Understanding Stationarity in Time Series Analysis

Stationarity
Denition, meaning and consequences
Matthieu Stigler Matthieu.Stigler@gmail.com
November 14, 2008
Version 1.1
This document is released under the Creative Commons Attribution-Noncommercial 2.5 India license.
Matthieu Stigler Matthieu.Stigler@gmail.com () Stationarity November 14, 2008 1 / 56
Outline
1 2 3
Preliminary Lectures Denitions Time series Description of a time series Stationarity Stationary processes Nonstationary processes The random-walk The random-walk with drift Trend stationarity Economic meaning and examples
Stationarity November 14, 2008 2 / 56
4 5
Matthieu Stigler Matthieu.Stigler@gmail.com ()
Economies of scale and increasing returns
So-called Nobel Prize has been attributed to Krugman: Economies of scale combined with reduced transport costs also help to explain why an increasingly larger share of the world population lives in cities and why similar economic activities are concentrated in the same locations. [This], in turn, stimulates further migration to cities.
Stationarity
November 14, 2008
3 / 56
The increasing returns are: Self production: Learning curve Common production: adoption/network externalities
Stationarity
November 14, 2008
4 / 56
How is a keyboard organised?

QWERTZ vs DVORAK costs
5 QWERTZ DVORAK
Average cost
0 0
10
20
30
Stationarity
40
50
November 14, 2008 5 / 56
Outline
1 2 3
4 5
1 2 3 4 5 6 7 8 9 10 11 12
Stationarity ARMA models for stationary variables Some extensions of the ARMA model Non-stationarity Seasonality Non-linearities Multivariate models Structural VAR models Cointegration the Engle and Granger approach Cointegration 2: The Johansen Methodology Multivariate Nonlinearities in VAR models Multivariate Nonlinearities in VECM models
Stationarity
November 14, 2008
7 / 56
Outline
1 2 3
4 5
Outline
1 2 3
4 5
What is a time serie?

In econometrics, we deal with three types of data: Cross individual data: Xi Times series data: Xt Panel data: Xit t : time i : individu t : time
Denition
A time series is a variable X indexed by the time t: Xt t = 1, 2, . . . , T .
Examples
The time t can be annual, monthly, daily... Let be Xt the annual GDP. Let Yt be the monthly temperature Zt be the daily stocks
Stationarity
November 14, 2008
10 / 56
Some distinctions
The time series can: Take discrete or continuous values. Be measured at discrete (monthly, daily) or continous time (signal processing, nance). Be measured at regular or irregular intervals. In this lecture, usually we will refer to time series that take continous values and are measured at discrete and regular intervalls.
Stationarity
November 14, 2008
11 / 56
Outline
1 2 3
4 5
Description of a time serie
To describe a time serie, we will concentrate on: Its Data Generating Process (DGP) The joint distribution of its elements Its moments :
Its expected value, E[Xt ] t . Its variance, Var[Xt ] 2 0 . Its autocovariance or autocorrelation of order k, Cov[Xt , Xtk ] k .
Stationarity
November 14, 2008
13 / 56
Variance
The second moments play an important role in time series and have a slight dierent denition:
Denition
The variance of Xt , denoted by 0 (t), is given by: Var(Xt ) E (Xt E(Xt ))2
Denition
The autocovariance of Xt of order k, denoted by k (t), is given by: Cov(Xt , Xtk ) E [(Xt E(Xt ))(Xtk E(Xtk ))]
Denition
The autocorrelation of Xt of order k, denoted by k (t), is given by: Cov(Xt ,Xtk ) Corr(Xt , Xtk )
Var(Xt ) Var(Xtk )
Stationarity
November 14, 2008
14 / 56
Outline
1 2 3
4 5
The stationarity is an essential property to dene a time series process:
Denition
A process is said to be covariance-stationary, or weakly stationary, if its rst and second moments are time invariant. E(Yt ) = E[Yt1 ] = Var(Yt ) = 0 < Cov(Yt , Ytk ) = k t t t, k
Stationarity
November 14, 2008
16 / 56
Weak stationarity
The third condition states that the autocovariances only depend on the decay in the time but not in the time itself. Hence, the structure of the serie does not change with the time. If a process is stationary, k = k
Stationarity
November 14, 2008
17 / 56
Mean reverting propriety
A stationary process has the propriety to be mean reverting : it will uctuate around its mean. This mean will act as an attractor. It will cross the mean line an innite number of ways.
Stationarity
November 14, 2008
18 / 56
Strong stationarity
Denition
f (Z1 , Z2 , ..., Zt ) = f (Z1+k , Z2+k , ..., Zt+k ) . Implies weak stationarity Denition not very useful as informations about the density function are dicult to obtain
Stationarity
November 14, 2008
19 / 56
ergodicity
Not absolutely useful...
Denition
limn E [|Y (yi , . . . , yi+k )Z (yi+n , . . . , yi+n+l )|] = E [Y (yi , . . . , yi+k )] E [Z (yi+n , . . . , yi+n+l )]
Stationarity
November 14, 2008
20 / 56
Outline
1 2 3
4 5
Example 1
We dene now the simplest stationary process:
Denition
An independent white noise process is a sequence of independant elements with same expected value and variance. We will denote it t iid(0, 2 )
Stationarity
November 14, 2008
22 / 56
Example 2
Consider the rst order auto-regressive process AR(1) with a constant: Yt = c + Yt1 + t t iid(0, 2 )
Theorem
An AR(1) process is asymptotically stationary if || < 1 We will need fot its proof to remember the properties of a geometric progression:
Lemma (Innite geometric progression)

|| < 1
i i=0
= 1 + + 2 + 3 + . . . =
1 1
Stationarity
November 14, 2008
23 / 56
Proof.
The AR(1) can be written as:
t1 t1
Yt = c
i=0
+ Y0 +
i=0
i ti
If || < 1, it can be simplied into: Yt = We have then: E(Xt ) =

c 1 2 12 j 2 12
c + 1
t1
i ti
i=0
Var(Xt ) =
Cov(Xt , Xtj ) =
Stationarity
November 14, 2008
24 / 56
Outline
1 2 3
4 5
Corollary
An AR(1) process is nonstationary if || 1. Furthemore, if: || = 1 it is dierence stationary || > 1 it is explosive.
Stationarity
November 14, 2008
26 / 56
Outline
1 2 3
4 5
Proof.
The AR(1) process without constant Yt = Yt1 + t with || = 1 can be rewritten as:
t1
t iid(0, 2 )
Yt = Y0 +
i=0
ti
We have then: E(Xt ) = Y0 Var(Xt ) = t 2 Cov(Xt , Xtj ) = (t j) 2
Stationarity
November 14, 2008
28 / 56
The AR(1) process with || = 1 is called a random walk. It is said to be dierence stationary.
Denition
The dierence operator takes the dierence between a value of a time serie and its lagged value. Xt Xt Xt1
Denition
A process is said to be dierence stationary if it becomes stationary after being dierenced once. Note: a dierence stationary process is also called integrated of order 1 and denoted by Xt I (1)
Stationarity
November 14, 2008
29 / 56
Theorem
A random walk is dierence stationary.
Proof.
Yt = Yt Yt1 = Yt1 + t Yt1 = t
Stationarity
November 14, 2008
30 / 56
Stationary AR(1) Yt = Yt1 + t || < 1, t iid(0, 2 ) Yt = Y0 +

t1 i i=0 t1
Random Walk Yt = Yt1 + t t iid(0, 2 ) Yt = Y0 +

t1 i=0 t1
E(Yt ) = 0( if Y0 = 0) Var(Yt ) =
2 12 j 2 12
E(Yt ) = 0( if Y0 = 0) Var(Yt ) = 2 t Cov(Yt , Ytj ) = (t j) 2 Corr(Yt , Ytj ) =

Yt ti ts t
Cov(Yt , Ytj ) =
Corr(Yt , Ytj ) = j
Yt ti
=1
Is mean reverting
Tends to move away from the mean

Simulation of AR(1) and RW

Random Walk and AR
RW AR(1) : = 0.4 0 RW 0 20 40 Matthieu Stigler Matthieu.Stigler@gmail.com () 15 10 5
60 Stationarity
80
100
November 14, 2008
32 / 56

Random Walk and AR
RW AR(1) : = 0.4 2 RW 0 20 40 Matthieu Stigler Matthieu.Stigler@gmail.com () 6 4 2 0
60 Stationarity
80
100
November 14, 2008
33 / 56

Random Walk and AR
2 RW AR(1) : = 0.9
RW
0 20 40 Matthieu Stigler Matthieu.Stigler@gmail.com ()
10
60 Stationarity
80
100
November 14, 2008
34 / 56
Stationary AR(1) Yt = Yt1 + t || < 1, t iid(0, 2 ) Yt = Y0 +

t1 i i=0 t1
Random Walk Yt = Yt1 + t t iid(0, 2 ) Yt = Y0 +

t1 i=0 t1
E(Yt ) = 0( if Y0 = 0) Var(Yt ) =
2 12 j 2 12
E(Yt ) = 0( if Y0 = 0) Var(Yt ) = 2 t Cov(Yt , Ytj ) = (t j) 2 Corr(Yt , Ytj ) =

Yt ti ts t
Cov(Yt , Ytj ) =
Corr(Yt , Ytj ) = j
Yt ti
=1
Is mean reverting
Tends to move away from the mean

Yt = 0 + 0t + 0.3Yt 1
20
Yt = 0 + 0t + 0.95Yt 1
5 10
5 0
20
40 t
60
80
100
10 0
10
20
20
40 t
60
80
100
Yt = 0 + 0t + 0.5Yt 1
50
Yt = 0 + 0t + 1Yt 1
15
Y 0 20 40 t 60 80 100
0 10 0
30
20
40 t
60
80
100
November 14, 2008 36 / 56
Stationarity
Outline
1 2 3
4 5
The random walk with drift

Xt = a + xt1 + t Can be rewritten as: t iid(0, 2 )
t1
Yt = at + Y0 +
i=0
ti
Expectation is also time-varying: E(Xt ) = Y0 + at Var(Xt ) = . . . = f (t) Cov(Xt , Xtj ) = . . . = f (t) But it is still dierence stationary:
Proof.
Yt = Yt Yt1 = a + Yt1 + t Yt1 = a + t
Stationarity
November 14, 2008
38 / 56
Outline
1 2 3
4 5
Trend stationarity
Yt = a + bt + Yt1 + t1
t1 t1 t1
Yt = a
i=0
i + a
i=0
i + t Y0 +
i=0
i ti
If || < 1, it can be simplied into: Yt = a +b 1

t1 t1
i (t i) +
i=0 i=0
i ti
Stationarity
November 14, 2008
40 / 56
Proof.
Yt = a + bt + Yt1 + t1 = a + bt + (a + b(t 1) + Yt2 + t2 ) + ti = a(1 + ) + bt + b(t 1) + 2 Yt2 + t2 + t1 = a(1 + ) + bt + b(t 1) + 2 (a + b(t 2) + Yt3 + t3 ) + t2 + t1 = a(1 + + 2 ) + b(t + (t 1) + 2 (t 2)) + 3 Yt3 + t3 + t2 + t1 = ...
t1 t1 t1
= Y0 + a a +b 1
+b
i=0 t1 i=0 i=0
(t i) +
i=0 t1
i ti
i (t i) +
i=0
i ti
November 14, 2008 41 / 56
Stationarity
Thus the model Yt = a + bt + Yt1 + t1 has: E(Xt ) = f (t) Var(Xt ) = f (t) Cov(Xt , Xtj ) = f (t) It is non-stationary as its expectation is time varying. However, its variance does not vary with time! This process is called trend-stationary: if one detrends it, the series is stationary:
Proposition
Yt a + bt + Yt1 + t1 is not stationary Yt bt = a + Yt1 + t1 is stationary
Stationarity
November 14, 2008
42 / 56
250
TS: y=0.5 t + e DS: y(t)=a+y(t1)+e
0 0
50
100
150
200
100
200
300
400
500
Stationarity
November 14, 2008
43 / 56
250
TS: y=0.5 t + e DS: y(t)=a+y(t1)+e
0 0
50
100
150
200
100
200
300
400
500
Stationarity
November 14, 2008
44 / 56
Comparisons between Trend and dierence stationarity: TS DGP: DGP E(Yt ) Var(Yt ) Cov(Yt , Ytj )
Yt ti
DS Yt = c + Yt1 + t Yt = Y0 + ct +
t1 i=0 t1
Yt = + t + t
+ t 2 0 0 = f (s)
Y0 + ct t 2 (t j) 2 =a = f (s)
E[yt+s yt+s|t ]2
Stationarity
November 14, 2008
45 / 56
Stationarity
November 14, 2008
46 / 56
So both TS and DS exhibit a trend tendency but with stable or increasing variance. This trend is said: Deterministic: TS process Stochastic: DS process
Stationarity
November 14, 2008
47 / 56
Yt = 0 + 0.5t + 0.3Yt 1
AR(1) AR(1)+e Y 100
Yt = 0.5 + 0t + 1Yt 1
AR(1) RW
60
20
40
20
40 t
60
80
100
0 0
20
60
20
40 t
60
80
100
Yt = 0 + 0.5t + 0.5Yt 1
100 AR(1) AR(1)+e Y 60
Yt = 0.5 + 0t + 1Yt 1
AR(1) RW AR(1)+e RW+e
60
20
20
40 t
60
80
100
0 0
20
40
20
40 t
60
80
100
November 14, 2008 48 / 56
Stationarity
Outline
1 2 3
4 5
Nelson and Plosser (1982) study

Nelson and Plosser (1982) investigate 14 time series: Real GNP Nominal GNP Real Per Capita GNP Industrial Production Index Total Employment Total Unemployment Rate GNP Deator Consumer Price Index Nominal Wages Real Wages Money Stock (M2) Velocity of money Bond Yield (30-year corporate bonds) Stock Prices
Matthieu Stigler Matthieu.Stigler@gmail.com () Stationarity November 14, 2008 50 / 56
Nelson and Plosser (1982) study

Real GNP
7.0 8.5 1860 1900 1940 1980 7.0 1860 7.5 5.0 8.0
Real Per Capita GNP
6.0
1900
1940
1980
Nominal GNP
11 12 13 14 15 5 0
Matthieu Stigler Matthieu.Stigler@gmail.com () 1860 1900 1940 1980
Industrial Production Index
1
Stationarity 1900 1860
1940
1980 November 14, 2008
51 / 56
Total Employment
10.0 10.5 11.0 11.5 6.0
GNP Deflator
1860
1900
1940
1980
3.0 1860
4.0
5.0
1900
1940
1980
Total Unemployment Rate

5.5 1860 1900 1940 1980 3.5 4.5
Consumer Price Index
0.5
1.5
2.5
1860
Stationarity
1900
1940
1980
November 14, 2008 52 / 56
Nelson and Plosser (1982) study 3

Nominal Wages
10
Money Stock (M2)
Bond Yield (30year corporate bonds)
1860 1900 1940 1980
1860 1900 1940 1980
2 1860 1900 1940 1980
Real Wages
4.2
Velocity of money
10
12
Stock Prices
1.5
3.8
1.0
3.4
0.5
3.0
Stationarity
November 14, 2008
53 / 56
Results of the test of Trend stationary vs Dierence stationary:
Results
13 series can be viewed as DS, one (unemployment) as TS. The distinction between the two classes of processes is fundamental and aceptance of the purely stochastic view of non-stationarity has broad implications for our understanding of the nature of economic phenomena
Stationarity
November 14, 2008
54 / 56
We conclude that macroeconomic models that focus on monetary disturbances as a source of purely transitory (stationary) uctuations may never be successful in explaining a very large fraction of output uctuations and that stochastic variation due to real factors is an essential element of any model of economic uctuations.
Stationarity
November 14, 2008
55 / 56
Fill var and cov page 33 Please check 35 and 36 Give title plots 38 39 Page 4: The axes should really begin with 0;0 so that only 2 lines of the rect should be visible
Stationarity
November 14, 2008
56 / 56

Understanding Stationarity in Time Series Analysis

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Understanding Stationarity in Time Series Analysis

Uploaded by

Copyright:

Available Formats

Stationarity

Denition, meaning and consequences

Matthieu Stigler Matthieu.Stigler@gmail.com

November 14, 2008

Matthieu Stigler Matthieu.Stigler@gmail.com ()

Economies of scale and increasing returns

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

How is a keyboard organised?

Matthieu Stigler Matthieu.Stigler@gmail.com ()

Matthieu Stigler Matthieu.Stigler@gmail.com ()

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Matthieu Stigler Matthieu.Stigler@gmail.com ()

Matthieu Stigler Matthieu.Stigler@gmail.com ()

What is a time serie?

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Matthieu Stigler Matthieu.Stigler@gmail.com ()

Description of a time serie

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Matthieu Stigler Matthieu.Stigler@gmail.com ()

The stationarity is an essential property to dene a time series process:

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Mean reverting propriety

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Not absolutely useful...

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Matthieu Stigler Matthieu.Stigler@gmail.com ()

We dene now the simplest stationary process:

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Lemma (Innite geometric progression)

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

If || < 1, it can be simplied into: Yt = We have then: E(Xt ) =

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Matthieu Stigler Matthieu.Stigler@gmail.com ()

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Matthieu Stigler Matthieu.Stigler@gmail.com ()

We have then: E(Xt ) = Y0 Var(Xt ) = t 2 Cov(Xt , Xtj ) = (t j) 2

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Matthieu Stigler Matthieu.Stigler@gmail.com ()

November 14, 2008

Matthieu Stigler Matthieu.Stigler@gmail.com ()