Nityananda Sarkar

Indian Statistical Institute

203 B. T. Road, Kolkata - 700108
 A time series is a sequence of observations
over time. What makes it distinguishable
from other statistical analyses is the explicit
recognition of the importance of the order
in which the observations are made. Also,
unlike many other problems where
observations are independent, in time
series observations are most often
dependent.
 Prediction of the future based on knowledge
of the past (most important).
 To control the process producing the series.
 To have a description of the salient features
of the series.
 Economic planning
 Sales forecasting
 Inventory (stock) control
 Evaluation of alternative economic
strategies
 Budgetting
 Financial risk management
 Model evaluation
 Old/Classical/Traditional Approach
 Modern/Stochastic Process Approach

Although the modern approach is a more

general one, it is not that the classical
approach has no importance at all in
analysing time series data.
As the saying goes, OLD IS GOLD.
 A Time series is a set of observations, each one being
recorded at a specific time. (Annual GDP of a country,
Sales figure etc)
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
1801
1807
1813
1819
1825
1831
1837
1843
1849
1855
1861
1867
1873
1879
1885
1891
1897
1903
1909
1915
1921
1927
1933
1939
Stock Returns

1945
1951
1957
1963
1969
1975
1981
1987
1993
1999
2005
Stock Returns
 Classical decomposition model: a time series
is considered to consist of three components
viz., Xt= mt+st+ct+t
where mt, st, ct and t are trend, seasonality,
cyclical and random (noise) components.
 0
50
100
200
250
300
350
400

150
01/01/1950
01/01/1953
01/01/1956
01/01/1959
01/01/1962
01/01/1965
01/01/1968
01/01/1971
01/01/1974
01/01/1977
01/01/1980
01/01/1983
01/01/1986
01/01/1989
Sunspots

01/01/1992
01/01/1995
01/01/1998
01/01/2001
01/01/2004
01/01/2007
01/01/2010
Sunspots
 Generally a plot speaks a thousands
words.
 However, plot needs to be studied
carefully.
 Following plots elaborate---plot does
not speak so many words as well----
need for mathematical modeling.
 y
0 20 40 60 80

0
5
t
10
15
20
y
0 2000 4000 6000 8000 10000

0
5
t
10
15
20
 Log y y
2.5 3 3.5 4 4.5 0 20 40 60 80

0
0

5
5

t
t

10
10

15
15

20
20

Log y y
2.5 3 3.5 4 4.5 0 20 40 60 80

0
0

1
1

Log t
Log t

2
2

3
3
 Cyclical Component: The oscillatory
component in a time series where the
period of oscillation is more than one
year.
One complete period is called a cycle.
Length of cycle, intensity of cycle,
occurrence of cycle are not fixed and
unpredictable.
EXAMPLE: Every economy goes through
prosperity (boom) and depression.
 0
50
100
200
250
300
350
400

150
01/01/1950

01/01/1953

01/01/1956

01/01/1959

01/01/1962

01/01/1965

01/01/1968

01/01/1971

01/01/1974

01/01/1977

01/01/1980

01/01/1983

01/01/1986

01/01/1989
Sunspots

01/01/1992

01/01/1995

01/01/1998

01/01/2001

01/01/2004

01/01/2007

01/01/2010
Sunspots
 0
50

-50
150
200
250
300
350

100
01/01/1955

01/02/1955
01/03/1955

01/04/1955
01/05/1955

01/06/1955
01/07/1955

01/08/1955

01/09/1955
01/10/1955

01/11/1955
01/12/1955

01/01/1956

01/02/1956
01/03/1956

01/04/1956
Sunspots

01/05/1956

01/06/1956
01/07/1956

01/08/1956

01/09/1956
01/10/1956

01/11/1956
01/12/1956
Sunspots
Trendline
Assume that Trend is linear.

Assume that there is no cyclical component.

We will discuss two methods of trend

determination (estimation).

Method1: Moving average Method (MA).

Method 2: Mathematical curve fitting method.

Step 1: We calculate the average of first ‘n’
observations.

Step 2: Next we drop the first observation and

include the ‘n+1th’ observation.

This process is repeated until the last

observation is reached.
Remarks: If ‘n’ is odd, the average of the first
‘n’ t h observations is tabulated with the
n 1
2
time point. The average of the next ‘n’
n  1  1 th
observations is tabulated with the 2 time
point and so on.

Step 2: Next we drop the first observation and

include the ‘n+1th’ observation.

This process is repeated until the last

observation is reached.
 0
1
2
3
4
5
6
7
8
9
1990m1
1990m9
1991m5
1992m1
1992m9
1993m5
1994m1
1994m9
1995m5
MONTHLY DATA

1996m1
1996m9
1997m5
1998m1
1998m9
1999m5
average
millions)
12 month Moving
Sale of warm blankets(in
 Determination of ‘n’.
For monthly data: n=12, 24, 36…
For Quarterly data: n=4,8,…
 Few data points are lost (at the beginning and
at the end)
 Cannot be used for forecasting.
 Does not work well when the trend line is
nonlinear.
 The method of moving average can successfully remove seasonal
variations if the period of the moving average is equal to or a
multiple of the period of the seasonal fluctuations; else the
fluctuations will not be removed completely.
10

8
Sale of warm blankets(in
6 millions)
5 month Moving
4
average
2 17 month Moving
Average
0
1990m11

1992m12

1993m10

1995m11

1997m12

1998m10
1995m6

1996m4
1990m1
1990m6

1991m4
1991m9
1992m2
1992m7

1993m5

1994m3
1994m8
1995m1

1996m9
1997m2
1997m7

1998m5

1999m3
1999m8
Note: These averages are so calculated that
seasonality does not get completely removed
from the data.
9
8
7
6
5 Sale of warm
4 blankets(in millions)
3 Average yearly values
2
1
0
1990m11

1992m12

1993m10

1995m11

1997m12

1998m10
1990m1
1990m6

1991m4
1991m9
1992m2
1992m7

1993m5

1994m3
1994m8
1995m1
1995m6

1996m4
1996m9
1997m2
1997m7

1998m5

1999m3
1999m8
 18000

16000

14000
Observation
Per Hector Wheat yield in Kg

12000

10000

Yield
8000

6000

4000

2000

0
1945 1950 1955 1960 1965 1970

time
 From the above plot, different people may fit
different line based on their naked eye
judgment.
 Which one is the best ‘in some sense’?
 The method is known as ‘Least Square
Method’.
 This method assumes a particular form of
trend, say, linear Trend, in the form as
Tt =a + b x t
where a and b are unknown, to be estimated
from the data.
 ILLUSTRATION ON TREND and ITS REMOVAL-I

Unemployment level for Belgium (Jan 1982 - Aug 1994)

640,000

600,000

560,000

520,000

480,000

440,000

400,000

360,000
82 83 84 85 86 87 88 89 90 91 92 93 94
Dependent Variable: Unemployment Level Belgium
Sample: 1982M01 1994M08
Included observations: 152

Variable Coefficient Std. Error t-ratio p-value

constant 524254.3 9224.385 56.83353 0.0000

t (time) 4904.821 530.7924 9.240564 0.0000
t2 -124.4672 8.183658 -15.20923 0.0000
t3 0.646678 0.035619 18.15533 0.0000

Note: Deterministic trend is removed first by the OLS method

 Residuals after deterministic trend fit
700,000

600,000

80,000
500,000

40,000
400,000

0
300,000

-40,000

-80,000
82 83 84 85 86 87 88 89 90 91 92 93 94

Residual Actual Fitted

 Detrended Unemployment level
80,000

60,000

40,000

20,000

-20,000

-40,000

-60,000

-80,000
82 83 84 85 86 87 88 89 90 91 92 93 94
Dependent Variable: Detrended Unemployment Level
Sample: 1982M01 1994M08
Included observations: 152

Variable Coefficient Std. Error t-ratio p value

constant 17779.69 1903.941 9.338364 0.0000

@SEAS(2) -17535.44 5280.581 -3.320740 0.0011
@SEAS(3) -31606.49 5280.581 -5.985418 0.0000
@SEAS(4) -44510.50 5280.581 -8.429091 0.0000
@SEAS(5) -55341.97 5280.581 -10.48028 0.0000
@SEAS(6) -58891.23 5280.581 -11.15241 0.0000

Note: Deterministic seasonality is removed from detrended (deterministic)

data. Only significant seasonal dummies are reported.
 Residuals of seasonality adjusted detrended unemployment level
80,000

40,000

0
60,000

40,000 -40,000
20,000
-80,000
0

-20,000

-40,000

-60,000
82 83 84 85 86 87 88 89 90 91 92 93 94

Residual Actual Fitted

Dependent Variable: Unemployment Level of Belgium
Sample: 1982M01 1994M08
Included observations: 152

Variable Coefficient Std. Error t ratio p-value

constant 516994.0 6323.958 81.75166 0.0000

@SEAS(4) -36511.80 19688.04 -1.854516 0.0657
@SEAS(5) -46263.03 19688.04 -2.349803 0.0201
@SEAS(6) -48686.10 19688.04 -2.472877 0.0145

Note: Instead of detrending first, deseasonalisation can as well be done in the

first step. This slide shows the computation to this end.
 Deseasonalized Unemployment Level (Belgium)
150,000

100,000

50,000

-50,000

-100,000

-150,000
82 83 84 85 86 87 88 89 90 91 92 93 94
Dependent Variable: Deseasonalized Unemployment Level
Sample: 1982M01 1994M08
Included observations: 152

Variable Coefficient Std. Error t-ratio p-value

constant 18895.48 6359.941 2.971015 0.0035

t 4909.100 365.9657 13.41410 0.0000

t2 -125.0055 5.642390 -22.15470 0.0000

t3 0.650591 0.024558 26.49159 0.0000

Note: Detrending of the deseasonalise series is done in the second step

 Residuals after a trend fit to the deseasonalized unemployment level
200,000

100,000

60,000
0
40,000

20,000 -100,000

0
-200,000
-20,000

-40,000

-60,000
82 83 84 85 86 87 88 89 90 91 92 93 94

Residual Actual Fitted

Dependent Variable: Belgium Unemployment Level
Sample: 1982M01 1994M08
Included observations: 152

Variable Coefficient Std. Error t-ratio p-value

constant 544141.9 5868.978 92.71493 0.0000

t 4822.603 326.8123 14.75649 0.0000
t2 -123.9371 5.037261 -24.60406 0.0000
t3 0.646789 0.021921 29.50483 0.0000
@SEAS(2) -17626.75 5332.856 -3.305311 0.0012
@SEAS(3) -31695.97 5330.839 -5.945774 0.0000
@SEAS(4) -44599.25 5329.683 -8.368088 0.0000
@SEAS(5) -55431.11 5329.379 -10.40105 0.0000
@SEAS(6) -58981.87 5329.927 -11.06617 0.0000
 Residuals of trend and seasonality fit to unemployment level
700,000

600,000

60,000 500,000
40,000

20,000 400,000

0
300,000
-20,000
-40,000
-60,000
-80,000
82 83 84 85 86 87 88 89 90 91 92 93 94

Residual Actual Fitted

Second Illustration

IP INDEX (Apr1995 - Sep 2008)

300

275

250

225

200

175

150

125

100
95 96 97 98 99 00 01 02 03 04 05 06 07 08
Dependent Variable: IP_INDEX
Sample: 1995M04 2008M09
Included observations: 162

Variable Coefficient Std. Error t-ratio p-value

constant 126.3000 2.114454 59.73170 0.0000

t 0.157815 0.060684 2.600619 0.0102
t2 0.005034 0.000365 13.80040 0.0000
 residuals after deterministic trend fit
300

250

40 200

30
150
20

10 100

-10

-20
95 96 97 98 99 00 01 02 03 04 05 06 07 08

Residual Actual Fitted

Dependent Variable: DETRENDED IP INDEX
Sample: 1995M04 2008M09
Included observations: 162

Variable Coefficient Std. Error t-ratio p-value

constant 8.762640 1.137970 7.700237 0.0000

@SEAS(2) -7.497238 1.971022 -3.803731 0.0002
@SEAS(3) 8.375794 1.971022 4.249467 0.0000
@SEAS(4) -12.23991 1.923521 -6.363286 0.0000
@SEAS(5) -10.73814 1.923521 -5.582547 0.0000
@SEAS(6) -14.98216 1.923521 -7.788925 0.0000
@SEAS(7) -13.60767 1.923521 -7.074356 0.0000
@SEAS(8) -14.40039 1.923521 -7.486477 0.0000
@SEAS(9) -12.20319 1.923521 -6.344193 0.0000
@SEAS(10) -14.10544 1.971022 -7.156408 0.0000
@SEAS(11) -11.78444 1.971022 -5.978847 0.0000
 residuals after fitting seasonality on detrended IP Index
40

30

20

10

30 0

20 -10

-20
10

-10

-20
95 96 97 98 99 00 01 02 03 04 05 06 07 08

Residual Actual Fitted

Dependent Variable: IP_INDEX
Sample: 1995M04 2008M09
Included observations: 162

Variable Coefficient Std. Error t-ratio p-value

constant 136.0743 1.740104 78.19895 0.0000

t 0.125352 0.039068 3.208549 0.0016
t2 0.005223 0.000235 22.23451 0.0000
@SEAS(2) -7.494443 1.979679 -3.785685 0.0002
@SEAS(3) 8.379950 1.979777 4.232775 0.0000
@SEAS(4) -12.30740 1.933625 -6.364936 0.0000
@SEAS(5) -10.80277 1.933489 -5.587188 0.0000
@SEAS(6) -15.04429 1.933421 -7.781177 0.0000
@SEAS(7) -13.66768 1.933421 -7.069173 0.0000
@SEAS(8) -14.45867 1.933489 -7.478022 0.0000
@SEAS(9) -12.26010 1.933625 -6.340474 0.0000
@SEAS(10) -14.11186 1.979777 -7.128004 0.0000
@SEAS(11) -11.78799 1.979679 -5.954494 0.0000
 Residuals after fitting Trend and Seasonality
300

250

30
200
20

150
10

0 100

-10

-20
95 96 97 98 99 00 01 02 03 04 05 06 07 08

Residual Actual Fitted

Dependent Variable: IP_INDEX
Sample: 1995M04 2008M09
Included observations: 162

Variable Coefficient Std. Error t-ratio p-value

constant 190.3462 13.43910 14.16361 0.0000

@SEAS(1) 0.392308 19.00576 0.020642 0.9836
@SEAS(2) -5.838462 19.00576 -0.307194 0.7591
@SEAS(3) 11.02308 19.00576 0.579986 0.5628
@SEAS(4) -12.80330 18.66328 -0.686015 0.4938
@SEAS(5) -10.35330 18.66328 -0.554741 0.5799
@SEAS(6) -13.63901 18.66328 -0.730794 0.4660
@SEAS(7) -11.29615 18.66328 -0.605261 0.5459
@SEAS(8) -11.11044 18.66328 -0.595310 0.5525
@SEAS(9) -7.924725 18.66328 -0.424616 0.6717
@SEAS(10) -16.30000 19.00576 -0.857635 0.3925
@SEAS(11) -13.03077 19.00576 -0.685622 0.4940
 Residuals after fitting seasonality to IP Index
300

250

120 200

80 150

40
100
0

-40

-80
95 96 97 98 99 00 01 02 03 04 05 06 07 08

Residual Actual Fitted

 Stationary series in obtained by removing the
three deterministic components mt,st and ct,
from xt either by “estimating” these
components and then subtracting these from
xt OR by taking differences
 xt= xt- xt-1,
2 xt=  xt- xt-1
=xt-xt-1-xt-1+xt-2
=xt-2xt-1+xt-2
Twelve period difference (to remove seasonality) of Belgium unemployment level
100,000

75,000

50,000

25,000

-25,000

-50,000

-75,000

-100,000
82 83 84 85 86 87 88 89 90 91 92 93 94
 First difference of the deseasonalized unemployment level
20,000

15,000

10,000

5,000

-5,000

-10,000

-15,000

-20,000
82 83 84 85 86 87 88 89 90 91 92 93 94
 twelve period difference(to remove seasonality) of IP index
60

50

40

30

20

10

-10
95 96 97 98 99 00 01 02 03 04 05 06 07 08
 first difference of deseasonalized IP Index
40

30

20

10

-10

-20

-30

-40

-50
95 96 97 98 99 00 01 02 03 04 05 06 07 08
 It is natural to assume that an observed series
(x1,x2,…,xn) is a particular realisation of a
stochastic process.
 A stochastic process is a family of random
variables defined on a probability space.
Specifying the complete form of the
probability distribution is very ambitious, and
we usually content ourselves with
concentrating attention on the first and the
second moments.
 However, it is impossible to make inferences
based on the first and the second moments
only. So we make a very important
assumption called stationarity.
 A stochastic process is said to be strictly
stationary if its properties are unaffected by a
change of time origin; in other words, the
joint probability distribution at any set of
times t1,t2, …,tk must be the same as the joint
probability distribution at times t1+m, t2+m, …,
tk+m, where m is any arbitrary shift in the time
axis. Obviously, strict stationarity is a strong
condition.
 A stochastic process is said to be weakly
stationary (or covariance stationary) if
(i) E(x1) = E(x2) = … = E(xn)=  (say)
(ii) V(x1) = V(x2) = … = V(xn)= 2 (say)
and (iii) Cov (xt,xt-k) = k, independent of t.
Obviously, strong stationarity  weak
stationarity but not vice versa. Only when
{xt) are jointly normal, then both are
equivalent.
Ergodicity
We also make the assumption of
ergodicity which refers to one type of
asymptotic independence. In words,
asymptotic independence means that two
realizations of a time series become even
closer to independence, the further they are
apart with respect to time. More formally,
asymptotic independence can be defined
as
 | F ( X 1 ,..., X n , X 1k , X 2k ,..., X nk )
 F ( X1,..., X n ) F ( X1 k ,..., X n  k ) | (1)

 0 as k  .
This means that the joint distribution of two
subsequences of a stochastic process { X t } is
equal to the product of the marginal distribution
functions the more distant the two subsequences
are from each other. A stationary stochastic
process is ergodic if
 1 n 
lim   E ( X t   ) ( X t  k   )  0 (2)

 k 1
n  n

holds. This condition stated above would be
satisfied if the autocovariances tend to zero
with increasing k.
 Sample Autocorrelations and Partial
Autocorrelations
While a time series plots give some idea about
the nature of the underlying times series, it is
not (always easy or possible) to conclude
whether the time series is stationary or not.
Therefore, it is useful to consider some
statistics related to a time series. To that end,
sample autocorrelations , ̂ k , are obtained as
ˆ k  ˆ k / ˆ 0 where ˆ k and ˆ0 are the estimates of

the autocovariance of lag k and variance , respectively, and

these are given by
n
ˆk   ( xt  x ) ( xt  k  x ) / n
t  k 1

n
and ˆ0   ( xt  x ) 2 / n
t 1
 n
where x   xt / n and n is the number of
t 1
observations in a given time series .
For a series with stationary DGP, the sample
autocorrelations typically die out quickly with
increasing k. In contrast, the sample autocorrelations
decays rather slowly for a non-stationary series. Note
that sample autocorrelations are estimates of the actual
(unknown) autocorrelations of the time series if the
process is stationary.
Partial autocorrelations are also useful
quantities that may convey useful information
on the properties of the DGP of a given time
series. The partial autocorrelation between X t
and X t  k is the autocorrelation between X t
and X t  k conditional on the intermediate
variables X t 1 , X t  2 , …, X t k 1 i.e.,
it measures the “true” correlation between X t
and X t  k after adjustments have been made
for the intervening lags. Formally, the k th
sample partial autocorrelation is the estimated
value of  kk , which is,
obtained as the ordinary least squares (OLS)
estimate of the coefficient  kk in the
following autoregressive model of order k
x     x  ...   x . u (3)
t k1 t 1 kk t  k t

For stationary processes, partial

autocorrelations also approach zero as k tends
to infinity, and hence the sample partial
autocorrelations should be small for large
lags.
It may be worthwhile to state that
autocorrelation functions and partial
autocorrelation functions give useful
information on specific properties (i.e., nature as
well as order of the underlying stationary model)
of a DGP other than stationarity.
 Further analysis is carried out with stationary
(weak) time series only. Using correlogram
analysis, as proposed by Box and Jenkins
(1970), which means using the
autocorrelation function (ACF) and partial
autocorrelation function (PACF), we can
tentatively identify a stationary time series
i.e., whether it is autoregressive process of
order p (AR (p)) or moving average process of
order q (MA(q)) or the mixed ARMA (p,q)
process.
 For an AR(p) process, the ACF is infinite in
extent, but the PACF cuts off at lag  p+1.
But, the picture is quite the opposite for an
MA(q) process. Here, the ACF cuts off at lag q
+1 onwards whereas the PACF is infinite in
extent. For the mixed ARMA process, none of
the criteria of ACF and PACF cuts off at some
lag value – both are infinite in extent.
 Turning now to calculating the acf, first calculate the
autocovariances:
1 = Cov(yt, yt-1) = E[yt-E(yt)][yt-1-E(yt-1)]
E(yt) = 0 and E(yt-1) = 0, so 1 = E[ytyt-1]
1 = E[ t 1 t 1 1 ut 2  ...) (ut 1  1ut 2  1 ut 3  ...) ]
    2 2
(u u

= E[ 1 ut 1  1 ut 2  ...  cross  products]

2 3 2

=  1 2
  
3 2
1  1   ...
5 2

1 2
= (1  12 )
For the second autocorrelation coefficient,
2 = Cov(yt, yt-2) = E[yt-E(yt)][yt-2-E(yt-2)]
Using the same rules as applied above for the lag 1 covariance
2 = E[ytyt-2]
= E[ (ut  1ut 1  12ut  2  ...) (ut 2  1ut 3  12 ut 4  ...)
]

= E[ 1 u t 2  1 u t 3  ...  cross  products]

2 2 4 2

= 12 2  14 2  ...

= 12 2 (1  12  14  ...)

12 2
=
(1  12 )
 0
0 = 1
0  2 2 
 1  
 
 1 2   2 
 2 
 (1  1 )   (1  1 ) 
1   2  
1 =   1 2 =   12
0 
 2 
 0  
 2 
 2  
 (1   1 ) 2 
 (1  1 ) 
3 = 3
1
 
 
…
s =  1
s

 ( 2  1 )
2
 22 0
(1   1 )
2

so PACF helps to determine the order of AR process for a finite order

stationary AR process.
 Autoregressive Process of order p (AR(p)):
xt=1xt-1+  2xt-2+… + p xt-p+at
Moving Average Process of order q (MA(q)):
xt=at-1at-1- 2at-2-… -qat-q

ARMA (p,q):
xt- 1xt-1-…- Pxt-p=at- 1at-1 -..-
qat-q.
where at’s are white noise (0, 2a).
Not-stationary== Non-stationary, when
distribution (parameters) changes over time.
Various important examples are: Deterministic
trend and Stochastic trend.
Trend, as we all know, is the long-run
smooth movement of a time series. While
the idea of deterministic trend has been
there for a long time, the concept of
stochastic trend is relatively new.
Deterministic trend can be represented in terms
of the following model:
xt  f (t )  ut (4)

where f (t )is any finite degree polynomial in

time i.e., f (t )represents the trend, and ut is a
stationary series with mean zero and variance
 u2 . Nelson and Plosser (1982) called such
processes as trend stationary processes (TSP).
It is obvious that trend defined in (4) is
deterministic in nature and uncertainty is
introduced only through the stationary
component ut which is temporary in nature,
and consequently, the uncertainty, as measured
by variance around forecast, is also bounded
even for the long-distance period.
If, on the other hand, xt is generated by the
model
xt    xt 1  ut (5)
where ut is, as before, a stationary series with
mean zero and variance  u2 , then xt is said to
follow a difference stationary process (DSP)
since the first differencing of xt makes the
series to be stationary, and the underlying trend
is then called the stochastic trend.
It is easy to find that in this case both mean
and variance are functions of time. Further,
unlike deterministic trend where the trend is
completely predictable, stochastic trend
exhibits a systematic variation though the
variation is hardly predictable.
As regards appropriate methods for
removing trend from these two processes
given in (4) and (5), note that the methods
differ.
In case of deterministic trend, f(t) is estimated
under a specific assumption about the trend
function and then it is subtracted from xt . For
stochastic trend, first differencing makes the
series free of trend. However, if differencing
is used for TSP, then it ensures stationarity of
the series, but introduces a non-invertible
moving average component involving ut .
In fact, Nelson and Kong (1981) have shown
that the use of first differencing to eliminate
the linear deterministic trend of a TSP series
with white noise error would create a spurious
first lag autocorrelation of the magnitude of
-0.5 among the residuals. Also, the presence of
unit root in the moving average component
creates a problem for maximum likelihood
estimation of the moving average parameter.
On the contrary, if the true process is DSP
(with errors not exhibiting any cycle) and
trend removal is done by a regression on time
(treating it as a TSP), then the detrended series
exhibits spurious periodicity (Nelson and
Kong (1981)). It is, therefore, advisable that
detrending of a time series so as to achieve
stationarity, is to be done with caution.
 Yt     t  ut
E (Yt )     t

See that mean changes over time.

One can apply OLS to estimate the

model parameters.
 Yt  Yt 1  ut Yt  Y0  u1  ...  ut 1  ut

E (Yt )  Y0  E (u1 )  ...  E (un )  Y0

 Y2  population variance of (Y0  u1  ...  ut 1  ut )
t

 population variance of (u1  ...  ut 1  ut )

  u2  ...   u2   u2
 t u2

This process is known as random walk.

 Yt    Yt 1  ut
Yt  t  Y0  u1  ...  ut 1  ut
E (Yt )  Y0  t

  t
2
Yt
2
u

This process is known as random walk with drift.

 Fig 2 Trend Stationary Process

250

200
values

150

100
1 101 201 301 401
Time
 Figure 1: Pure Random Walk

130

120
Values

110

100

90
1 101 201 301 401

Time
 Fig 3: Random Walk with Drift

700

600

500
Values

400

300

200

100
1 101 201 301 401
Time
20

15
Random walk with drift
10

5
Stationary process
0
1 11 21 31 41 51 61 71 81 91
-5

-10
Random walk

-15
 This is the most important question before
undertaking any time series analysis.
 The answer requires an understanding of the
modern concept of trend. Trend can be of
two types. Examples of these two types of
trend are given below.
 Deterministic trend:yt=+bt+ct2+at
 Stochastic trend : yt=+yt-1+at where at is a
stationary process.
 Note that in the latter, yt can be expressed as
yt=y0+ t+ aj
so that it also has linear trend as in the case
of deterministic trend. The difference lies in
the fact that the noise term i.e.,  aj is no
longer a stationary process since its mean
and variance are now functions of t. Hence, in
this case although a systematic variation is
exhibited, it is not predictable.
 Underlying model:
yt=+yt-1+at i.e., yt=yt-1+at, = -1
 Null hypothesis H0: =1 / =0
 Alternative hypothesis H1: <1 / <0
 Estimating equation:
yt = yt-1+ j yt-j+at
 The test, called the augmented Dickey – Fuller
test (ADF) which is originally due to Dickey &
Fuller (1979, 1981), requires testing for the
significance of  .
 Main problem: The distribution of the test
statistic does not follow any standard
distribution under H0. The critical values have
been computed by Dickey and Fuller. Note
that the most general form of the estimating
equation is
yt=0+ 1t+yt-1+ j yt-j+at .
The critical values are now different from the
original DF critical values.It may be noted that
seasonal dummies may also be included; but
this would not result in further changes in the
limiting distributions, and hence in the
critical values.
 LNEX_RATE
3.9

3.8

3.7

3.6

3.5

3.4
95 96 97 98 99 00 01 02 03 04 05 06 07 08
Null Hypothesis: LNEX_RATE has a unit root

Exogenous: Constant, Linear Trend

Lag Length: 1 (Automatic based on SIC, MAXLAG=13)

t-ratio p-value

Augmented Dickey-Fuller
test statistic -2.237333 0.4653

Test critical values:

1% level -4.016433
5% level -3.438154
10% level -3.143345
 DLNEX_RATE
.08

.06

.04

.02

.00

-.02

-.04

-.06

-.08
95 96 97 98 99 00 01 02 03 04 05 06 07 08
Null Hypothesis: DLNEX_RATE has a unit root

Exogenous: Constant, Linear Trend

Lag Length: 0 (Automatic based on SIC, MAXLAG=13)

t-ratio p-value

Augmented Dickey-Fuller
test statistic -8.700974 0.0000*

Test critical values:

(* indicates significance) 1% level -4.016433

5% level -3.438154

10% level -3.143345

 One important drawback of this type of time
series modelling is the fact that whereas in
such analyses the parameters are assumed to
be constant all throughout, in reality
parameters may not remain constant over the
entire sample period for which observations
are available, especially if the span of the
data is (moderately) large. The problem in its
entirety is called the problem of STRUCTURAL
CHANGE.
 Exchange rate return (3rd JAN 2000 - 2nd JULY 2009)
4

-1

-2

-3

-4
500 1000 1500 2000
Quandt-Andrews unknown breakpoint test
Null Hypothesis: No breakpoints within trimmed data
Equation Sample: 3 2392
Test Sample: 362 2033

Statistic Value Prob.

Maximum LR F-statistic (Obs. 439) 9.282480 0.0367

Maximum Wald F-statistic (Obs. 439) 9.282480 0.0367

Exp LR F-statistic 2.172070 0.0427

Exp Wald F-statistic 2.172070 0.0427

Ave LR F-statistic 2.748680 0.0556

Ave Wald F-statistic 2.748680 0.0556
 In the context of unit root tests also, the
implicit assumption is that the deterministic
trend is correctly specified. Perron (1989)
argued that if there is a break in the
deterministic trend, the unit root test will lead
to a misleading conclusion that there is a unit
root, when in fact, there is not.
 This problem is extremely important both at
the level of stationary as well as
nonstationary data analysis. At the level of
stationary data, it was Chow (1960) who first
talked about this problem and proposed a
test under the assumption of an a priori
known break point and assumption of
constant variances. The assumption of known
break point has been severely criticized.
 Now, we have Andrews’ (1993, 2003) test to
take care of this problem. Further, Bai (1997)
has estimated the break point(s). As regards
unit root tests under structural change,
Perron (1989) has proposed a modified ADF
test under three different types of
deterministic trend function:
 I. Crash model: A one – time change in the
intercept of the trend function.
 II. Changing growth model: This allows for a
change in the slope of the trend function
without any sudden change in the level at the
time of the break.
 III. Both effects are allowed.