n m
∆Ct = ∑ βci ∆Ct −i + ∑ β yi∆Yt −i + λ (Ct −1 − kYt −1 ) + ε t (2)
i i
by imposing the equilibrium condition that all the
PACF changes in the variables are zero, i.e., from the term in
ACF
the last part in (2), since in equilibrium, equation (2) will
be
0 = 0 + 0 + λ(Ct – 1 – kYt – 1) + 0
∴ C* = kY*.
Thursday, November Vijayamohan: CDS M Phil 103: Time 19 11/7/2013 10:19:14 AM Vijayamohan: M Phil: Time Series 20
07, 2013 Series
GETS Approach VAR Models
The expression in the lagged level variables, Developed as an alternative to the large scale
econometric models based on the Cowles Commission
λ(Ct – 1 – kYt – 1) in equation (2)
approach.
n m
∆C t = ∑ β ci ∆ C t − i + ∑ β yi ∆ Y t − i + λ (C t − 1 − kY t − 1 ) + ε t Sims (1980) argued:
i i
is known as the error correction term and models that • The classification of variables into endogenous and
exogenous,
include it are known as the error correction models
• The constraints implied by the traditional theory on the
(ECM).
structural parameters, and
It implies that departures from the equilibrium position in • The dynamic adjustment mechanisms used in the large
the immediate past period will be offset in the current scale models, are all arbitrary and too restrictive.
period by λ proportion. ∴ Include all variables as endogenous.
11/7/2013 10:19:14 AM
Note that λ should be negative.
Vijayamohan: M Phil: Time Series 21 Thursday, November Vijayamohan: CDS M Phil 103: Time 22
07, 2013 Series
VAR Models VAR Models
For a set of n time series variables Consider a two-variable VAR(1) with k =2.
y t = ( y1t , y 2t , ..., y kt )' y t = b12 zt + c11 y t −1 + c12 zt −1 + ε yt
a VAR model of order p (VAR(p)) can be written zt = b21y t + c21y t −1 + c22 zt −1 + ε zt
as:
with ε it ~ i .i .d(0 , σ εi ) and cov( ε y , ε z ) = 0
2
y t = A1y t −1 + A2 y t −2 + ... + Ap y t − p + ε t
where the Ai’s are (n x n) coefficient matrices
In matrix form:
and
ε t = ( ε 1t , ε 2t ,..., ε kt )' 1 − b12 y t c11 c12 y t −1 ε yt
= +
− b21 1 zt c21 c22 zt −1 ε zt
is an unobservable i.i.d. zero mean error term.
Thursday, November Vijayamohan: CDS M Phil 103: Time 23 Thursday, November Vijayamohan: CDS M Phil 103: Time 24
07, 2013 Series 07, 2013 Series
4
11/7/2013
VAR Models VAR Models
1 − b12 y t c11 c12 y t −1 ε yt y t a11 a12 y t −1 e1t
− b21
=
1 zt c21
+
c22 zt −1 ε zt = +
zt a21 a22 zt −1 e2t
BX t = Γ1 X t −1 + ε t
These error terms are composites of the structural
Structural VAR (SVAR) or the Primitive System innovations from the primitive system.
1 1 b12
To normalize the LHS vector, we need to multiply et = B −1ε t B −1 =
(1 − b21 b12 ) b21 1
the equation by inverse B:
B −1BX t = B −1Γ1 X t −1 + B −1ε t
e1t 1 1 b12 ε yt
X t = A1 X t −1 + et e = 1 ε zt
2t (1 − b21 b12 ) b21
VAR in standard form (unstructured VAR: UVAR).
Thursday, November Vijayamohan: CDS M Phil 103: Time 25 Thursday, November Vijayamohan: CDS M Phil 103: Time 26
07, 2013 Series 07, 2013 Series
VAR Models VAR Models
ε yt + b12ε zt But covariances are not zero:
e1t =
∆ ∆ = 1 − b21b12
Cov ( e1t , e2t ) = E ( e1t e2t ) =
b21ε yt + ε zt
e2t =
∆ E [( ε yt + b12ε zt )( ε zt + b21ε yt )]
=
E ( eit ) = 0 ∆2
( b12σ z2 + b21σ y2 )
≠0
E(ε 2
+b ε )
2 2
σ +b σ
2 2 2 ∆2
Var ( e1t ) = E ( e ) = =
2 yt 12 zt y 12 z
1t
∆2 ∆2
So the shocks in a standard VAR are correlated.
The only way to remove the correlation and make
time independent, and the same is true for the the covar = 0 is to assume that the
other one. contemporaneous effects are zero: b12 = b21 = 0
y t = b12 zt + c11y t −1 + c12 zt −1 + ε yt zt = b21y t + c21 y t −1 + c22 zt −1 + ε zt
Thursday, November Vijayamohan: CDS M Phil 103: Time 27 Thursday, November Vijayamohan: CDS M Phil 103: Time 28
07, 2013 Series 07, 2013 Series
y t a11 a12 y t −1 e1t
VAR / GETS / VECM z = a +
VAR Models t 21 a22 zt −1 e2t
The simple VAR models do not identify structural
Same regressors for all equations; coefficients (hence ‘Structural VAR models’)
equations;
so model estimation is straightforward: nor do they take seriously the relevance of unit root
straightforward:
tests.
ML estimator → OLS estimator for each equation.
In GETS, although there is some awareness of the unit
This property → popularity of VAR models.
models. root characteristics of the variables, the crucial
theoretical relationship, in the error correction part, is
Note: All variables in the reduced form equations are
specified in the levels of the variables.
endogenous,
endogenous, and hence the equations can be seen as
the basic ARDL formulation of a simple VAR model.
model. In contrast, VECM, like VAR, treats all variables as
endogenous,
but limits the number of variables to those relevant for
a particular theory.
Thursday, November Vijayamohan: CDS M Phil 103: Time 29 Thursday, November Vijayamohan: CDS M Phil 103: Time 30
07, 2013 Series 07, 2013 Series
5
11/7/2013
VECM
VECM
Granger Representation Theorem (Engle and Granger
This method developed by Johansen (1988) is 1987):
undoubtedly the most widely used method in applied
work. If a set of variables are cointegrated,
cointegrated, then there
exists a VARMA representation for them and an
Models using this approach are also known as ‘error-
error-correcting’ mechanism (ECM); for example,
Cointegrating VAR (CIVAR) models.
If Yt and Xt are both I(1) and have constant means and
VECM can be seen as scaled down (reduced form)VAR are cointegrated,
cointegrated, then there exists an ECM,
ECM, (with the
model in which the structural coefficients are identified. equilibrium error Ut = Yt – β Xt ), of the form:
form:
lagged{∆Yt, ∆Xt} + θ (L)ε1t,
The theoretical basis of VECM:
∆Yt = –λ1 Ut–1+ lagged{∆
Granger Representation Theorem. Where θ (L) is a finite polynomial in lag operator L and
ε1t is a white noise.
Thursday, November Vijayamohan: CDS M Phil 103: Time 31 Thursday, November Vijayamohan: CDS M Phil 103: Time 32
07, 2013 Series 07, 2013 Series
VECM VECM in Stata
lagged{∆Yt, ∆Xt} + θ (L)ε1t,
In ∆Yt = –λ1Ut–1+ lagged{∆
Statistics > Multivariate time series > Vector error-
period, Ut–1,
The ‘equilibrating’ error in the previous period, correction model (VECM)
captures the adjustment towards long-
long-run equilibrium,
equilibrium,
and the expected –ve sign ⇒ error would correct in the
long-run.
long-run.
Note: No feedback assumed from Yt to Xt.
If feedback assumed from Yt to Xt, an additional
equation:
lagged{∆Yt, ∆Xt} + θ (L)ε2t,
∆Xt = –λ2Ut–1+ lagged{∆
where ε2t is white noise.
noise. Level variables
Thursday, November Vijayamohan: CDS M Phil 103: Time 33 7 November 2013 Vijayamohan: CDS M Phil: Time 34
07, 2013 Series Series 7
VECM in Stata VECM in Stata
provides information about the sample, the model fit,
and the identification of the parameters in the The main estimation table
cointegrating equation. the estimates of the short-run parameters, + their
standard errors and confidence intervals.
Thursday, November Vijayamohan: CDS M Phil 103: Time 35 Thursday, November Vijayamohan: CDS M Phil 103: Time 36
07, 2013 Series 07, 2013 Series
6
11/7/2013
VECM in Stata VECM in Stata
. predict ce, ce
. line ce time
10
Predicted cointegrated equation
-5 0 5
The second estimation table
the estimates of the parameters in the cointegrating
-10
equation, + their standard errors and confidence
intervals.
-15
1950q1 1960q1 1970q1 1980q1 1990q1
Thursday, November Vijayamohan: CDS M Phil 103: Time 37 Thursday, November time
Vijayamohan: CDS M Phil 103: Time 38
07, 2013 Series 07, 2013 Series
VECM
VECM
Remember: VECM, like VAR, treats all variables as
endogenous, Thus we need to test for direction of feedback:
endogenous,
but limits the number of variables to those relevant for a From Xt to Yt: whether Yt is exogenous to Xt;
particular theory. From Yt to Xt: whether Xt is exogenous to Yt;
For example, with two variables: Yt and Xt,
two endogenous variables,
variables, two possible CVs; That is, we need to test for ‘exogeneity’ of variables:
If our equation of interest is Yt = f(Xt) only,
only, that is,
If No feedback assumed from Yt to Xt, or Xt = f(Yt), we This exogeneity test is:
need to consider and estimate only one, first, equation,
equation, Granger non-‘causality’ test (Granger 1969):
unlike in VAR.
Thursday, November Vijayamohan: CDS M Phil 103: Time 39 Thursday, November Vijayamohan: CDS M Phil 103: Time 40
07, 2013 Series 07, 2013 Series
VECM: Granger non-‘causality’ test: VECM: Granger non-‘causality’ test:
an Exogeneity test an Exogeneity test
Consider the following equations:
equations: Yt = ΣαiYt−i + Σβ iXt−i + e1t , (1)
Yt = ΣαiYt−i + ΣβiXt−i + e1t , (1) We have the following implication:
implication:
Xt = ΣγiYt−i + ΣδiXt−i + e2t , ( 2)
X does not ‘Granger-
‘Granger-cause’ Y if, and only if, βi ≡ 0,
if,
where the summations are for some lag length k, for all i, as a group;
group;
and e1t and e2t are independently distributed
white noises.
noises. Thus the measure of linear feedback from X to Y is zero
(Geweke 1982)
1982).
(1) hypothesises that the current value of Y is
That is, the past values of X do not help to predict Y.
related to past values of Y itself and those of X,
In this case, Y is exogenous with respect to X
while (2) postulates a similar behaviour for X.
(Engle et al.
al. 1983)
1983).
Thursday, November Vijayamohan: CDS M Phil 103: Time 41 Thursday, November Vijayamohan: CDS M Phil 103: Time 42
07, 2013 Series 07, 2013 Series
7
11/7/2013
VECM: Granger non-‘causality’ test: VECM: Granger non-‘causality’ test:
an Exogeneity test an Exogeneity test
Xt = ΣγiYt−i + ΣδiXt−i + e2t , ( 2) Yt = ΣαiYt−i + ΣβiXt−i + e1t , (1)
Similarly, we have the following implication:
implication:
Xt = ΣγiYt−i + ΣδiXt−i + e2t , ( 2)
‘Granger-cause’ X, if, and only if, γi ≡ 0 for all
Y does not ‘Granger-
i as a group;
group; the measure of linear feedback from Y to X
If the lagged terms have significant non-
non-zero
is zero.
zero.
coefficients,
coefficients,
That is, the past values of Y fail to help predict X.
Here X is exogenous with respect to Y. then there is ‘causality’ or feedback in both
directions.
directions.
Thursday, November Vijayamohan: CDS M Phil 103: Time 43 Thursday, November Vijayamohan: CDS M Phil 103: Time 44
07, 2013 Series 07, 2013 Series
Granger non-‘causality’ test in STATA Granger non-‘causality’ test in STATA
In Stata, first run a VAR model:
Statistics >
Multivariate time
series > Vector
autoregression
(VAR)
Level
variables
7 November 2013 Vijayamohan: CDS M Phil: Time 45 Thursday, November Vijayamohan: CDS M Phil 103: Time 46
Series 7 07, 2013 Series
Granger non-‘causality’ test in STATA Granger non-‘causality’ test in STATA
Thursday, November Vijayamohan: CDS M Phil 103: Time 47 Thursday, November Vijayamohan: CDS M Phil 103: Time 48
07, 2013 Series 07, 2013 Series
8
11/7/2013
Granger non-‘causality’ test in STATA Granger non-‘causality’ test in STATA
In Stata H0: x does not Granger-cause y.
Statistics > Multivariate time series > VAR
diagnostics and tests > Granger causality tests
Y X
H0: x does not Granger-cause y.
Thursday, November Vijayamohan: CDS M Phil 103: Time 49 Thursday, November Vijayamohan: CDS M Phil 103: Time 50
07, 2013 Series 07, 2013 Series
VECM: Granger non-‘causality’ test: VECM: Granger non-‘causality’ test:
an Exogeneity test an Exogeneity test
This cointegrtaed system can be written in terms of ECM
‘Granger causality’:
causality’: concerned with only short run as:
as:
forecastability, while
forecastability,
Cointegration:
Cointegration: concerned with long run equilibrium;
equilibrium;
∆yt = −λ1ut−1 + lagged{∆yt , ∆xt} + θ1(L)ε1t, ( 1)
Error correction model (ECM) brings the two
(different) concepts together.
together.
∆xt = −λ2ut−1 + lagged{∆yt , ∆xt} + θ2(L)ε2t, ( 2)
Suppose Yt and Xt are both I(1) series and they are
cointegrated such that ut = Yt − βXt is I(0).
where θ1(L)ε1t and θ2(L)ε2t are finite order moving
averages and one of λ1, λ2 ≠ 0.
Thursday, November Vijayamohan: CDS M Phil 103: Time 51 Thursday, November Vijayamohan: CDS M Phil 103: Time 52
07, 2013 Series 07, 2013 Series
VECM: Granger non-‘causality’ test: VECM: Granger non-‘causality’ test:
an Exogeneity test an Exogeneity test
∆yt = −λ1ut−1 + lagged{∆yt , ∆xt} + θ1(L)ε1t, ( 1)
∆yt = −λ1ut−1 + lagged{∆yt , ∆xt} + θ1(L)ε1t, ( 1)
∆xt = −λ2ut−1 + lagged{∆yt , ∆xt} + θ2(L)ε2t, (2)
∆xt = −λ2ut−1 + lagged{∆yt , ∆xt} + θ2(L)ε2t, (2)
That is, the coefficient of EC contains information on
term, Ut−1,
In the ECM, the error correction term,
whether the past values of the variables ‘affect’
affect’ the
‘Granger causes’ ∆Yt or ∆Xt (or both). current values of the variable under consideration.
As Ut−1 itself is a function of Yt−1 and Xt−1, This then implies that there must be some ‘Granger
either Xt is ‘Granger
‘Granger caused’ by Yt−1 or Yt by Xt−1. causality’ between the two series in order to induce them
towards equilibrium.
equilibrium.
Thursday, November Vijayamohan: CDS M Phil 103: Time 53 Thursday, November Vijayamohan: CDS M Phil 103: Time 54
07, 2013 Series 07, 2013 Series
9
11/7/2013
Granger non-‘causality’ test:
Granger non-‘causality’ test: ‘Causality’- test ?
‘Causality’- test ?
Adrian Pagan (1989) on Granger causality:
First suggested by Wiener (1956):
“There was a lot of high powered analysis of this
More properly called Wiener-Granger non-‘causality’
test. topic, but I came away from a reading of it with the
Economists (e.g., Zellner 1979) and even philosophers feeling that it was one of the most unfortunate
(e.g., Holland 1986) question the very term ‘causality’: turnings for econometrics in the last two decades,
To mean ‘cause-effect’ relationship, when there is only and it has probably generated more nonsense results
temporal lead-lag relationship? than anything else during that time.”
Not ‘causality’ but ‘precedence’ as suggested by
Pagan, A.R. (1989), '20 Years After: Econometrics
Edward Leamer.
1966-1986,' in B. Cornet and H. Tulkens (eds).,
Unfortunately several studies to infer ‘cause-effect’ Contributions to Operations Research and
relationship! Econometrics, The XXth Anniversary of CORE,
Thursday, November Vijayamohan: CDS M Phil 103: Time 55 Thursday, November Vijayamohan: CDS M Phil 103: Time 56
07, 2013 Series (Cambridge,
07, 2013 Ma., MIT Press). Series
Consider a (linear) filter – Consider a (linear) filter –
Yt = ∑Φ
k
k X t −k Yt = ∑Φ
k
k X t −k
࣐k = response in the output at time t to a unit
pulse in the input at time t – k:
The impact multiplier is:
one-period multiplier or transient response or
∂Yt impulse response:
∂X t
∂Yt ∂Yt + k
=
∂X t − k ∂X t
7 November 2013 Vijayamohan: CDS MPhil: Time 57 7 November 2013 Vijayamohan: CDS MPhil: Time 58
Series 2 Series 2
The first order Autorregresive process AR(1):
Given the linear filter
ACF = ρk = γk / γ0 = Φk, k = 0, 1, 2, …..
Yt = ∑Φ
k
k X t −k
Impulse responses
Impulse response function = ACF ↓, as k ↑, since |Φ| < 1: Sign of stationarity.
The time path of all the periodical impulse +Φ: direct convergence;
responses: negative Φ: oscillatory convergence.
∂Yt ∂Yt + k
=
∂X t − k ∂X t
7 November 2013 Vijayamohan: CDS MPhil: Time 59 7 November 2013 Vijayamohan: CDS MPhil: Time 60
Series 2 Series 3
10
11/7/2013
In PcGive
First run a VAR
Thursday, November Vijayamohan: CDS M Phil 103: Time 61 Thursday, November Vijayamohan: CDS M Phil 103: Time 62
07, 2013 Series 07, 2013 Series
SYS( 1) Estimating the system by OLS (using Data.in7)
The estimation sample is: 1953 (3) to 1992 (3)
URF equation for: CONS URF equation for: INFLAT
Coefficient Std.Error t-value t-prob Coefficient Std.Error t-value t-prob
CONS_1 0.720835 0.1408 5.12 0.000 CONS_1 0.00227184 0.02679 0.0848 0.933
CONS_2 0.114970 0.1219 0.943 0.347 CONS_2 0.000632921 0.02319 0.0273 0.978
INC_1 0.119269 0.08579 1.39 0.166 INC_1 0.0103773 0.01632 0.636 0.526
INC_2 0.028640 0.07645 0.375 0.708 INC_2 -0.0157283 0.01455 -1.08 0.281
INFLAT_1 -1.14639 0.3395 -3.38 0.001 INFLAT_1 1.54190 0.06460 23.9 0.000
INFLAT_2 -0.222212 0.3554 -0.625 0.533 INFLAT_2 -0.648024 0.06764 -9.58 0.000
Constant U 14.1709 15.12 0.937 0.350 Constant U 2.41866 2.877 0.841 0.402
sigma = 1.87171 RSS = 525.4948372 sigma = 0.356165 RSS = 19.02800862
URF equation for: INC
Coefficient Std.Error t-value t-prob
CONS_1 -0.201124 0.2311 -0.870 0.386
CONS_2 0.294033 0.2001 1.47 0.144 Vector Portmanteau(12): 78.1333
Vector AR 1-5 test: F(45,395)= 0.84719 [0.7481]
INC_1 0.847601 0.1408 6.02 0.000
Vector Normality test: Chi^2(6) = 9.1045 [0.1678]
INC_2 -0.0203814 0.1255 -0.162 0.871 Vector hetero test: F(72,723)= 1.2828 [0.0642]
INFLAT_1 0.470490 0.5574 0.844 0.400 Vector hetero-X test: F(162,695)= 1.1138 [0.1819]
INFLAT_2 -1.26264 0.5836 -2.16 0.032
Constant U 73.8946 24.83 2.98 0.003
sigma = 3.0729 RSS = 1416.406296
Thursday, November Vijayamohan: CDS M Phil 103: Time 63 Thursday, November Vijayamohan: CDS M Phil 103: Time 64
07, 2013 Series 07, 2013 Series
Quarterly data Response of 1953 (1) to 1992 (3) Accumulated Response of
Consumption Income Inflation Consumption Income Inflation
1.0 0.4 0.050 4
CONS (CONS eqn) INC (CONS eqn) INFLAT (CONS eqn) 6 cum CONS (CONS eqn) cum INC (CONS eqn) cum INFLAT (CONS eqn)
0.75
Shock from0.5 Shock from 3
0.2
0.025 4 0.50
Consumption Consumption 2
0.0 0.25
0.0 2 1
0.000
0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150
1.0
0.4 CONS (INC eqn) INC (INC eqn) INFLAT (INC eqn) 10 cum CONS (INC eqn) 10 cum INC (INC eqn)
0.0 cum INFLAT (INC eqn)
0.00
Shock from Shock from -0.1
0.5
Income 0.2
-0.02
Income 5 5 -0.2
-0.3
0.0 0.0
0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150
1.0 0 0 10.0
CONS (INFLAT eqn) 0 INC (INFLAT eqn) INFLAT (INFLAT eqn) cum CONS (INFLAT eqn) cum INC (INFLAT eqn) cum INFLAT (INFLAT eqn)
0.0
-50 7.5
Shock from-2.5 0.5 Shock from -50
-2 5.0
Inflation Inflation -100
-5.0 2.5
-4 0.0 -150 -100
0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150
Thursday, November Vijayamohan: CDS M Phil 103: Time 65 Thursday, November Vijayamohan: CDS M Phil 103: Time 66
07, 2013 Series 07, 2013 Series
11
11/7/2013
In Stata
First run a VAR or VECM
. var cons inc inflat, noconstant lags(1/2)
Thursday, November Vijayamohan: CDS M Phil 103: Time 67 Thursday, November Vijayamohan: CDS M Phil 103: Time 68
07, 2013 Series 07, 2013 Series
In Stata
First run a VAR or VECM
. var cons inc inflat, noconstant lags(1/2)
. irf create order1, step(10) set(myirf1)
(file myirf1.irf created)
(file myirf1.irf now active)
(file myirf1.irf updated)
. irf graph oirf, impulse(inc) response(cons)
Thursday, November Vijayamohan: CDS M Phil 103: Time 69 Thursday, November Vijayamohan: CDS M Phil 103: Time 70
07, 2013 Series 07, 2013 Series
order1, inc, cons
-.5
-1
-1.5
-2
0 5 10
step
95% CI orthogonalized irf
Graphs by irfname, impulse variable, and response variable
Thursday, November Vijayamohan: CDS M Phil 103: Time 71 Thursday, November Vijayamohan: CDS M Phil 103: Time 72
07, 2013 Series 07, 2013 Series
12
11/7/2013
1. All the series are I(0)
3. All the series are integrated of the same order, and they
are cointegrated.
Simply model the data in their levels, using OLS
estimation.
estimate two types of models:
(i) An OLS regression model using the levels of the data.
2. All the series are integrated of the same order (e.g., I(1)), → the long-run equilibrating relationship between the
but not cointegrated. variables.
(ii) An error-correction model (ECM), estimated by OLS.
Just difference (appropriately)each series, and
estimate a standard regression model using OLS. → the short-run dynamics of the relationship between
the variables.
Thursday, November Vijayamohan: CDS M Phil 103: Time 73 Thursday, November Vijayamohan: CDS M Phil 103: Time 74
07, 2013 Series 07, 2013 Series
4. Finally, ARDL= Autoregressive-Distributed Lag.
Some of the variables in question may be stationary, in use for decades,
but in more recent times provide a very valuable
some may be I(1) vehicle for testing for the presence of long-run
relationships between economic time-series.
and there may be cointegration among some of the
I(1) variables.
Thursday, November Vijayamohan: CDS M Phil 103: Time 75 Thursday, November Vijayamohan: CDS M Phil 103: Time 76
07, 2013 Series 07, 2013 Series
In its basic form, an ARDL regression model : Pesaran MH and Shin Y. 1999.
(MARMA model) “An autoregressive distributed lag modelling
approach to cointegration analysis.”
yt = β0 + β1yt-1 + .......+ βkyt-p
+ α0xt + α1xt-1 + α2xt-2 + ......... + αqxt-q + εt Chapter 11 in
Econometrics and Economic Theory in the 20th
where εt is a random "disturbance" term. Century: The Ragnar Frisch Centennial Symposium,
"autoregressive“ = yt is "explained (in part) by lagged Strom S (ed.). Cambridge University Press:
values of itself. Cambridge.
"distributed lag" = successive lags of the explanatory
variable "x" .
Sometimes, the current value of xt is excluded
Thursday, November Vijayamohan: CDS M Phil 103: Time 77 Thursday, November Vijayamohan: CDS M Phil 103: Time 78
07, 2013 Series 07, 2013 Series
13
11/7/2013
Two sets of asymptotic critical values :
“Bounds Testing Approaches to the Analysis one when all regressors are purely I(1)
of Level Relationships” and the other if they are all purely I(0).
M. Hashem Pesaran, Yongcheol Shin and Richard J. Smith
Journal of Applied Econometrics
16: 289–326 (2001) These two sets of critical values provide critical value
bounds for all classifications of the regressors into
purely I(1), purely I(0) or mutually cointegrated.
“… a new approach to testing for the existence of a
relationship between variables in levels
Accordingly, various bounds testing procedures are
which is applicable irrespective of
proposed.
whether the underlying regressors are
purely I(0), purely I(1) or mutually cointegrated.”
Thursday, November Vijayamohan: CDS M Phil 103: Time 79 Thursday, November Vijayamohan: CDS M Phil 103: Time 80
07, 2013 Series 07, 2013 Series
For instance:
The ARDL / Bounds Testing methodology of
It can be used with a mixture of I(0) and I(1)
Pesaran and Shin (1999) and Pesaran et al.
data.
(2001)
has a number of features that many
It involves just a single-equation set-up,
researchers feel give it some advantages over
making it simple to implement and interpret.
conventional cointegration testing.
Different variables can be assigned different
lag-lengths as they enter the model.
Thursday, November Vijayamohan: CDS M Phil 103: Time 81 Thursday, November Vijayamohan: CDS M Phil 103: Time 82
07, 2013 Series 07, 2013 Series
Step 1:
A conventional ECM for cointegrated data :
use the ADF/PP/KPSS tests to check that none of the
∆yt = β0 + Σ βi∆yt-i + Σγj∆x1t-j + Σδk∆x2t-k + φzt-1 + et ; series are I(2).
The ranges of summation : from 1 to p, 0 to q1, and 0 to q2 Step 2:
respectively. Formulate the following model:
z, the "error-correction term", is the OLS residuals series ∆yt = β0 + Σ βi∆yt-i + Σγj∆x1t-j + Σδk∆x2t-k
from the long-run "cointegrating regression",
+ θ0yt-1 + θ1x1t-1 + θ2 x2t-1 + et ;
yt = α0 + α1x1t + α2x2t + vt
Thursday, November Vijayamohan: CDS M Phil 103: Time 83 Thursday, November Vijayamohan: CDS M Phil 103: Time 84
07, 2013 Series 07, 2013 Series
14
11/7/2013
∆yt = β0 + Σ βi∆yt-i + Σγj∆x1t-j + Σδk∆x2t-k ∆yt = β0 + Σ βi∆yt-i + Σγj∆x1t-j + Σδk∆x2t-k
+ θ0yt-1 + θ1x1t-1 + θ2 x2t-1 + et ; + θ0yt-1 + θ1x1t-1 + θ2 x2t-1 + et ;
almost like a traditional ECM. Step 3:
The error-correction term, zt-1 replaced with the terms
yt-1, x1t-1, and x2t-1 from yt = α0 + α1x1t + α2x2t + vt The ranges of summation :
from 1 to p, 0 to q1, and 0 to q2 respectively
the lagged residuals series would be
zt-1 = (yt-1 - a0 - a1x1t-1 - a2x2t-1), Maximum lags are determined by using one or more of
where the a's are the OLS estimates of the α's. the "information criteria" : AIC, SC (BIC), HQ, etc.
"unrestricted ECM", or an "unconstrained ECM". Remember: Schwarz (Bayes) criterion (SC) is a
Pesaran et al. (2001) call this a "conditional ECM". consistent model-selector.
Thursday, November Vijayamohan: CDS M Phil 103: Time 85 Thursday, November Vijayamohan: CDS M Phil 103: Time 86
07, 2013 Series 07, 2013 Series
Step 4:
A key assumption in the ARDL / Bounds Testing Step 5:
methodology of Pesaran et al. (2001) : We have a model with an autoregressive structure,
so we have to be sure that
The errors of the equation must be serially independent. the model is "dynamically stable".
Once an apparently suitable version of the equation Check that
has been estimated, use the LM test to test all the
the null hypothesis that associated with the model
the errors are serially independent, .
against the alternative hypothesis that
the errors are (either) AR(m) or MA(m), for m = 1, 2, 3, ...
Thursday, November Vijayamohan: CDS M Phil 103: Time 87 Thursday, November Vijayamohan: CDS M Phil 103: Time 88
07, 2013 Series 07, 2013 Series
Step 6: "Bounds Testing"
Step 6:
Now perform the " " yt = β0 + Σ βi∆yt-i + Σγj∆x1t-j + Σδk∆x2t-k
+ θ0yt-1 + θ1x1t-1 + θ2 x2t-1 + et ;
∆yt = β0 + Σ βi∆yt-i + Σγj∆x1t-j + Σδk∆x2t-k
+ θ0yt-1 + θ1x1t-1 + θ2 x2t-1 + et ; As in conventional cointegration testing,
we're testing for
Do a "F-test" of the hypothesis, between the variables.
H0: θ0 = θ1 = θ2 = 0 ;
against the alternative that H0 is not true. This absence coincides with zero coefficients for
yt-1, x1t-1 and x2t-1 in the equation:
H0: θ0 = θ1 = θ2 = 0
A rejection of H0 implies that we have a long-run
relationship.
Thursday, November Vijayamohan: CDS M Phil 103: Time 89 Thursday, November Vijayamohan: CDS M Phil 103: Time 90
07, 2013 Series 07, 2013 Series
15
11/7/2013
Exact critical values for the F-test not available for an
arbitrary mix of I(0) and I(1) variables. If the computed F-statistic falls below the lower bound
Pesaran et al. (2001) supply bounds on the critical values conclude : the variables are I(0),
for the asymptotic distribution of the F-statistic. so no cointegration possible, by definition.
lower and upper bounds on the critical values. If the F-statistic exceeds the upper bound,
conclude : we have cointegration.
the lower bound is based on the assumption that all of
the variables are I(0), Finally, if the F-statistic falls between the bounds,
and the upper bound : all the variables are I(1). the test is inconclusive.
Thursday, November Vijayamohan: CDS M Phil 103: Time 91 Thursday, November Vijayamohan: CDS M Phil 103: Time 92
07, 2013 Series 07, 2013 Series
Step 7:
If the bounds test proves cointegration, Step 8:
estimate the long-run equilibrium relationship between
the variables: “Extract" long-run effects from the unrestricted ECM.
yt = α0 + α1x1t + α2x2t + vt ; ∆yt = β0 + Σ βi∆yt-i + Σγj∆x1t-j + Σδk∆x2t-k
+ θ0yt-1 + θ1x1t-1 + θ2 x2t-1 + et ;
as well as the usual ECM:
at a long-run equilibrium, ∆yt = 0, ∆x1t = ∆x2t = 0,
∆yt = β0 + Σ βi∆yt-i + Σγj∆x1t-j + Σδk∆x2t-k + φzt-1 + et ;
the long-run coefficients for x1 = -(θ1 / θ0)
where zt-1 = (yt-1 -a0 - a1x1t-1 - a2x2t-1),
and the a's are the OLS estimates of the α's . and x2 = -(θ2 / θ1).
Thursday, November Vijayamohan: CDS M Phil 103: Time 93 Thursday, November Vijayamohan: CDS M Phil 103: Time 94
07, 2013 Series 07, 2013 Series
Thursday, November Vijayamohan: CDS M Phil 103: Time 95 Thursday, November CDS MPhil FQ 96
07, 2013 Series 07, 2013 Vijayamohan
16
11/7/2013
Thursday, November CDS MPhil FQ 97
07, 2013 Vijayamohan
17
11/14/2013
ARCH
Time Series Econometrics
9
Vijayamohanan Pillai N
Vijayamohan: CDS M Phil : Time Series 9 1 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 2 Thursday, November 14, 2013
Empirical applications
SURVEYS:
• to capture the volatility clustering
Bollerslev, Chou and Kroner(1992), Journal of
Econometrics, Special Issue on ARCH Models in which can be observed in
Finance macro-economic series such as
Bollerslev, Engle and Nelson,(1994), ARCH inflation (Engle, 1982),
Models, in Handbook of Econometrics, volume or financial time-series such as
IV, (eds. Engle and McFadden), Elsevier. exchange rates and
Engle, ARCH: SELECTED READINGS, Oxford stock market returns.
University Press, 1995
4
Vijayamohan: CDS M Phil : Time Series 9 3 Thursday, November 14, 2013
Modelling Volatility Modelling Volatility
High volatility periods alternate with periods of
Volatility refers to the degree of (typically short- relative tranquility.
term) unpredictable change over time of a certain But ‘volatility clusters’:
variable. Large changes in asset prices follow large changes
and small changes follow small ones.
It may be measured via the standard deviation of
⇒ current levels of volatility positively correlated
a sample. with its level during the preceding periods.
High volatility periods alternate with periods of
7.5
relative tranquility. 5.0
2.5
0.0
-2.5
-5.0
-7.5
0 20 40 60 80 100 120 140 160 180 200
Vijayamohan: CDS M Phil : Time Series 9 5 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 6 Thursday, November 14, 2013
1
11/14/2013
Modelling Volatility
A key distinction between the
conditional and
unconditional variance.
the unconditional variance =
the standard measure of the variance
Volatility is var(x) =E{x -E(x)}2
autocorrelated!
This phenomenon is Conditional variance:
modelled in
ARCH framework. Var(yt | yt-1)
Vijayamohan: CDS M Phil : Time Series 9 7 Thursday, November 14, 2013
Modelling Volatility: ARCH Model Modelling Volatility: ARCH Model
The regression function ‘Conditional error variance’ = time-varying.
Yt =β 0 + β1Xt + εt.
This revolutionary notion (Robert F. Engle
⇒ the conditional mean: Yt = E{Yt| Xt} + εt. 1982): made it possible to explain
systematic variation in variance and to
E{εt | Xt}= E(εt ) = 0, and E(εt εs) = 0, for t ≠ s;
estimate parameters of
In estimation: homoscedasticity assumption;
conditional variance jointly with those of
True for ‘unconditional error variance’; conditional mean.
‘Conditional error variance’ = time-varying.
Vijayamohan: CDS M Phil : Time Series 9 9 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 10 Thursday, November 14, 2013
Modelling Volatility:
Robert F. Engle III Conditional and Unconditional Variance
Distinction between conditional and
The Sveriges Riksbank Prize in Economic unconditional variance:
Sciences in Memory of Alfred Nobel 2003 Consider the AR(1) model:
"for methods of analyzing economic time Yt = Φ0 + Φ1Yt-1 + ε t, |Φ1|< 1;
series with time-varying volatility (ARCH)“
εt is a white noise process with Var(εt) = σε2.
USA Now E{Yt | Yt – 1} = Φ0 + Φ1Yt-1; and
New York University
New York, NY, USA Var(Yt | Yt – 1) = E{εt2 | Yt – 1} = σε2.
b. 1942
Short-run, conditional, variance.
Vijayamohan: CDS M Phil : Time Series 9 11 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 12 Thursday, November 14, 2013
2
11/14/2013
Modelling Volatility: Modelling Volatility: ARCH Model
Conditional and Unconditional Variance
Conditional error variance : time varying =
Unconditional variance of the AR(1) model:
heteroscedastic.
Yt = Φ0 + Φ1Yt-1 + εt, |Φ1|< 1; is
Modelling heteroscedasticity: as a function of
Var(Yt) = Var(Φ0 + Φ1Yt-1 + εt) another, exogenous variable:
= Φ12Var(Yt-1) + Var(εt) = σε2/ (1 – Φ12)
ht = Var{εt} = σt2 = cXt2.
Long-run, unconditional, variance.
Depends on knowledge of causes of
Since 1/ (1 – Φ12) > 1,
unconditional error variance > heteroscedasticity.
conditional error variance;
So conditional error variance preferable.
Vijayamohan: CDS M Phil : Time Series 9 13 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 14 Thursday, November 14, 2013
Modelling Volatility: ARCH Model Modelling Volatility: ARCH Model
ht = Var{εt} = σt2 = cXt2 In ht = Var{εt} = α0 + α1ε2t– 1 + α2ε2t– 2 +… +
Depends on knowledge of causes of αqε2t– q + υt
heteroscedasticity.
If α1, α2, …, αq all equal zero,
Engle (1982) suggests: heteroscedasticity the estimated variance = α0, constant.
depends on past values of errors: Thus using
Thus the conditional variance of Yt evolves according
the square of the estimated residuals:
to the above AR process:
ht = Var{εt} = σt 2 = α0 + α1ε2t– 1 + α2ε2t– 2 Hence it is called Autoregressive conditional
heteroscedastic (ARCH) model of order q.
+… + αqε2t– q + υ t,
υt = a white noise process.
Vijayamohan: CDS M Phil : Time Series 9 15 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 16 Thursday, November 14, 2013
Modelling Volatility: ARCH Model:
ARCH Non-negativity Constraints
• AUTOREGRESSIVE- In ht = Var(εt) = α0 + α1εt2– 1 + α2εt2– 2 +… + αqεt2– q + υt,
depends on its own past ht must be non-negative; usually positive; so
imposes conditions on the coefficients:
• CONDITIONAL- 1. α0 ≥ 0: If αi = 0, i = 1, .., q, then the conditional variance
variance depends upon past is α0 that must be ≥ 0;
information
2. αi ≥ 0, i = 1, .., q; εt2– i is always non-negative;
• HETEROSKEDASTICITY- so for the product αiεt2– i to be non-negative,
non-constant variance αi ≥ 0, ∀ i = 1, .., q;
Vijayamohan: CDS M Phil : Time Series 9 18 Thursday, November 14, 2013
3
11/14/2013
Modelling Volatility: ARCH Model: Modelling Volatility: ARCH Model:
Testing Testing
Given εt 2 = α0 + α1εt2– 1 + α2εt2– 2 +… + αqεt2– q
We can test for ARCH effects (of order q) :
1. Estimate the model of interest via OLS and save the 3. The test statistic = TR2;
squared residuals, εt2 so find the TR2 value
2. Perform an auxiliary regression in which the current (sample size times unadjusted R2)
squared residual is regressed on a constant and q lags from the auxiliary regression.
of itself: εt2 = α0 + α1εt2– 1 + α2εt2– 2 +… + αqεt2– q The test statistic = TR2:
Breusch and Pagan (1980);
Vijayamohan: CDS M Phil : Time Series 9 19 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 20 Thursday, November 14, 2013
Modelling Volatility: ARCH Model: Modelling Volatility: ARCH Model:
Testing
εt2 = α0 + α1εt2– 1 + α2εt2– 2 +… + αqεt2– q + υt,
Given εt2 = α0 + α1εt2–1 + α2εt2–2 +… + αqεt2–q
Many possible applications for ARCH models,
4. Compare the TR2 value with
since the residuals in the above equation can
the χ2 distribution with q degrees of freedom,
come from an autoregression, an ARMA model,
and if the p-value is small enough,
or a standard regression model.
reject the null hypothesis of homoscedasticity
in favor of the alternative of ARCH(q).
H0: α1 = α2 = …= αq = 0;
H1: α1 ≠ α2 ≠ … ≠ αq ≠ 0
Vijayamohan: CDS M Phil : Time Series 9 21 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 22 Thursday, November 14, 2013
εt2 = α0 + α1εt2– 1 + α2εt2– 2 +… + αqεt2– q + υt, Multiplicative conditionally heteroscedastic
model
Such linear specification not the most
convenient. Because the model for {yt} and the
conditional variance are best estimated ε t = υ t α 0 + α1ε t2−1
simultaneously using ML methods.
Where υt = white noise process such that
Hence specify υt as a multiplicative disturbance.
σν2 = 1,
Multiplicative conditionally heteroscedastic
E(υt, εt – 1) = 0, and
model proposed by Engle (1982) is:
αo, α1 constants such that
ε t = υ t α 0 + α1ε t2−1
αo > 0, and 0 < α1 < 1.
Vijayamohan: CDS M Phil : Time Series 9 23 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 24 Thursday, November 14, 2013
4
11/14/2013
White noise process υt ε t = υ t 1 + 0 . 6 ε t2− 1 White noise process υt ε t = υ t 1 + 0 . 6 ε t2− 1
εt2 = 1 + 0.01εt2– 1 + υt εt2 = 0.01εt2– 1 + υt yt = 0.2yt – 1 +εt yt = 0.9yt – 1 +εt
Vijayamohan: CDS M Phil : Time Series 9 25 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 26 Thursday, November 14, 2013
Modelling Volatility: GARCH Model Modelling Volatility: GARCH Model
Bollerslev (1986) extended Engle’s ARCH model: The GARCH(p, q) model
allows the conditional variance to be an ARMA
(where p is the order of the GARCH terms and q is the
process:
order of the ARCH terms) is given by
If an ARMA process is assumed for the
conditional error variance, the model is a
generalized autoregressive ht = σt2 = α0 + α1εt2–1 + … + αqεt2–q + β1 ht– 1 +
conditional heteroskedasticity
(GARCH) model. …+β p ht– p + υt ,
Vijayamohan: CDS M Phil : Time Series 9 27 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 28 Thursday, November 14, 2013
Modelling Volatility: GARCH Model Modelling Volatility:
(G)ARCH Variations
ht = α0 + α1εt2– 1 + … + αqεt2– q + β 1ht– 1 +
Absolute (G)ARCH: (AGARCH): If absolute
…+βp ht– p + υt,
values |εt| are used.
If all β i = 0; then no MA term in the
Exponential GARCH: (EGARCH) (Nelson 1991):
conditional variance and
exponential (log) specification;
GARCH(p, q) = ARCH(q) model.
even if the parameters are negative,
∞);
GARCH: more parsimonious for an ARCH(∞ ht = σt2 will be +ve;
avoids overfitting; no need to impose non-negativity constraints.
less likely to breach non-negativity
constraints.
Vijayamohan: CDS M Phil : Time Series 9 29 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 30 Thursday, November 14, 2013
5
11/14/2013
Modelling Volatility: Modelling Volatility:
(G)ARCH Variations: (G)ARCH Variations
EGARCH
Integrated GARCH (IGARCH): (Engle and
Exponential GARCH: (EGARCH) (Nelson 1991): Bollerslev 1986; Nelson 1990):
exponential (log) specification: a restricted version of the GARCH model,
ln(ht ) = α0 + α1 {εt–1 /(ht–1 )0.5} where the sum of the persistent parameters
+ λ1| εt–1 /(ht–1 )0.5| + β1 ln(ht– 1) sum up to one.
(G)ARCH in Mean: (G)ARCH-M (Engle, Lilien
and Robins 1987):
allows the mean of a sequence to depend on
its own conditional variance.
Vijayamohan: CDS M Phil : Time Series 9 31 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 32 Thursday, November 14, 2013
Yt
Squared Yt
ACF and PACF of Yt ACF and PACF of
Squared Yt
Vijayamohan: CDS M Phil : Time Series 9 33 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 34 Thursday, November 14, 2013
Modelling Volatility: Modelling Volatility:
in PcGive in PcGive
Vijayamohan: CDS M Phil : Time Series 9 35 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 36 Thursday, November 14, 2013
6
11/14/2013
Modelling Yt by restricted GARCH(0,1) (Data1) ht = α0 + α1εt2– 1 Residuals
The estimation sample is: 1 to 200
Coefficient Std.Error robust-SE t-value t-prob
Constant X -0.0428385 0.07188 0.07561 -0.567 0.572
alpha_0 H 0.831143 0.1415 0.1271 6.54 0.000
alpha_1 H 0.631354 0.1486 0.1194 5.29 0.000
log-likelihood -331.112707 HMSE 1.76418
mean(h_t) 2.07511 var(h_t) 4.27002
no. of observations 200 no. of parameters 3
AIC.T 668.225414 AIC 3.34112707
mean(yt) -0.0164891 var(yt) 1.96391
alpha(1)+beta(1) 0.631354 alpha_i+beta_i>=0,
← Restriction on IGARCH)
alpha(1)+beta(1) <1 (← Squared Residuals: Conditional Variance
Vijayamohan: CDS M Phil : Time Series 9 37 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 38 Thursday, November 14, 2013
ACF and PACF of Residuals ACF and PACF of squared residuals:
Conditional variance
Vijayamohan: CDS M Phil : Time Series 9 39 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 40 Thursday, November 14, 2013
Modelling Yt by restricted GARCH(1,1) (Data1) Modelling Yt by EGARCH(0,1) (Data1)
The estimation sample is: 2 to 200 ht = α0 + α1εt2– 1 + β1ht– 1 The estimation sample is: 1 to 200
Coefficient Std.Error robust-SE t-value t-prob
Coefficient Std.Error robust-SE t-value t-prob
Constant X -0.0388542 0.07239 0.07660 -0.507 0.613 Constant X -8.638e-007 0.01101 0.001653 -0.000522 1.000
alpha_0 H 0.892488 0.1983 0.1700 5.25 0.000 alpha_0 H 0.421241 0.1390 0.1278 3.30 0.001
alpha_1 H 0.628874 0.1480 0.1176 5.35 0.000 eps[-1] H 0.186865 0.08641 0.06599 2.83 0.005
beta_1 H -0.0299775 0.07239 0.04900 -0.612 0.541 |eps[-1]| H 0.980867 0.1531 0.1277 7.68 0.000
log-likelihood -330.154416 HMSE 1.74143 log-likelihood -328.518742 HMSE 1.6768
mean(h_t) 2.07558 var(h_t) 4.13637 mean(h_t) 1.90533 var(h_t) 2.04582
no. of observations 199 no. of parameters 4 no. of observations 200 no. of parameters 4
AIC.T 668.308832 AIC 3.35833584 AIC.T 665.037485 AIC 3.32518742
mean(y) -0.016572 var(y) 1.97378 mean(yt) -0.0164891 var(yt) 1.96391
alpha(1)+beta(1) 0.598897 alpha_i+beta_i>=0,
ln(ht ) = α0 + α1 {εt–1 /(ht–1)0.5} + λ1| εt–1 /(ht–1 )0.5|
alpha(1)+beta(1)<1
Vijayamohan: CDS M Phil : Time Series 9 41 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 42 Thursday, November 14, 2013
7
11/14/2013
Residuals ACF and PACF of Residuals
Squared Residuals: Conditional Variance
Vijayamohan: CDS M Phil : Time Series 9 43 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 44 Thursday, November 14, 2013
Modelling Yt by EGARCH(1,1) (Data1)
ACF and PACF of squared residuals: The estimation sample is: 2 to 200
Conditional variance
Coefficient Std.Error robust-SE t-value t-prob
Constant X 0.00753892 0.07226 0.07108 0.106 0.916
alpha_0 H 0.339108 0.1410 0.1277 2.66 0.009
eps[-1] H 0.186005 0.1039 0.07598 2.45 0.015
|eps[-1]| H 1.01518 0.1649 0.1282 7.92 0.000
beta_1 H 0.185387 0.1582 0.1436 1.29 0.198
log-likelihood -327.180358 HMSE 1.64902
mean(h_t) 1.94404 var(h_t) 2.28722
no. of observations 199 no. of parameters 5
AIC.T 664.360715 AIC 3.33849606
mean(y) -0.016572 var(y) 1.97378
ln(ht ) = α0 + α1 {εt–1 /(ht–1)0.5}+ λ 1| εt–1 /(ht–1 )0.5| + β 1 ln(ht– 1)
Vijayamohan: CDS M Phil : Time Series 9 45 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 46 Thursday, November 14, 2013
Modelling Volatility: Modelling Volatility:
in Stata in Stata
Vijayamohan: CDS M Phil : Time Series 9 47 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 48 Thursday, November 14, 2013
8
11/14/2013
Modelling Volatility: . arch y, arch(1/1)
Modelling Volatility:
in Stata in Stata
ARCH family regression Modelling Yt by restricted GARCH(0,1)
The estimation sample is: 1 to 200
Sample: 1994w1 - 1997w44 Number of obs = 200
Distribution: Gaussian Wald chi2(.) = . Coefficient Std.Error
Log likelihood = -331.1127 Prob > chi2 = . Constant X -0.0428385 0.07188
alpha_0 H 0.831143 0.1415
------------------------------------------------------------------------------ alpha_1 H 0.631354 0.1486
| OPG
y | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
y |
_cons | -.0428354 .0707218 -0.61 0.545 -.1814475 .0957768
-------------+----------------------------------------------------------------
ARCH |
arch |
L1. | .6313589 .1866612 3.38 0.001 .2655096 .9972082
|
_cons | .831128 .1609185 5.16 0.000 .5157335 1.146523
------------------------------------------------------------------------------
Vijayamohan: CDS M Phil : Time Series 9 49 Thursday, November 14, 2013 Vijayamohan: CDS M Phil : Time Series 9 50 Thursday, November 14, 2013
Thursday, November 14, Vijayamohan: CDS M Phil : Time 51
2013 Series 9
9
View publication stats
You might also like
- An Introduction To Bayesian VAR (BVAR) Models R-EconometricsNo ratings yetAn Introduction To Bayesian VAR (BVAR) Models R-Econometrics16 pages
- Chapter 5 Exponential Smoothing Methods L 2015No ratings yetChapter 5 Exponential Smoothing Methods L 201519 pages
- The Analysis of Time Series An Introduction by Chris Chatfield 5th Edition 11 18 PDFNo ratings yetThe Analysis of Time Series An Introduction by Chris Chatfield 5th Edition 11 18 PDF8 pages
- Basic Concepts of Time Series Modeling-NotesNo ratings yetBasic Concepts of Time Series Modeling-Notes4 pages
- Hannan E.J., Krishnaiah P.R., Rao M.M.-Handbook of Statistics, Vol. 5. Time Series in The Time Domain (1985) PDFNo ratings yetHannan E.J., Krishnaiah P.R., Rao M.M.-Handbook of Statistics, Vol. 5. Time Series in The Time Domain (1985) PDF482 pages
- Arma Models: This Project Is About The Time Analysis Based Model ARMA. Which Is A ForecastingNo ratings yetArma Models: This Project Is About The Time Analysis Based Model ARMA. Which Is A Forecasting23 pages
- Manonmaniam Sundaranar University: B.Sc. Statistics - Iii YearNo ratings yetManonmaniam Sundaranar University: B.Sc. Statistics - Iii Year88 pages
- Module 02.1 Time Series Analysis and Forecasting AccuracyNo ratings yetModule 02.1 Time Series Analysis and Forecasting Accuracy11 pages
- Journal Time Series Analysis - 2023 - Armillotta - Count Network AutoregressionNo ratings yetJournal Time Series Analysis - 2023 - Armillotta - Count Network Autoregression29 pages
- John W. Tukey's Work On Time Series and Spectrum AnalysisNo ratings yetJohn W. Tukey's Work On Time Series and Spectrum Analysis25 pages
- TOD 212-PPT 2 For Students - Monsoon 2023No ratings yetTOD 212-PPT 2 For Students - Monsoon 202326 pages
- Stationarity Issues in Time Series Models: David A. Dickey North Carolina State UniversityNo ratings yetStationarity Issues in Time Series Models: David A. Dickey North Carolina State University17 pages
- Sta 445 1 Stationarity and Non-StationarityNo ratings yetSta 445 1 Stationarity and Non-Stationarity15 pages
- Time Series Analysis Using SAS The Augmented Dickey-Fuller (ADF) TestNo ratings yetTime Series Analysis Using SAS The Augmented Dickey-Fuller (ADF) Test6 pages
- The Annals of Statistics 2002, Vol. 30, No. 6, 1595-1618No ratings yetThe Annals of Statistics 2002, Vol. 30, No. 6, 1595-161824 pages
- Bangladesh Climate and Disaster Risk Atlas: Hazards—Volume IFrom EverandBangladesh Climate and Disaster Risk Atlas: Hazards—Volume INo ratings yet
- Multihazard Risk Atlas of Maldives: Climate and Geophysical Hazards—Volume IIFrom EverandMultihazard Risk Atlas of Maldives: Climate and Geophysical Hazards—Volume IINo ratings yet
- Asset Rotation: The Demise of Modern Portfolio Theory and the Birth of an Investment RenaissanceFrom EverandAsset Rotation: The Demise of Modern Portfolio Theory and the Birth of an Investment RenaissanceNo ratings yet
- YaffeeRA2010ForecastEvaluationwithStataNo ratings yetYaffeeRA2010ForecastEvaluationwithStata63 pages
- Panel Causality Example Xtabond2 - 11 Asamalar YESNo ratings yetPanel Causality Example Xtabond2 - 11 Asamalar YES33 pages
- Binomial, Poisson, and Normal DistributionsNo ratings yetBinomial, Poisson, and Normal Distributions3 pages
- An Introduction To Probability and Statistics - 2015 - Rohatgi - Subject IndexNo ratings yetAn Introduction To Probability and Statistics - 2015 - Rohatgi - Subject Index11 pages
- Operational Risk Modeling in Insurance and BankingNo ratings yetOperational Risk Modeling in Insurance and Banking13 pages
- Introduction To Computational Finance and Financial Econometrics With RNo ratings yetIntroduction To Computational Finance and Financial Econometrics With R527 pages
- MS5030: Data Analysis For Management: Rahul R MaratheNo ratings yetMS5030: Data Analysis For Management: Rahul R Marathe39 pages
- SEHH1008 Chapter 06 Binomial DistributionNo ratings yetSEHH1008 Chapter 06 Binomial Distribution28 pages
- Sta301 Final Term Solved Mcqs Mega File-1-Converted-1No ratings yetSta301 Final Term Solved Mcqs Mega File-1-Converted-140 pages
- Rounding Error Propagation - Bias and UncertaintyNo ratings yetRounding Error Propagation - Bias and Uncertainty34 pages
- Discrete Probability Distributions Problem SetNo ratings yetDiscrete Probability Distributions Problem Set7 pages
- Syllabus Mathematics For Communication 2024-25No ratings yetSyllabus Mathematics For Communication 2024-254 pages
- Test Bank For Stats Data and Models 5th by de Veaux100% (47)Test Bank For Stats Data and Models 5th by de Veaux9 pages
- Computer Science CPSC 322: Bayesian Networks: ConstructionNo ratings yetComputer Science CPSC 322: Bayesian Networks: Construction70 pages
- Lecture 07 - BKF4310 - Financial Modelling II - Brownian MotionNo ratings yetLecture 07 - BKF4310 - Financial Modelling II - Brownian Motion37 pages
