Time Series

Stationary and Non-Stationary Time Series

In regression analysis involving time series data, a critical assumption is that the time
series under consideration is stationary. Broadly speaking, a time series is stationary if
its mean and variance are constant over time.
A time series is an example of what is called a stochastic process, which is a sequence
of random variables ordered in time.2
The importance of stationary time series
First, if a time series is nonstationary, we can study its behavior only for the
period under consideration, such as the one in our dollar/euro exchange rate. Each
time series will therefore be a particular episode. As a result, it is not possible to generalize
it to other time periods. For forecasting purposes, therefore, nonstationary time
series will be of little practical value.
Second, if we have two or more nonstationary time series, regression analysis involving
such time series may lead to the phenomenon of spurious or nonsense regression.
That is, if you regress a nonstationary time series on one or more nonstationary
time series, you may obtain a high R2 value and some or all of the regression coefficients
may be statistically significant on the basis of the usual t and F tests. Unfortunately,
in cases of nonstationary time series these tests are not reliable, for they assume
that the underlying time series are stationary.
Tests of stationarity
For the reasons just stated, it is important to find out if a time series is stationary.
There are basically three ways to examine the stationarity of a time series:
(1) graphical
analysis, (2) correlogram, and (3) unit root analysis.
 lnex
0.6

0.5

0.4

0.3

0.2

0.1

0
0 500 1000 1500 2000 2500

-0.1

-0.2

-0.3
 log(ex)t-log(ex)t-1
Stationary and Non-stationary
Stochastic Process 0.03

Stationary Stochastic Process 0.02

A time series is stationary if its mean and variance 0.01

are constant over time (do not vary systematically
0
over time)

133
199
265
331
397
463
529
595
661
727
793
859
925
991
1

1057
1123
1189
1255
1321
1387
1453
1519
1585
1651
1717
1783
1849
1915
1981
2047
2113
2179
2245
2311
67
-0.01
In short, if a time series is stationary, its mean,
variance, and autocovariance (at various lags) -0.02
remain the same no matter at what point we
measure them; that is, they are time invariant. -0.03

Non-Stationary Stochastic Process lnex

0.6
If a time series is not stationary in the sense just
0.5
defined, it is called a nonstationary time series. In
other words, a nonstationary time series will have a 0.4
time varying mean or a time-varying variance or 0.3
both
0.2

0.1

0
0 500 1000 1500 2000 2500
-0.1

-0.2

-0.3
 lnex
Stationary and Non-stationary 0.6

Stochastic Process 0.5

0.4

Most economic time series in level form are 0.3

nonstationary. 0.2

Such series often exhibit an upward or downward 0.1

trends over a sustained period of time.
0
0 500 1000 1500 2000 2500
-0.1

-0.2

-0.3

250

200

Profits, Dividends
150

PROFITS
100
DIVIDENDS

50

0
1965 1970 1975 1980 1985 1990 1995
Year
 Why it is important to have a stationary stochastic
process?
If we want to understand the relationship between two or more variables using
regression analysis, we need to assume some sort of stability over time. If we allow
the relationship between two variables (say, yt and xt) to change arbitrarily in each
time period, then we cannot hope to learn much about how a change in one variable
affects the other variable if we only have access to a single time series realization. In
stating a multiple regression model for time series data, we are assuming a certain
form of stationarity in that the βj do not change over time.
Regressing a nonstationary time series on one or more nonstationary time series
may often lead to the phenomenon of spurious or meaningless regression.
if a time series is nonstationary, we can study its behavior only for the time period
under consideration. Each set of time series data will therefore be for a particular
episode. As a consequence, it is not possible to generalize it to other time periods.
Therefore, for the purpose of forecasting, such (nonstationary) time series may be of
little practical value.
 Yt = β1 + β2t + β3Yt−1 + ut
β1 = 0, β2 = 0, β3 = 1 β1 ≠ 0, β2 = 0, β3 = 1, β1 ≠ 0, β2 ≠ 0, β3 = 0 β1 ≠ 0, β2 ≠ 0, β3 = 1
Yt = β1 +Yt−1 + ut Yt = β1 + β2t + ut Yt = β1 + β2t + Yt−1 + ut
Yt = Yt−1 + ut

Deterministic trend Random walk with drift

Random walk without drift (trend stationary and deterministic trend
And trend (Pure random walk) Random walk with drift process)
Yt = b2t+Yt−1 + ut

Yt = β2t + ut
DIAGNOSTIC TOOLS: TESTING FOR
STATIONARITY

1.Correlogram

2.Unit Root Test (ADF test)

Unit root test
Yt = β1 + β2t + β3Yt−1 + ut
Test for β3=1
If accept, series is non-stationary

Yt - Yt−1 = β1 + β2t + β3Yt−1- Yt−1 + ut

ΔYt = β1 + β2t + (β3-1)Yt−1+ ut
ΔYt = β1 + β2t + cYt−1+ ut ; where (β3-1)=c
Test for c=0
 UNIT ROOT TEST OF STATIONARITY
➢The unit root test for the exchange rate can be expressed as follows:
LEX t = B1 + B2t + B3 LEX t −1 + ut
➢We regress the first differences of the log of exchange rate on the trend variable and the
one-period lagged value of the exchange rate.
➢The null hypothesis is that B3, the coefficient of LEXt-1, is zero.
➢This is called the unit root hypothesis.
➢A non-rejection (acceptance) of the null hypothesis would suggest that the time series
under consideration is nonstationary.
➢We cannot use a t test because the t test is valid only if the
underlying time series is stationary.
➢Use the τ (tau) test, also known as the Dickey-Fuller (DF)
test, whose critical values are calculated by simulations and
modern statistical packages
Inference

ADF Test

criteria schwert
drift yes
trend yes
lag 26
alpha 0.05
For series to be non-stationary:
tau-stat -2.46515 Absolute value of tau-stat has to be less than the tau-critical value
tau-crit -3.41221
stationary no
aic -6.23728
bic -6.22994
lags 0
coeff -0.00602
p-value > .1

Question fromAcademic year 2022-23 to be given

2. AUTOCORRELATION FUNCTION (ACF)
AND PARTIAL AUTOCORRELATION
FUNCTION(PACF)
Correlogram alpha 0.05
Correlogram alpha 0.05
lags pacf lower upper
lags acf lower upper 0 1
0 1
1 0.998486 -0.04039 0.040388
1 0.998486 -0.04039 0.040388
2 0.003673 -0.04039 0.040388
2 0.996986 -0.06988 0.069884
3 -0.01726 -0.04039 0.040388
3 0.995436 -0.09015 0.090147
4 0.011685 -0.04039 0.040388
4 0.993924 -0.10658 0.106579
5 0.992371 -0.12076 0.120756
5 -0.01388 -0.04039 0.040388
6 0.990859 -0.1334 0.133397 6 0.01207 -0.04039 0.040388
7 0.989286 -0.14491 0.144906 7 -0.01991 -0.04039 0.040388
8 0.987664 -0.15553 0.155533 8 -0.01828 -0.04039 0.040388
9 0.986059 -0.16545 0.165448 9 0.006151 -0.04039 0.040388
10 0.984461 -0.17477 0.174772 10 0.001163 -0.04039 0.040388
11 0.982867 -0.18359 0.183594 11 0.000546 -0.04039 0.040388
12 0.981203 -0.19199 0.191986 12 -0.02382 -0.04039 0.040388
13 0.979487 -0.2 0.199998 13 -0.01881 -0.04039 0.040388
14 0.977765 -0.20768 0.207676 14 -0.00105 -0.04039 0.040388
15 0.976093 -0.21505 0.215054 15 0.015681 -0.04039 0.040388
16 0.974403 -0.22216 0.222163 16 -0.0069 -0.04039 0.040388
17 0.972687 -0.22903 0.229028 17 -0.01021 -0.04039 0.040388
18 0.971032 -0.23567 0.23567 18 0.020193 -0.04039 0.040388
19 0.969341 -0.24211 0.242109 19 -0.01134 -0.04039 0.040388
20 0.96764 -0.24836 0.248359 20 -0.0045 -0.04039 0.040388
21 0.965923 -0.25443 0.254434
21 -0.00569 -0.04039 0.040388
22 0.964233 -0.26035 0.260347
22 0.006341 -0.04039 0.040388
23 0.962523 -0.26611 0.266109
23 -0.00534 -0.04039 0.040388
24 0.960767 -0.27173 0.271728
24 -0.0162 -0.04039 0.040388
25 0.958952 -0.27721 0.277214
26 0.957167 -0.28257 0.282573 25 -0.02049 -0.04039 0.040388
27 0.955385 -0.28781 0.287813 26 0.009074 -0.04039 0.040388
PACF removes the correlation of the inbetween lags. So for
28 0.953625 -0.29294 0.292941 27 0.000618 -0.04039 0.040388 example, PACF of Yt and Yt-2, doesn’t include the
29 0.951842 -0.29796 0.297962 28 0.006545 -0.04039 0.040388 correlation of Yt-1 and Yt-2.
30 0.950099 -0.30288 0.302881 29 -0.00898 -0.04039 0.040388
PACF suggests that Yt is significantly correlated with Yt-1
30 0.011644 -0.04039 0.040388 only.
Trend Stationary
Yt = β1 + β2t + ut
While running the regression use t=1,2,3,…. as one independent
variable.
Difference Stationary
Yt = Yt−1 + ut
OR
Yt = β1 +Yt−1 + ut

While running the regression use Yt - Yt−1 as dependent variable.

 0
0.01
0.02
0.03

-0.02

-0.03
-0.01
1
58
115
172
229
286
343
400
457
514
571
628
685
742
799
856
913
970
1027
1084
1141
dlnex

1198
1255
1312
1369
1426
1483
1540
1597
1654
1711
1768
1825
1882
1939
1996
2053
2110
2167
log(ex)t-log(ex)t-1

2224
2281
2338
Integrated Time Series
• A time series is said to be integrated if the series becomes stationary
after differencing.
- Integrated of order 1, I(1): if series becomes stationary after
differencing it once.
- Integrated of order 2, I(2): if series becomes stationary after
differencing it twice (difference of difference)