Meet this gentleman!!!

He is
struggling
with …

He is Member, CENTRAL BOARD OF DIRECT TAXES.

 200000
400000
600000
800000
1000000
1200000
1400000
1600000

0
1950-51

1952-53

1954-55

1956-57

1958-59

1960-61

1962-63

1964-65

1966-67

1968-69

1970-71

1972-73

1974-75

1976-77

1978-79
Can you

1980-81
he is doing

1982-83
imagine what

with this data?

1984-85

1986-87

1988-89

1990-91
TOTAL TAX COLLECTED - INDIA

1992-93

1994-95

1996-97

1998-99

2000-01

2002-03

2004-05

2006-07

2008-09

2010-
11(R.E.)
Revenue Forecasting is highly
technical and professional job!!!
Let’s meet another person. He is working in
the Corporate Planning Cell of NTPC.
 (In GWH)

20000
40000
60000
80000
100000
120000

0
1950-51

1955-56

1960-61

1965-66

with this data?

1968-69

what he is doing
Can you imagine
1973-74

1978-79

1979-80

1984-85

1989-90

1991-92

1996-97

2000-01

2001-02

2002-03
Total Electricity Generation Units in India

2003-04

2004-05

2005-06

2006-07

2007-08

2008-09

2009-10
He struggling to forecast
future electricity generation so
that he can plan for capacity
additions for NTPC!!!
 1,200
1,300
1,400
1,500
1,600
1,700
1,800
1,900

1-Jan-13
8-Jan-13
15-Jan-13
22-Jan-13
29-Jan-13
5-Feb-13
12-Feb-13
19-Feb-13
26-Feb-13
5-Mar-13
12-Mar-13
19-Mar-13
26-Mar-13
2-Apr-13
9-Apr-13
16-Apr-13
23-Apr-13
share? Imagine…

30-Apr-13
7-May-13
14-May-13
21-May-13
28-May-13
4-Jun-13
11-Jun-13
18-Jun-13
25-Jun-13
2-Jul-13
9-Jul-13
The WHOLE CLASS WILL THINK

16-Jul-13
23-Jul-13
30-Jul-13
Maruti - Close Price

6-Aug-13
13-Aug-13
20-Aug-13
27-Aug-13
3-Sep-13
10-Sep-13
17-Sep-13
What can one do with the Price Data of a

24-Sep-13
1-Oct-13
8-Oct-13
15-Oct-13
22-Oct-13
29-Oct-13
5-Nov-13
12-Nov-13
19-Nov-13
26-Nov-13
3-Dec-13
10-Dec-13
17-Dec-13
24-Dec-13
31-Dec-13
7-Jan-14
Once issues are clear to us…

Then • We can move ahead

and enter the area of
… …
 “Forecasting is very
dangerous, especially about
the future.” --- Samuel Goldwyn

And, we are daring

to enter into it!!!
Entering into
the WORLD
of time
series!!!
Let’s reiterate our issues related to Time Series

 Issues:

 UNDERSTANDING THE BEHAVIOUR

 ESTIMATION & FORECASTING

 ESTABLISHING THE RELATION

After understanding the
basics of a time series, let’s
move to the challenges …
What do say
about the
Prediction?
?
What is the
challenge?

What do you
say about
this point?

o
It is a STRUCTURAL CHANGE POINT!!!

Such structural break points

make our job of prediction,
estimation and establishing
relation more difficult.
How to take care of
such structural break is
one of the challenge
with which we shall be
confronting!
Now, let’s move to …
SOME INTERESTING TIME
SERIES GRAPHS!
Let’s look at the following graph of a time series…

A SIMULATED TIME SERIES

What do say about the

Prediction related to
such a time series?
Let’s look at another graph of a time series…

x(t)

600 What do say about

the Prediction
500 related to such a
time series?
400

300

200

100

0
1 501 1001 1501 2001 2501 3001 3501 4001 4501 5001

-100

Time
One more graph of a time series…

What do say about

the Prediction
related to such a
time series?
These charts are showing a fact that time series
are not stationary.

x(t)

600

500

400

300

200

100

0
1 501 1001 1501 2001 2501 3001 3501 4001 4501 5001

-100

Time
What if …you perform
regression analysis on the non-
stationary time series?....You
will be trouble!
We may get…

SPURIOUS REGRESSION !!!

 … ONE MUST ENSURE
To avoid SPURIOUS REGRESSION …

THAT THE SERIES IS

STATIONARY!!!
To get consistent forecasts and other

… ONE MUST ENSURE

results …

THAT THE SERIES IS

STATIONARY!!!
And, this takes to the discussion on …

STATIONARITY
 STATIONARITY …

 …is an idea which talks about the

stationarity of the parameters of a time
series.
 Loosely speaking a stationary process is one
whose statistical properties do not change
over time

 We have two concepts of Stationarity …

STRONG STATIONARITY …

 A time series Yt is said to be having

STRONG STATIONARY –

if the joint probability distribution of

any m observations made at times t1, t2,

…, tm is the same as the joint statistical

distribution m observations made at
times t1 + k, t2 + k, …, tm + k
STRONG STATIONARITY …

 This is an extremely strong definition:

it means that all moments of all
degrees (expectations, variances,
third order and higher) of the
process, anywhere are the same.
STATIONARITY …

 Since the definition of strict stationarity is

generally too strict for everyday life a
weaker definition of second order or weak
stationarity is usually used
Weak Stationarity

 A time series is said to be weakly

stationary if it has –
• Constant mean
• Constant variance

• Cov (rt, rt-l) = γl, which only depends on l .

Or we can say

 Weak stationarity means that mean

and the variance of a stochastic
process do not depend on t (that is
they are constant)

 The autocovariance between Xt and

Xt+τ only can depend on the lag τ

More formally, it may be defined as …

 A time series Yt is said to be weakly

stationary if for all t the following holds

1. E(Yt) = m < ∞
true:

2. V(Yt) = E(Yt - m)2 = g0 < ∞

3. Cov(Yt , Yt-k) = E{(Yt - m)(Yt-k – m ) = gk

 It is also called Second Order Stationarity

or Covariance Stationarity.
Kindly note that…

Stationary Stationary
Process Time Series
Characteristics of a Stationary Process

 It is mean reverting.

 It has a finite memory.

 The impact of a shock dies out with time.

 It has a finite variance.

 It has a finite mean.

Stationary

 Stationary can be
 Trend stationery
 Intercept Stationary
 Trend and intercept Stationary
TRENDS…

 … can create a problem in obtaining the

consistent and normally distributed OLS
estimates.

 …trends may make our estimates at times

unstable.
TREND IS …

 … A persistent upward or downward

movements of variables over a period of
time.

 …Observed not in short-term but over a

long period of time.
TRENDS…

 Trends are of TWO types:

 Deterministic Trends: E (Y ) - E (Y ) 
t t -1 1
• The trending variable changes by a constant
amount each period
 Stochastic Trends:
E(Yt ) - E(Yt -1 )   vt
• The trending variable changes by a random
amount each period (vt)
 Trends …

SALES - DETERMINISTIC TREND

35,000

30,000

25,000

20,000 SALES - STOCHASTIC

TREND
15,000 160,000

140,000
10,000

120,000
5,000
100,000
0
0 10 20 30 40 50 60 70 80 90 100 80,000

60,000

40,000

20,000

0
1 7 13 19 25 31 37 43 49 55 61 67 73 79 85 91 97
 Non-Stationary Time Series…

Trend Stationary: If Yt contains a deterministic trend and

{Yt – Trend} becomes stationary; then Yt is known as trend
stationary.

Difference Stationary: If Yt contains a stochastic trend and

{Yt – Yt-1 } becomes stationary; then Yt is known as
difference stationary.

Remember that these can be used

as a strategy to make a non-
stationary time series as
STATIONARY TIME SERIES
Time series plot…
 A Simple Time Series…

REALIZATION OF A VALUE WHICH IS PURELY RANDOM FOLLOWING NORMAL

PROBABILITY DISTRIBUTION
4.0000

3.0000

2.0000

1.0000

0.0000

-1.0000

-2.0000

-3.0000

-4.0000
Another Time Series…..

REALIZATION OF A VALUE WHICH IS PURELY RANDOM FOLLOWING

NORMAL PROBABILITY DISTRIBUTION WITH A DRIFT
33.0000

32.0000

31.0000

30.0000

29.0000

28.0000

27.0000

26.0000
Look at this graph?

REALIZATION OF A VALUE WHICH IS HAVING SHOCKS

What’s your reaction?

REALIZATION OF A VALUE WHICH IS HAVING SHOCKS

Look at this graph?

REALIZATION OF A VALUE WHICH IS HAVING SHOCKS

What’s your reaction?

REALIZATION OF A VALUE WHICH IS HAVING SHOCKS

HOW TO TEST WHETHER IS
STATIONARY?
Testing for
STATIONARITY if
caused by Deterministic
or Stochastic Trend.

This takes us to the discussion

on UNIT ROOT TEST
Let’s introduce Autoregressive Model – AR(1)

 For the AR(1) Model, we have…

Yt = a + r Yt-1 + et

 Its moments are…


E (Yt )  From it, we can infer
1 
that |r|<1, if we want
2 to have a series as
Var (Yt )  stationary.
1  2
 s 2
Cov(Yt ,Yt  s ) 
1  2
UNIT ROOT Stationarity Tests…

 Dickey-Fuller Test

 Augmented Dickey-Fuller Test

 The Phillips-PerronTest
 UNIT ROOT Stationarity Tests…

 Dickey-Fuller Test …presumes that the error

term in the Model are not correlated.
 Augmented Dickey-Fuller Test …presumes
that the error term in the Model are correlated.
 The Phillips-PerronTest …presumes that the
error term in the Model are correlated and they
are heteroscedastic.

Null Hypothesis in these tests are -

The series has UNIT ROOT.
GOODNESS OF FIT

INDICATORS:
Adjusted R2
Standard Error of Regression
Akaike Information Criterion (AIC)
Schwarz Bayesian Criterion or Bayesian Information Criterion (BIC)

Overall Test: F-Test

t-Test for Individual Estimates
INFORMATION-BASED Criteria…

 Information-Based criteria are measures of the goodness of fit

of an estimated statistical model.

 They are founded on the concept of Entropy.

 Offers a relative measure of the information lost

 Describes the tradeoff between simplicity and complexity of

the model.
INFORMATION-BASED Criteria…

 Given a data set, several competing models may be ranked

according to their Information Criterion.

 Lower the value of an information based criterion of a model, the

better is model in terms of Goodness of fit.
INFORMATION-BASED Criteria… (Continued…)

 An Information-Based criteria rewards goodness of fit, but also

includes a penalty that is an increasing function of the number of
estimated parameters.

 This penalty discourages overfitting.

 The Information Criterion based methodology attempts to find the

model that best explains the data with a minimum of free
parameters.
We have the following information based criteria…

 Akaike Information Criteria (AIC)

 Bayesian Information Criteria (BIC) or Schwarz

criterion (SIC)

 Hannan-Quinn Criteria (HQ)

 Akaike Information Criteria (AIC)

 The Akaike information criterion (AIC) is a measure of the relative

quality of statistical models for a given set of data

 AIC estimates the quality of each model, relative to each of the other
models. Hence, AIC provides a means for model selection.
 Akaike Information Criteria (AIC)

 There are several information criteria available to determine the order

p of an AR process, all of them are likelihood based.
 For example, Akaike information criterion (AIC) (Akaike, 1973) is
defined as.

 AIC =
 K= k be the number of estimated Beta(s) + 1
 Log(l) = l be the maximum value of the likelihood function
Schwarz Information Criterion (SIC)

 The Schwarz Information Criterion (SIC), also known as the Bayesian

Information Criterion (BIC), is used for model selection. It balances the goodness
of fit and model complexity.
 Formula for BIC/SIC:

 BIC =
 where:
• n = number of observations (sample size)
• k= number of parameters in the model
• RSS= Residual Sum of Squares
Hannan-Quinn Information Criterion (HQIC)

 The Hannan-Quinn Information Criterion (HQIC) is a statistical criterion used

for model selection, particularly in time series and econometrics. It is similar to the
Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC),
but it has a different penalty term.
 The formula for HQIC is:

 HQIC =
 where:
• n = number of observations (sample size)
• k= number of parameters in the model
• RSS= Residual Sum of Squares