You are on page 1of 25

APPLICATIONS OF STATISTICS

In this Section we intend to publish articles/reports on


INTERESTING LIVE APPLICATIONS OF STATISTICS.
The emphasis will not be so much on theoretical originality, but rather
on novelty of application, solution of real-life problems and readabil-
ity. If necessary, representative supporting data may be included as
part of the communication. The object is to open a forum for inter-
action between Statisticians on the one hand and Scientists working
in diverse fields on the other, as well as among Statisticians with dif-
ferent specializations.

Each communication to this Section should ideally cover not


more than fifteen printed pages inclusive of representative data sets
(However, the upper limit is relaxable at the discretion of the Edito-
rial Board). At the time of submission the authors should specifically
mention that the communication is meant for the Section 'Applica-
tions of Statistics'.

The Council of Calcutta Statistical Association has instituted


an annual award entitled
BOSE- NANDI AWARD
in memory of late Professors P.K. Bose and H.K. Nandi. The award
would consist of a cash prize of Rupees Ten Thousand only*
(payable in Indian currency) and a certificate as a token of recog-
nition and would be given to the best publication in the Section 'Ap-
plications of Statistics' in the issues of each volume of the Bulletin
containing such a Section. The best publication in each volume will
be chosen by an independent board of jury to be appointed by the
Council every year.

Scientists in the fields of Agriculture, Life Sciences, Physical


Sciences, Meteorology, Earth Sciences, Social Sciences and other ap-
plie~ areas are welcome to contribute to this Section. Modalities for
communicating papers are given within 'Instructions to Authors'.

* The Council of Calcutta Statistical Association in its meeting has


resolved that the cash prize money for Bose-Nandi Award be raised
to Rupees Ten thousand only with effect from Volume 51.
Calcutta Statistical Association Bulletin
Vol. 53, Sept. & Dec. 2002, Nos. 211-212

FORECASTING INFLATION RATE BY STAR


MODEL: AN INDIAN EXPERIENCE

G.P. SAMANTA
Dept. of Statistical Analysis and Computer Services,
Reserve Bank of India , Mumbai, India

ABSTRACT: In this empirical study, an attempt has been made to model


non-linear dynamics of inflation rate in India through Smooth Transition /
Threshold Auto-Regression (STAR). Inflation is measured based on weekly
data on Wholesale Price Index (WPI) fur a period of seven years from the
week ended April 2, 1994 to the week ended March 31, 2001. The log(WPI)
series is detected to be a Difference-Stationary process, indicating that the se-
ries is non-stationary but its first-order difference is stationary. The generat-
ing process of the transformed-stationary series is identified to be non-linear.
Six variants of STAR model are estimated for transformed-stationary series
and are used to forecast \V'PI and annual inflation rate. Empirical assessment
of out-of-sample forecast errors • eveals that estimated STAR models perform
reasonably well in generating short-run forecasts of both the variables.

Keyw01·ds and phrases: Forecasting, Time Series Analysis, Stationarity,


Nonlinearity, Smooth Transition Auto-Regression (STAR)Model

AMS (2000} Subject Classification: Primary 91B84; Secondary 621\110

1. INTRODUCTION

The expected level of future inflation rate plays an i~1portant role


in the decision making process of economic agents and policy makers.
The assessment of expected level, however, is a challenging task. A
convenient strategy in this regard has been to build a model for ex-
plaining inflation behaviour and use the model to generate forecasts .
For model building one may take help of existing economic theories
of inflation. But researchers are well aware that such economic-theory
266 Calcutta Statistical Association Bulletin

based models, particularly macro-econometric models consisting of si-


multaneous equations, perform well to explain but not necessarily to
forecast inflation. In order to improve forec'ast accuracy, in recent years,
considerable attention is being paid to statistical and/or completely
data-driven models. In this context, time series analysis is gaining
increasing importance and modelling non-linear dynamics of inflation
through univariate time series models has been an important theme in
recent years.
A number of univariate models, such as, Bilinear (Granger and
Anderson, 1978; Subba Rao, 1981), State Dependent Model (Priest-
ley, 1980), Exponential Auto-Regression (EAR) (Haggan and Ozaki,
1981), Auto-Regressive Conditional Heteroscedasticity (ARCH) and
generalized-ARCH models (Bollerslev, 1986; Engle, 1982), Self-Exciting
Threshold Auto-Regression (SETAR) (Tong, 1983), Smooth Transi-
tion/Threshold Auto-Regression (STAR) (Chan and Tong, 1986; Grange;
and Terasvirta, 1993; Terasvirta, 1994), are useful for modelling non-
linearity. But a major problem in handling non-linearity lies in the
difficulty in identifying the best-suited non-linear form (such as Bilin-
ear, EAR, SETAR, etc.) for a time series.
For the lack of any clear-cut guideline on how to detect appropri-
ate form of non-linearity, a popular strategy has been to build different
types of non-linear model and to choose the best one based on some cri-
teria, say out-of-sample forecast accuracy. Alternatively, forecasts from
alternative models can be suitably combined to arrive at a final forecast
(Bradley, 1982, 1986; Tibilletti, 1994; Deutch, et.al. 1994). Another
sensible strategy would involve building a non-linear model, which can
approximate a wider class of non-linearity. In this case, one may hope
that though the exact type/form of non-linearity remains unknown,
the approximating model would take care of that to a gr~at extent. A
useful model in this regard is Threshold Auto-Regression (TAR), which
can approximate a class of non-linearity covering Bilinear, EAR, State
Dependent Model, etc., almost surely (Petruccelli, 1992). As SETAR
represents the univariate form of TAR model, it is likely that SETAR
model approximates a wide class of univariate non-linearity. In real life
application, Samanta (1999) also found the suitability of SETAR model
for forecasting a number of important Indian economic variables. As
STAR is further generalization to SETAR, it is expected that STAR
will be reasonably good to handle unknown non-linearity. With this
background, this paper remains at building STAR model for forecast-
ing inflation rate in Indian economy. The organization of the rest of the
paper is as follows. Section 2 provides different forms of STAR models,
Inflation Rate by STAR Model 267

and the model building strategy adopted in this paper. Section 3 be-
gins with a discussion on the database used in this·study and thereafter
examines the stationarity and linearity /non-linearity of the underlying
series. In Section 4, six variants of STAR model are estimated and
forecast performances of fitted models are assessed. Finally, Section 5
concludes.

2. SMOOTH THRESHOLD/TRANSITION AUTO-REGRESSION


(STAR) MODEL

2.1 Structures of Different STAR Models

A simple STAR model of order p for a time series Xi is expressed


as

where o:i 's and /3i 's, j 2: 0, are constants; p is a positive integer; F[.] is
a continuous function taking values from [0,1] and is called a transition
function; d is a positive integer and is called a delay parameter; r is
the threshold parameter; 'Y > 0 is scale/smoothness parameter and et
is the usual residual series.
The model in equation (2.1) can be called 'smooth threshold auto-
regression' (Chan and Tong, 1986; Tong, 1990) or 'smooth transition
al!to-regression' (Granger and Teriisvirta, 1993; Terasvirta, 1994). Re-
placing the parameter 'Y by its reciprocal, 'F[.] can also be expressed as
a function of 'Y•(Xt-d- r), where 'Y* = 1/'Y· Moreover, different re-
strictions may be imposed on the function F[.] . For example, Chan and
Tong (1986) and Tong (1990) assume F[.] to be cumulative distribu-
tion function of N(O, 1) variable and hence a non-decreasing continuous
function whereas Granger and Teriisvirta (1993) and Teriisvirta (1994)
assume F[.] to be continuous function, which may also be an even func-
tion. In any case, if F[.] tends to zero (this may occur, for instance,
when F[.] is non-decreasing, Xt-d < r and 'Y is adequately close to
zero), the coefficient of Xt-j tends to O:j. Alternatively,· when F[.] is
close to another extreme value 1, the coefficient of Xt-j is close to
(o:j + /3j) · At any time point, the value of F[.]lies in the interval [0,1]
and a STAR model has coefficients in between two extreme regimes
characterized by the values O:j and O:j + /3j, j 2: 0. In limiting case if
F[.] takes only two possible values, 0 and 1, then equation (2.1) repre-
sents a specific form of a SETAR model.
268 Calcutta Statistical Association Bulletin

In equation (2.1) non-linearity is triggered by the gap between Xt-d


and a threshold-value r. A further generalization to this model would
be (Granger and Terasvirta, 1993)
p p p
Xt =no+.~ aiXt-j+(.Bo+ .~ .BjXt-j)F[(Xt-d-ao- .~ ajXt-d-j)h]+et
J=1 J=1 J=1
(2.2)
The non-linearity in equation (2.2) is due to the deviation of Xt-d
from a linear function/path. This type of STAR model is denoted by
STAR-Deviation or simply STAR-dD.
The lag-structure in equations (2.1) and (2.2) includes all consectt-
tive lags from 1 to p. In practice, one may prefer subset lag selection
by dropping insignificant intermediate lags. In either case, depending
on the functional form ofF[.] and the determining/triggering factor of
non-linearity, a number of STAR models can be formulated. Six vari-
ants of STAR model proposed in the literature, are given in Table 1.

Table 1 : Different Forms of STAR Model


Type/Name of STAR Model Notation Used in Functional form of F[.J
this study
(A) STAR Models Using Equation {2.1) (Simple-STAR Models)
Normal • STAR NSTAR ~l':.J. is distribution function
of N(O, 1) random variable
Logistical STAR LSTAR F[zj I ex z
Exponential STAR ESTAR F[z] = 1 - exp( -z")
(B) STAR Models Using Equation (2.2} (STA R-D Models}
Normal• STAR-Deviation NSTAR-D F!·J. is distribution function
of N(O, 1) random variable
Logistical STAR-Deviation LSTAR-D F[z] I+ex z
Exponential STAR-Deviation ESTAR-D F[z] - 1 - exp( -z')

Note: '*'The wm·d 'normal' is used to emphasize that the corresponding transition
function F[.] is the cumulative distribution function of N(O , l) vm·iable.

2.2 Model Identification and Estimation

A major problem in modelling time series is that the series has a sin-
gle replication. The stationarity is a convenient assumption/requirement,
which ensures statistical stability of the behaviour of the series over
time and enables one to identify and estimate a model based on a sin-
gle replication. Thus model identification and estimation is done for
stationary time series.
Inflation Rate by STAR Model 269

2.2.1 Available Strategies and Techniques

For building a STAR model, one has to identify /select the model
structure (i.e. the form of STAR model and its structural parameters
lag-structurefp and d) and to estimate unknown parameters for the
identified model structure. The algorithm proposed by Chan and Tong
(1986) is useful in this regard. For a NSTAR model, they estimated
the parameters <p = (a 0 , a: 1 , ... , a:p), '1/• = ((30 ,!31 , . . . , f3v) and 6 = (1·, '/')
by minimizing the conditional sum of squares Qn(<p, '1/J, 6), defined as

~
j=m+l
(xt- (a:o + f
j=l
ajXt-j)
2

-(fJo + jEl ,ajxt-j)F[(Xt-d- 1·)hl) (2.3)

where, n= no. of observations, m= maximum of p and d.


When the delay parameter (d) and lag-structure/order are known, the
algorithm works as follows. For fixed 6, Qn<p, '¢, 6) is function of (<p, '1/J)
only and (<p, '¢) may be estimated under ordinary least-square (OLS)
framework by solving linear normal equations. Let, (cp(6), '1/~(6)) be the
estimate of (<p, '1/J) for given 6. Now consider the variable projectional
functional

Qn(cp(6), ~(6), 6) =min Qn(<p, 'l/1, 6)


<p,,P

If Qn(cp(6), ~(6), 6) has its mininum at 6 = J, then (cp(J), ~(J), J) is the


estimate of (<p, '1/J, 6) . For minimizing Qn(cp(6), ~(6), 6) over 6, Chan
and Tong (1986) used the derivative-free algorithm proposed by Pow-
ell (1964) and discussed by Quandt (1983). The parameter 6 may be
estimated through grid-search method also. The threshold parame-
ter (r) can be estimated, as generally is the practice in case of SETAR
model (Tong, 1983, 1990; Chan and Cheung, 1994), through grid-search
over a set of quantilesforder-statistics or some values in between min-
imum and maximum of the empirical distribution of Xt. Similarly, 1'
can be estimated using a grid around the sample standard deviation
(Terasvirta, 1994).
In practice, both lag-structure and d are unknown and one has to
estimate these parameters. Chan and Tong (1986) suggested to adopt
grid-search over a set of potential values of (p, d), on the basis of some
model selection criteria, such as AIC, BIC, SC (Judge, et.al. 1988,
Tong, 1983).
The strategy proposed by Chan and Tong (1986), involving esti-
mation of (i) (<p, '1/J) by solving linear normal equations and (ii) 6 by
270 Calcutta Statistical Association Bulletin

derivative-free method, has straightforward application to estimating


any simple-STAR model. However, there are certain problems in ap-
plying this technique to a STAR-D model. for in a STAR-D model
there is no threshold parameter, so o contains only the smoothing pa-
rameter "(, and assuming m = (p +d), the expression for conditional
sum of squares Qn(cp, 'l/1, o) based on n observations would be modified
as

In this case, for given structural parameters and fixed o,mimmiza-


tion of conditional sum of squares requires to solve non-linear normal
equations and therefore, one has to apply a general optimization tech-
nique. For so doing , one finds several techniques - starting from gen-
eral minimization principle using both first and second-order deriva-
tives, to those using only first-order derivatives. to the Powell's (1964)
derivative-free method (Granger and Terasvirta, 1993) . A systematic
step-by-step strategy for identifying structural parameters of a STAR
model is available in Granger and Teriisvirta (1993) and Terasvirta
(1994).

2.2.2 Building STAR Models - Strategy Adopted

In this empirical study. we followed Chan and Tong (1986) for esti-
mating (cp, 'lj;, o) for any Simple-STAR model. In case of STAR-D mod-
els, we decided to estimate all of cp, ·lj1 and o using direct application
of derivative-free method (Powell, 1964; also see Quandt, 1983). The
identification of unknown lag-structure and d are carried out by adopt-
ing grid-search method. To simplify this task , we followed a strategy
broadly similar to that of Granger and Terasvirta (1993) and Terasvirta
(1994) . Accordingly, we first identify lag-structure of best-fitted linear
auto-regressive model using the method proposed by Box and Jenkins
(1976) . The identified lag-structure is then supplied to all STAR mod-
els and the delay parameter d is estimated through grid-search using
AIC.
For implementing Powells' (1964) algorithm, we need to prov:ide the
initial estimates of parameters of interest. We consider sample mean
and standard deviation as initial estimates of 7' and "' respectively. In
the case of a STAR-D model, we need to provide the initial estimate of
Inflation Rate by STAR Model 271

(r.p, '1/J) also. But the literature is not very clear about how to derive this
initial estimate. As a heuristic choice, initial estimate of (r.p, 'ljJ) is cho-
sen as follows; for given (p, d) and initial estimate of 15, initial estimate
of r.p is first obtained by regressing X 1 on (1,X 1 _ 1,X1 _ 2 .•• ,X1-p)·
Let. u 1 be the corresponding residual series. Now regressing n 1 on
(F[·ut-d/T]), Xt-1F[ut-d/T], Xt_zF[ut-dh], ... , Xt-pF[nt-d/T]) initial
estimate of '1/J is derived. For general lag-structure (which need not in-
clude consecutive lags), the process may be modified by dropping the
unimportant lags in both these regressions.

3. DATABASE AND EMPIRICAL INVESTIGATION ON STATIONARITY


AND NON-LINEARITY

As is well-known, the term 'inflation' is an economic concept, which


refers to sustained rise in the general price level of an economy. The
empirical measurement of the concept, however, is very difficult and
as of now none of the available methodologies can measure the ground
reality. General practice is to proxy general price level by a suitable
price index and consider the change in the chosen price index as the
measure of inflation rate. Thus weekly {annual) percentage change in
the index is weekly (annual) inflation rate and in general, k-period of
percentage change in the index is the estimate of k-period inflation rate.

3.1 Derivation of Inflation Rate

In India, official measure of inflation is derived from Wholesale Price


Index (WPI); we also used WPI for the purpose. The latest series on
WPI in India pertains to base year 1993-94 (Government of India, 1999)
and weekly information on the series is available since the week ended
April 2, 1994. This study uses WPI for 366 weeks, from the week
ended April 2, 1994 to the week ended March 31, 2001 and focusses
on forecasting WPI and annual inflation rate as both these variables
are important for policy purposes. Based on our database, annual (52-
weeks) inflation rate is simply defined as,

1rt = [w Pit- w Pit-sz] x 100 {3.1)


WPit-52
where 7rt and WPI 1 denote annual inflation rate and WPI respectively
for t-th week. The weekly information on WPI1 and 7rt are plotted
in Graph 1 and Graph 2 respectively. From these graphs it appears
272 Calcutta Statistical Association Bulletin

that though WPI follows an increasing path over time, no clear trend
is exhibited by 7rt· Thus WPI series appears to be non-stationary and
before building any STAR model there is a need to examine the series
for stationarity.

Graph 1: Weekly WPI (Base 1993-94=100)


170 .
160 1

~ 1so 1 ,...__/"·~

~::~k
:!: 120 ------
110

100 ... ... ...


...~----------------------------------------------------------
:g :g ,.._ ,.._ ,.._ ,.._ co co co co "' "' co "'
.
10 10 10
c;
<D <D 0 0 0 0

"'ci "'"'...,..!.:::> "'~t; "'"'u "'"'.!. "'"'3 b. u


10
"'
m
"'m "'
.. ..., '1b. u .. '1 9u (ij
0> 0> 0> 0> 0> 0> 0> 0> 0> 0> 0> 0> C> 0 0 0 0
....<:>
c "'0.
0> 0> m m m co m
0> 0> 0> 0> 0> 0> 0> 0> 0> 0 0 0 0

c ,,_ .,
c:J
~ N N
·.:, C: 6. u
~

.!. l;. .!. c 0. 0Q> ~


N q 9 1 7 en 0
:::> :::> ...,
:::> ...,:::> (/)<U
'" ...,::> '"
Q> Q. Q> Q> Q> Q>
Q> Q> Q> Q>
::;; <U
::;;
1 ~
~
0 7 ,.._ ~ 0 ~ 0 0

"'Week«>(Ended
(J) (J)
0 6 0 co .... «> cD cD co ,.:_ ,.:_ cD v'> .;, .;, .;. .;, .;, .;.
;;;
~
N
M M M N N N N N N N N N N N N N N N N
"" N N N

the Date)·····>

Graph 2: Annual (52-Weeks) Inflation Rate


12 .00

i
1\
10.00

-
Cll 8.00

..
r;_ 6 .00
g 400
nl
;;:::: 2.00
c:
000
,.._ ,.._ ,.._ ,.._ co co co co
:g
.. ...,. b. 2l . .
m m m <::> c;
"'
10 It) II) <D <D <D 0> 0 0 0
m m m ·O> m m m m m

..
0>

..
0> 0> 0> 0> 0> 0> 0> 0> 0> 0 0 0 0
m m 0> 0> 0> 0> 0> 0> 0> 0> 0> 0> m 0> 0> 0> m O:• 0> 0 0 0 0 0
~
u c:::J b. u c u ~ c:::> a. u
N
~
6. u
N N
~ ~ ~
..!.
0. .!. 0 .!. .!. 0. .!.
ci :::> Q>
c
:::J .,.., 2l ~ ...,:::> Q>
..., t?<U 0 Q>
c
:::> Q> ~
:z ... ::;;
Q> Q> Q>
7
Q> Q>
::;;
1
"' "' :b
~ 0
«> '1
~ ~
,.:_ cD .;, .;, v'> 7
0 0 (J)
,.._ ,.:_ «>
(J)
0 0 0 «> ,.:_ cD .;, .;, .;.
""
M M M N N N N N N N N N N N N N N N N N N N

Week (Ended the Date)·····>

3.2 Empirical Investigation on Stationarity

As is well known, for stability of univariate time series model, sta-


tionarity is a basic required assumption. Thus before building any
STAR model, it is necessary to examine whether the underlying series
is stationary. If original series is non-stationary, there is a need to ap-
ply suitable transformation to achieve stationarity. As WPI constitutes
the basic database (inflation rate is a derived series using WPI) in the
Inflation Rate by STAR Model 273

present study, the same has been examined for stationarity. Augmented
Dickey-Fuller (ADF) test has been used for the purpose.

3.2.1 Augmented Dickey-Fuller (ADF)Test

The Augmented Dickey-Fuller (ADF) tests are extensively used in


the literature for examining stationarity of a series and also for discrim-
inating between two non-stationary classes; Difference Stationary (DS)
and Trend Stationary (TS) (Dickey and Fuller, 1979, 1981). When
a series is non-stationary, knowledge on whether it belongs to DS or
TS class is helpful to suggest transformation for achieving stationarity.
The ADF tests can be carried out based on following general regression
equation for a time series X 1 ;

(3.2)

where a, /3, p and 8k 's are unknown constants, I:k indicates summa-
tion over suitably chosen lags, ~ is the first-difference operator ((i.e.
~Xt = Xt- Xt-d and et is the usual error series.

The regression equations without intercept and/or without time


trend (i.e. when value of !3 .is set to zero) may also be considered
(Dickey and Fuller, 1979, 1981). However, for simplicity in illustration
and practical implementation the general fol'm will be considered in this
study. Here the possibilities of p > 1 and p ::=; -1 are kept outside the
domain of analysis as for such cases X 1 represents an explosive series
for which it is difficult to carry out any formal analysis. Depending
on the value of (a,/3,p) in equation (3.1), three alternative scenarios
may arise. First, if !3 = 0, p < 1, the X 1 is a stationary series. Second,
if !3 = 0, p = 1, then X 1 is a DS series. In addition, a f= 0, the DS
series posseses a non-zero drift term. The first order difference of aDS
series produces a stationary series. Third, (3 f= 0 and pj1 then X 1 is a
TS series. To achieve stationarity in this case, one needs to fit a time
trend for X 1 and subtract the fitted trend from X 1 .
A number of procedures are available for testing the hypothesis re-
lating to alternative scenarios and we employed three of the several test
statistics suggested by Dickey and Fuller (1981), namely the statistics,
(i) <}) 2 for testing the joint hypothesis (a, /3, p) = (0, 0, 1); (ii) <}) 3 for
testing the joint hypothesis (a,/3,p) = (a,O, 1) and (iii) the t-ratio as-
sociated with the estimate of /3 (denoted by the symbol Tf3t), under
the hypothesis (a, /3, p)=(0,0,1). For conducting these tests, one may
use the finite sample percentiles/critical values for relevant test statis-
274 Calcutta Statistical Association Bulletin

tics estimated by Dickey and Fuller (1981) based on simulation exercise.

3.2.2 Results of ADF Test

Empirical results of ADF tests using <I>:.!. <1> 3 and TfJt are presented in
Table 2. Before carrying out these tests, log-transformation of WPI is
taken to reduce/eliminate possible problem due to heteroscedasticity,
changing seasonal effect (Chatfield, 1989) . For loge (WPI) series, it is
seen that while <1> 2 is statistically significant at 1 percent level, <1> 3 is in-
significant at 5 percent level (hence insignificant at 1 percent level also).
Thus, data accept the hypothesis (a,f],p) = (a,O, 1) but not the hy-
pothesis (a,f],p) = (0,0, 1), identifying loge(WPI) as aDS series with
possible non-zero drift. However, the test statistic Tf3t indicates a mild
possibility of a different scenario. Though the observed value of Tf3t is
insignificant at 1 percent level, it is not so far 5 percent level of signifi-
cance, pointing towards the possible presence of time trend (i.e. fJ f:. 0).
We resolve the issue by giving more preference to the joint tests based
on <1> 2 and <1> 3 statistics. Accordingly, we consider loge (WPI) as a DS
process. Therefore, the first order differencing of loge (WPI) would pro-
duce a stationary series. Note that this transformed stationary series
actually represents an approximation to proportional weekly change in
WPI and thus is a proxy for weekly inflation rate. Henceforth, the
transformed series is denoted by ~ log(WPI), where ~ denotes first or-
der difference operator. For further confirmation on stationarity, ADF
tests were also carried out for ~ log(WPI) series. In this case both
<1> 2 and <1> 3 are strongly significant (at 1 percent level) but Tf3t is in-
significant at 5 percent level (Table 2). Significance of <1> 2 indicates
(a, /3, p) f:. (0, 0, 1), significance of <1> 3 means (a, /3, p) f:. (a, 0, 1), and
insignificance of Tf3t means fJ = 0. These results, therefore, indicate
that fJ = 0 and p < 1, which confirm the stationarity of ~ log(WPI)
series.

Table 2: Results of ADF Tests

Variable Optimal Observed Value of the Statistics


Xt Lags$ <1>2 <1>3 Tf3t
log(WPI) 1,4,52 6.7048** 5.9131# 3.4260*
~log (WPI) 1,11,51 85.0542** 127.5413** 0.4320#
m equat1on (2.4); '* ' Insignificant at 1
r
Note: '$' Ind1cate the selected lags of AXt
percent level but significant at 5 percent level; '** ' Significant at 1 percent level of
significance (and hence at 5 percent level also); '#'Insignificant at 5 per·cent level
(and hence insignificant at 1 percent level also).
Inflation Rate by STAR Model 275

3.3 Empirical Investigation on Linearity /Non-Linearity

While building non-linear model like STAR, it is important to exam-


ine whether the transformed (stationary) series exhibits non-linear be-
haviour. If data do not support non-linearity, then logically, one would
model the underlying series under linear framework . In this study, two
general linearity tests, viz., (i) Modified Regression Error Specification
Test (Modified-RESET) and (ii) the test proposed by Tsay (1986) are
used (for empirical application linearity tests, see Nachane and Ray,
1993, 1997-98, among others) .

3.3.1 Modified Regression Error Specification Test (Modified


RESET) for Linearity

The modified RESET for linear-specification of a time series (Ram-


sey, 1969; Thursby and Schmidt, 1977; Granger and Terasvirta, 1993)
can be carried out in three steps:

Step 1: Let the series Xt haven observations. Fix a positive integer


q. Regress Xt on the vector (1mXt-1,Xt-2, .. . , Xt-q) · Let et be the
estimated residual series.

Define S S Ro = :E e;
t

Step 2: For an integer h > 1, regress et on (1, Xt-1, Xt-z, .. . , Xt-q,


X[, Xf, ... , Xf). Let Ut be the corresponding residual series. Define
SSR = :E u;_ It is seen that test with h = 4 gives best result (Thursby
t
and Schmidt, 1977; Granger and Terriivirta, 1993) .

Step 3: Define F = [(SSRo-SSR)/Ih- 1)]_ Under some regularity condi-


SSR/(n-q-h}
tions, F follows F-distribution with (h -1, n-q- h) degrees offreedom
if X t is linear.

3.3.2 Non-Linearity Test Suggested by Tsay

A general non-linearity test suggested by Tsay (1986) involves fol-


lowing three steps.

Step 1: Let the series Xt haven observations. Fix a positive integer


q. Regress Xt on the vector (1, Xt-1, Xt-1, ... , Xt-q) by least square
276 Calcutta Statistical Association Bulletin

technique. Let e; be the estimated residual series.

Step 2: Define Ut = (Xt-I,Xt-2 , ... ,X1-q) and Z 1= transpose of


Vech( u;, Ut) where Vech denotes half-stacking vector and u; is the
transpose of Ut . Now regress the vector Zt on (1, Xt-1, Xt-2, ... , Xt-q)
using least square technique in multivariate framework . Let ~~ be the
corresponding residual vector.

Step 3: Regress e;
on ~~ and let F* be the F-ratio of the mean
square of regression to the mean square error. In other words, fit the
regression equation e;
= ~t!J + C:t, C:t being the residual series and define
the F-ratio as

where Q = q(q + 1)/2. Under the null hypothesis of linearity of X~, F*


follows F -distribution with (Q, n - q - Q - 1) degrees of freedom.

3.3.3 Results of Linearity Tests

The linearity /non-linearity tests discussed above are performed us-


ing weekly data on A log(WPI), which is identified as a stationary pro-
cess earlier. Tests are carried out for four sets of residuals as obtained
by fixing q = 4, 8, 12, 16. Observed values of the test statistics and asso-
ciated p-values for modified~RESET are presented in Table 3. Similar
results for Tsay's test are given in Table 4. As seen from these tables,
the null hypothesis of linearity could not be accepted at 1 percent level
of significance (as corresponding p-values are less than 0.01) in all cases
except when q = 4 for modified-RESET test. Thus there is substantial
empirical evidence to consider A log(WPI) series as a non-linear pro-
cess, though the exact form/type of non-linearity remains unknown.
Inflation Rate by STAR Model 277

Table 3: Results of Modified-RESET Test for Transformed


Series
Value of h-3 h-4 h-5
q F-Statistics p--value F -Statistics p-value F -Statistics p--value
( l) egrees of (Degrees of (Degrees of
Freedom) Freedom) Freedom )
4 2.8477 0.0593 2.9083 0 .0346 3.2402 0.0125
(2,354) (3 ,353 ) (4,352)
8 5.1046 0 .0065 5.2054 0.0016 4.8206 0.0009
(2,346) (3,345) {4,344)
12 7.4716 0.0007 7.5626 0.0001 9.9044 0.0000
(2,338) (3, 337) (4,336)
16 9.3773 0.0001 9.9182 0.0000 14.0720 0.0000
(2,330) (3,329) (4 ,328)

Table 4 : Results of Tsay's Test of Non-Linearity for


Transformed Series
Value of q F-St.atistics (Degrees of Freedom) p--value
4 35.0000 {10,350) 0.0000
8 8 .8889 {36,320) 0 .0000
12 3.5128 {78,274) 0.0000
16 1.5588 (136,212) 0.0019

4. STAR MODELS FOR INFLATION RATE IN INDIA

We build six variants of STAR model for capturing the non-linear


dynamics of ~ log(VlPI) series. Forecast performances of fitted mod-
els are evaluated in terms of the magnitudes of out-of-sample forecast
errors.

4.1 Estimated STAR Models

For building a STAR model, we first identify the lag-structure. Us-


ing Box-Jenkins approach (Box and Jenkins, 1976), the lags of the
best-fitted linear-AR model for ~ log(WPI) series are identified to be
1, 4, 11 and 52. These lags, as stated earlier, are considered for esti-
mating different STAR models. The delay parameter d for a11y model
is estimated through grid-search (based on AIC) over 1, 4, 8, 13, 26,
39 and 52, which are possibly important seasonal lags in weekly data.
For calculating AIC, one essentially uses the estimated residual series.
As seen earlier, for given (p, d), optimization problem to be solved for
278 Calcutta Statistical Association Bulletin

estimating (<p, '1/J, 8) of a STAR model is conditional on first few obser-


vations - the required number of conditioning observations varies on
choice of model-type and (p, d) - and residual can not be estimated
for those conditioning observations. In order to ensure comparability
of estimated AIC for different types of STAR model and for different
alternatives of d, it may be appropriate to use estimated residuals for
same set of observations. This has been accomplished by fixing m-
[maximum chosen lag + maximum of alternative values of d] in both
equation (2.3) and equation (2.4). In our case, maximum lag= 52 and
maximum chosen d= 52, thus m = 104, indicating that the starting
point of estimated residual series corresponds to the 105-th observa-
tion of ~ log(WPI) series (i.e. the week ended April 6, 1996 in our
data period).
It is seen that estimates of AIC (Table 5) for different model types
and possible choices of d are practically same; these estimates vary
within -11.7688 to -11.6540. However, going by the numerically min-
imum AIC, the ESTAR model with (d = 39) gives the best within-
sample fit to the ~ log(WPI) series, followed by ESTAR-D (d= 39)
and LSTAR (d= 39). Also note that for the identified lag-structure,
optimal delay parameters for NSTAR, LSTAR, ESTAR, NSTAR-D and
ESTAR-D models are identified to be 13, 39, 39, 13, 13 and 39 respec-
tively.

Table 5: Estimated AIC-Six Models for Different Delay


Model Delay Parameter (in Weeks)
Type 1 4 8 13 26 39 5:l
NSTAR -11.6594 -11.6579 -11.6786 -11.6824. -11.6684 -11.6810 -11.6fih ..
LSTAR -11.6587 -11.6580 -11.6783 -11.6870 -11.6662 -11.6887' -11.G6!.1c
ESTAR -11.6655 -11.6696 -11.6757 -11.6732 -11.6798 -11.7688' -11.6717 I
NSTAR-D -11.6586 -11.6582 -11.6579 -11.6596' -11.6576 -11.6594 -11.Gf>lll l
LSTAR-D -11.6586 -11.6584 -11.6581 -11.6590' -11.6576 -11.6586 -11.6:,;,~o 1
ESTAR-D -11.6579 -11.6722 -11.6666 -11.6695 -11.6660 -11.7360' -11.66:!1 1
Note : •• ' mdtcates mtmmum m the row of the Table.

Estimated STAR models are presented in Table 6. Actual values


~ log(WPI) series and its fitted values through estimated models are
plotted in Graphs 3.1 to 3.6. For the reason stated above,~ log(WPI)
will be fitted from its 105-th observation, which corresponds to the
week ended April 6, 1996. Fitted values of WPI are generated by ap-
plying inverse-transformation to fitted ~ log(WPI) series; the same for
six STAR models (along with actual WPI) are plotted in Graphs 4.1
to 4.6.
Inflation Rate by STAR Model 279

Table 6: Estimated STAR Models for ~ log(WPI)

NSTAR Model (Delay= 13 weeks)


Xt 0.0001 + 0.5030Xt-1 + 0.3122Xt-4- 0.1193Xt-11
+0.4614Xt-52 + (0.0006- 0.7605Xt-1- 0.3377Xt-4
+0.2998Xt-u- 0.4146Xt-s2)F(Xt-13- .0001)/0.0048] + et
where F[.J is the cumulative distribution function of N(0,1) variable and et is usual
residual series.

LSTAR. Model (Delay = 39 weeks)


Xt = 0.0005 + 0.0976Xt.-1 + 0.1165Xt-4 + 0.3185Xt-ll- 0.3109Xt-52
+(0.0000 + 0.0254Xt-1 + 0.0130Xt-4- 0.6362Xt-11
+1.1707Xt-s2)F(Xt-39- .0021)/0.0048] + et
where F[ .] is the logistic function defined earlier and et is usual residual series.

ESTAR Model (Delay= 39 weeks)


Xt 0.0007 + 0.0909Xt-1 + 0.1369Xt-4 - 0.0195Xt-u - 0.1311Xt-sz
+( -0.0005 + 0.01987Xt-l - 0.0808Xt-4 + 0.2915Xt-ll
+1.2064Xt-sz)F[Xt-39- .0021)/0.0048} + et
where F[ .] is the exponential function defined earlier and et is usual residual series.

NSTAR.-D Model (Delay= 13 weeks)


Xt 0.0005 + 0.1118Xt-1 + 0.1230Xt-4 + 0.0417Xt-ll + 0.2104Xt-52
+(0.0000- 0.0453Xt-l - 0.0249Xt-4 + 0.0585Xt-ll .
-0.0655Xt-s2)F[(Xt-13- Lt-13)/0.0048] + et
where F[.) is the cumulative distribution function of N(O, 1) variable; Lt = 0.0005+
0.1118Xt-1 + 0.1230Xt-4 + 0.0417Xt-ll + 0.2104Xt-sz; and et is usual residual
series.

LSTAR Model (Delay = 13 weeks)


Xt = 0.0005 + 0.1118Xt-1 + 0.1230Xt-4 + 0.0417Xt-ll + 0.2104Xt-52
+(0.0000- 0.0243Xt-1- 0.0161Xt-4 + 0.0479Xt-u
-O.OI-53Xt-s2)F((Xt-13- Lt-13)/0.01) + et
where F(.) is the logistic function define of N(O, 1) variable; Lt = 0.0005+0.1118X1 _, +
0.1230Xt-4 + 0.0417Xt-ll + 0.2104Xt-s2; and et is usual residual series.

ESTAR-D Model (Delay = 39 weeks)


Xt 0.0004 + 0.1078Xt-l + 0.1230Xt-4 + 0.0463Xt-u + 0.2003Xt-52
+(0.0004 + 0.1856Xt-l- 0.0998Xt-4 + 0.0210Xt-ll
+0.3949Xt-sz)F[(Xt-39- Lt-39)/0.0051] + et
where F(.) is the exponential function defined earlier; Lt = 0.0004 + 0.1078Xt-l +
0.1230Xt-4 + 0.0463Xt-ll + 0.2003Xt-s2; and et is usual residual series.
Transformed-WPI -----> Transformed-WPI "---->
b b 0 0 0 0 0
0 0 0 0 0 6 b 0 0 0 0 0
~ 00 0 ~ ~
0 1\J 0 0 0 0 0 0
0 U'l 0 U'l 0 U'l 0 ~
0 ~ ~
0 Transformed-WP~ 1:-.J
0 0 1\J -----> 00
6-ApT-1996 I 0 U'l 0 c.n 0 01 0 0
6-Apr-1996 b b 0 0 0 0 0
0
~
0 0 0 0
~
0 0
6-Jul-1996 0 0 0 N
4t 6-Jul-1996 0 U'l 0 U'l 0 ;;, 0
6-Apr-1996
5-0ct-1996 I .a- G)
... G)
OJ 5-0ct- 1996 ... 6-Jul-1996
OJ ..,G)
4-Jan-1997 I <I' "0
~ "0 Ql
4-Jan-1997 ~ "'0
w 5-0ct-1996
5-Apr-19971 w
~ 5-Apr-1997 ~ 4-Jan-1997 w
5-Jul-1997
j_ ...-t
OJ
-t ..:....
:I
5-Jul-1997 iil 5-Apr-1997 ..,-t
:I Ql
4-0ct-1997 I ~ :I

:e 3-Jan-1.998
...~
3 :E
4-0ct-1997 ~
...
5-Jul-1997
r
<II CD :; 3-Jan-1998 3 4-0ct-1997 ..,~ 0e:.
~
Q. CD
Q. :E
~
3 c
:, 4-Apr- 1998 CD .....
m4·Apr-1998 :; 3-Jan-1998 c. .....
~
gt
:e
"tJ :I
:e
"tJ
~
(/)
Q. 4-Jul-1998 :e
<II Ql g. 4-Jul-1998 m4-Apr-1998 ""0 .....
OJ :I
Ill
.....
:I c.
~ 3-0ct-1998 Q. :I u;·
g. 4-Jul-1998
-ItO
:z :r 3-0ct-1998 c. .....
<II Q. ;:: ;::;·
iJf CD ;:: Ill
i' 2-Jan-1999 Ill - 3-0ct- 1998 e:.
"T1 o 2-Jan-1999 ~ "T1
<II
.. I ;:;: OJ "T1 CD ;:;:
-;; 2-Jan-1999 I I Ul
>
3-Apr-1999 I CT I CT Ul
f '< ~ 3-Apr-1999 CT 0 I 0
I OJ '<
m
- I '<
a; 3-Apr-1999 I
~
it 3-Jul-1999 I r z ~-
(/) (/)
.;.,
I
)> I
I 3-Jul-1999 (/) . I .....
;:::;,; 0 -t v . I -t
- "T1 )> -t o·
~~ )> "T1)> I 3-Jul-1999 ;:;: (") )> ::::1
2-0ct-1999 I (1) c ;::::;.: n )> I
a. Q) :::0 2-0ct-1999 :::0
(1) c :::0 v 2-0ct- 1999
roa. cQ) t::t:l
---
1-Jan-2000 I ~ ~ a. Q) - ~
0 1-Jan-2000 ~ 0
E.
---
c. 0 c. en-
CD a. 1-Jan-2000 CD
.....
CD s·
1-Apr-2-~~~ I ~ - 1-Apr-2000
'-' - 1-Apr-:1opo
1-JUI-LOUU
1-Jul-2000
. 1-Jul-2000
30-Sep-2000 I 30-Sep-2000
30-Sep-2000
30-Dec-2000 I 30-Dec-2000 ·30-Dec-2000
:ll ·Milr-2001 I 31 -Mar-2001 31 -Mar-2001
i f.=-
Transformed-WPI -----> Transformed-WPI -----> Transformed-WPI ----->
6 6 0 0 0 0 0 6 6 0 0 0 0 0 6 6 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0
~
0 0 0 0 0 0 ::l
~
0 0 0 0 ~ ~
0 0 0 N ~
0 0 0 ~
0 0 0 N :::ll
()1 0 ()1 0 ()1 N 0 ()1 0 ()1 0 ()1 0
6-Apr-1996
0 0
6-Apr-1996
0 ()1 0 <.To 0 ()1 0
6-Apr-1996
-
~
.,....

::l
6-Jul-1996
..._., 6-Jul-1996 6-Jul-1996 i ::0
G) G) ll'
.,....
5-0ct-1996
.., .., 5-0ct-1996 (D
rl AI 5-0ct-1996 AI G)
"tt "tt
.., c:r
=r AI «
4-Jan-1997 ~ 4-Jan-1997 =r 4-Jan-1997 "tt
w ;r r:n
~
5-Apr-1997 ..en 5-Apr-1997 ~ 5-Apr-1997 ~ :;2
-1
.., ~ ::0
j_ AI
..,-1
5-Jul-1997 5-Jul-1997 AI 5-Jul_.1997 ..,-1 ~
;:, ;:, AI 0
;:, 0..
4-0ct-1997 3-.., 4-0ct-1997 4-0ct-1997 ~
..,3-
~ 3
C1)
~ ~ ..,3-
CD C1)
3-Jan-1998 3-Jan-1998
3C1)
Q. C1) CD 3-Jan-1998
CD c. C1)
3C1)
';11:'
';11:' c.
!:: 4-Apr-1998 .,~ 4-Apr-1998 4-Apr-1998
m
;:,
m
;:, m
-AI ;:,
Q. c.
.,-~ .,~
4-Jul-1998 4-Jul-1998 AI c. 4-Jul-1998
C1) ;:, C1) I
;:, CD
Q. c. c. c. ' !i'O
~ 3-0ct-1998 3-0ct-1998 3-0ct-1998 J
=r
C1) iii :r
C1) ~ =r =
C1)
0 2-Jan-1999 "TI
0
AI
2-Jan-1999 I

I
::!! 0
- 2-Jan-1999 [
"TI
"'
;:;:
AI ... AI I
I 0" C" C"
~ 3-Apr-1999 C1) 3-Apr-1999 I '<
'< '< CD 3-Apr-1999
I
I ,.... I
I
m ' I z
(fl (fl
- --
I I ,)>
:' (fl
: 3-Jul-1999 t
-1
3-Jul-1999 · - ·o
--
''
3-Jul-1999 ' I -1
I I -1 I' "'1'1 )>
v '
"TI )> )> v' ~- )> v ;::;.: () )>
2-0ct-1999 ::.: () ;o 2-0ct-1999 CDc ;o :::0
CD c
a.ru 2-0ct-1999 <Dc
a.~
a. (U
6 0' 0'
1-Jan-2000 ('.;)
1-Jan-2000
---
!or ~ 1-Jan-2000 s:
0
s: 0 00
0 ......
a. .. A-- .... 1"\1"\1"\ a.
i -AI)r-2000 '=4-··-· · C1) 1- /""\f.JI-L..VVV c.. CD
~- C1)
l"Aor-2000
I
1-Jul-2000 1-Jul-2000 I
1-Jul-2000 i
30-S~p-20UU :O-Scp-2000 I
30-Sep-2000 i
l 1
30-Dec-2000 30-Dec-2000
30-Dec-2000 I
31-Mar-2001 31-Mar-2001
31 -Mar-2001 L_ -
r f
WPI -----> WPI -----> WPI ----->
NNWWAAU'1U10'lO'l
......
.-. .- ......
~ N w w ~ ~ ~ ~ m m
NNWWAAU1U10'lO'l
OU'10U'10U'10U'10U1
-
OU'10UlOU'10UlOU1 0 U1 0 U1 0 (J1 0 L"l 0 (J1 1-.:.
00
6-Apr-1996 6-Apr-1996 6-Apr-1996 - "-=l

6-Jul-1996 6-Jul-1996 6-Jul-1996 ·

5-0ct-1996 5-0ct-1996 5-0ct-1996

4-Jan-1997 4-Jan-1997 4-Jan-1997


C)
C)
5-Apr-1997 C) 5-Apr-1997 iil 5-Apr-~997
iil -a iil
~ -a
-a 5-Jul-1997 ~
5-Jul-1997 ~ 5-Jul-1997 ~
~ !':' ~
4-0ct-1997 w )>
.....
4-0ct-1997 i 4-0ct-1997
n )> ()
::E c n
::E 3-Jan-1998 \J ~ 3-Jan-1998 J Ill ~ 3-Jan-1998 e:..
«< c n
«< Cll !
- <D Ill c
.,....
Ill .,....
_!: 4-Apr-1998 !: 4-Apr-1998 j :e\J !:4-Apr-1998 l;ll
m a m 1 m
:I '
:e\J w
Ill .,....
5. 4-Jul-1998 · ~ 4-Jul-1998 j a 4-Jul-1998
<D
ii :I Cll •go ~
., Q. I ~
a. ;:;: a.. u:r
.,....
... 3-0ct-1998 - 3-0ct-1998 : g. 3-0ct-1998 ii c:;·
~ tr ii .,
<D '< ;- I 'TI <D e:..
;:;:
0 2-Jan-1999 o 2-Jan-1999 ! ;:;: 2-Jan-1999
Dl rn 111 tr Ill tr >
,..
II
'<
...0 '<
iii 3-Apr-1999 ~ 3-Apr-1999 j ~3-Apr-1999 z "'"'0
i!:::0 r en o.
I en i -'1 ~
! 3-Jul-1999 s: 3-Jul-1999 ~ 3-Jul-1999 )> ::::-.
v 0
I
v
!I i! ,l> :::0 §
a. ;l> :::0 ;::;: 0·
2-0ct-1999 ; l> <D ;::;: 0 2-0ct-1999 a; c
;::;.: 0 2-0ct-19991 s: s: to
CDC
0..!!!_ 0
a.. !!!.. 0· c
CD C ~ Q. -
1-.Jan-2000 a. n• 1-Jan-2000 ; ij) 1.. Jan-2CCO .,....
!\ =- i "' c:o

1-Apr-2000 1-Apr-2000 !I 1-Apr-2000
i
1-Ju!-2000 i -Jui-2000 : ~ i-jul-2000
\
30-Sep-2000 30-Sep-2000 i 30-Sep-2000

30-Dec-2000 30-Dec-2000 30-Dec-2000

31-Mar-2001 31-Mar-2001 31-Mar-2001 ~


'
WPI -----> WPI----> WPI----->
NNWW.I>..I>.<.n<.nOlOl NNWWJ:a..l>.<.n<.nOlOl ~;;:iz;;WAA<nc:;;mo;
--~-------
oc.noc.noc.noc.noc.n oc.noc.noc.noc.noc.n
---------- oc.noc.noc.noc.noc.n
6-Apr-1996 t:"
6-Apr-1996 6-Apr-1996 Eb
..,
<""
6-Jul-1996 6-Jul-1996 · 6-Jul-1996 o·
::;l

5-0ct-1996 5-0ct-1996 . 5-0ct-1996 ' :=tl


~
('!)
4-Jan-1997 4-Jan-1997 ' G') 4-Jan-1997 : V'
'<!
G') Al en
5-Apr-1997 5-Apr-1997 "0 5-Apr-1997 ...C)
Al ::r • Ill
"0 'tl ~
5-Jul-1997 ::r 5-Jul-1997 ~ 5-Jul-1997 ::r :::0
~
~
4-0ct-1997 ~
"" 4-0ct-1997 4-0ct-1997
""~ 0
c.
~ ()
~ > ~
~ 3-Jan-1998 ::2 ~ 3-Jan-1998 : I» ~ 3-Jan-1998 ·
CD CD c
CD CD CD Ill
!: 4-Apr-1998

~
:1("
- 4-Apr-1998 ,
~
!: 4-Apr-1998
-
m a. m . ::2.
~ . ~ I»
gw .,~
a. 4-Jul-1998 a. 4-Jul-1998 ' ~ Q. 4-Jul-1998 ·
CD CD Q. CD QO
i.,
a. ;; Q. Q.
:T 3-0ct-1998 - 3-0ct-1998 3-0ct-1998 i
::r i' :T
CD .i CD
c 2-Jan-1999
~I
"TI)>
;::;;o 2-Jan-1999 '
.,;;i CD
2-Jan-1999 =;""
I» me: m ~ CT ~ •CJ'
0.!!!._ '< "TI)> '< .
.!. 3-Apr-1~99 ~ CD 3-Apr-1999 r- CD
.._. 3-Apr-1999 ;:::;: n z
AI (/j co.
-I 0 0.!!!._
c· en
~ 3-Jul-1999 vj 3-Jul-1999
. ·' ~ 3-Jul-1999
3: ~I ;p t i!:;o
0 "TI)>
2-0ct-1999 a. 2-0ct-1999 ;::.:o c 2-0ct-1999 b
CD . me
0.!!!._
3: 3:
1-Jan-2000 1-Jan-2000 : 0 1-Jan-2000 0
Q. Q. h.?
CD CD 00
w
1-Apr-2000 ~ .J\pr-2000 1·Apr-2COO

1-Ju!-2000
" 1-Jul-2000 1-Ju'l-iooo

30-Sep-2000 30-Sep-2000 · 30-Sep-2000

30-Dec-2000 30-Dec-2000 · 30-Dec-2000 '


31-Mar-2001 31-Mar-2001 31-Mar-2001
\ '~
284 Calcutta Statistical Association Bulletin

4.2 Out-of-Sample Forecast Performance

Forecast performances of identified STAR models are assessed based


on the estimated magnitudes of errors in out-of-sample forecasts of
WPI and annual inflation rate. The strategy adopted for assessing the
forecast performance of a model is a..<> follows . Let a time series X 1
have n observations X 1 , X 2 , .. . , X n. For a positive integer l , let Xt (l)
denote the forecast of Xt+l generated using information upto time point
t . The error (E) and percentage error (PE) associated with this l-th
step ahead forecast are

Et(l) = Xt+l-Xt(l); PEt(l) = 100x [Xt+l-Xt(l)]/ Xt+l ; l = 1, 2, 3 . . .


(4.1)
For assessing the overall magnitude of forecast error for a lead period
l, forecasting exercise can be repeated sequentially for several times.
Let forecasting for leads 1 to L be repeated R times (see Ray, 1988
and Samanta, 1999 for the strategy of repeating forecasting exercise).
In each repetition, model structure (i.e. model type, lag-structure and
delay parameter) is assumed to be same with that obtained using all
n observations. Other parameters, viz., tp, 1/J and 6 are re-estimated in
each repetition. In each repetition , one may of course adopt a rigorous
model building strategy rather than assuming a common lag-structure
and delay parameter. For the sake of simplicity and convenience, how-
ever, such an extensive analysis is not attempted here. Based on the
Et(l) and PEt(l) values in different repetitions, two measures of out-
of-sample forecast error. viz. , Root-Mean-Square-Error (RMSE) and
Root-Mean-Square-Percentage-Error (RMSPE) are defined as below;

1 n-L ]
RMSE(l) = E
[ R. t=n-R-L+l [Et(lW ; l=1,2, . . . ,L (4.2)

RMSPE(l) = 1 n-L
E
[ R t=n-R-L+l (PE1 (l)]-·>] ; l = 1,2, ... , L (4.3)

It is easy to see that lower values of RMSE and RMSPE will indicate
better forecast performance. If the variable under forecast exhibits
trend , use of RMSPE as an estimate of forecast error magnitude is
more meaningful (Ray, 1988) . For a stationary series or growth rate
series RMSE is generally used. Accordingly, we calculated RMSPE
values for forecasts of WPI and RMSE values for forecasts of annual
inflation rate. Relevant results for forecast leads 1, 4, 13, 26, 39 and
52 weeks are presented in Table 7 and Table 8. From these tables, it
Inflation Rate by STAR Model 285

appears that the estimated STAR models perform well in generating


forecasts \VPI as well as annual inflation rate in short-run, particu-
larly with leads upto 26-weeks (i.e. half-year). It may be noted that
for weekly data, lead period increases very rapidly even for generating
very short run forecast. For example, 6-months ahead forecast from
weekly data corresponds to 26-s~eps ahead forecast, which for quar-
terly data corresponds to lead of only 2-period. As multi-steps-ahead
forecasting is difficult. these points need to be kept in mind while in-
terpreting forecast errors for weekly series.

Table 7: RMSPE in Out-of-Sample Forecast of WPI


Type of Delay Forecast Lead (I)
STAR (weeks) 1-week 4-wceks 13-weeks 26-weeks 39-wceks 52-weeks
Model• (1-month) (!-Quarter) (Half-year) (9-rnonth) (1-year)
NSTAH. 13 0.31 0 .80 1.51 1.28 1.99 2.09
LSTAR 39 0.30 0.79 1.48 1.28 1.88 1.80
ESTAR 39 0.30 0.77 1.47 1.:.!4 1.91 2.01
NSTAR-D 13 0.30 0.79 1.48 1.25 1.90 1.83
LSTAR-D 13 0.30 0.79 1.48 1.24 1.89 1.81
ESTAR-D 39 0.30 0.80 1.49 1.24 1.89 1.98

Note : '*'Estimates o/1·espcctive Models a1·e given in Table 6.

Table 8: RMSE in Out-of-Sample Forecast of Annual


(52-weeks)lnflation Rate
Type of Delay Forecast Lead (I)
STAR (weeks) 1-week 4-weeks 13-weeks 26-weeks 39-wceks 52-weeks
Model• (1-month) (!-Quarter) (Half-year) (9-month) (1-yearj
NSTAR 13 0.32 0.84 1.59 1.35 2.13 2.25
LSTAR 39 0.32 0 .83 1.56 1.35 2.01 1.94
ESTAR 39 0 .32 0.81 1.55 1.31 2.04 2.16
NSTAR-D 13 0.32 0.83 1.56 1.32 2.03 1.97
LSTAR-D 13 0.32 0.83 1.5() 1.31 2.02 1.95
ESTAR-D 39 0.32 0.80 1.49 1.31 2.02 2.13

Note : '*'Estimates of respective Models are given in Table 6.

5. CONCLUDING REMARKS

In the forecasting literature, a number of non-linear univariate time


series models are proposed. The STAR model is an important non-
linear model, which is generalization to SETAR. As SETAR can ap-
proximate a class of non-linearity almost surely, STAR would be useful
in handling unknown non-linearity. In this empirical study, we examine
the usefulness of STAR model for capturing inflation dynamics in In-
dia. Following the official practice in India, inflation has been measured
based on WPI.
286 Calcutta Statistical Association Bulletin

Using weekly data, it is seen that log(WPI) SERIES is non-stationary


and belongs to a Difference Stationary class. Thus, the first order
difference of log(WPI) is stationary. The generating process of the
transformed-stationary series is detected to be non-linear. Empirical
results show that STAR performs well in capturing this non-linear dy-
namics. Forecast. performances of six variants of STAR model esti-
mated here are quite impressive-error in out-of-sample forecasts of both
WPI and inflation rate are reasonably low, at least for leads upto 26-
weeks. However, the results presented here are exploratory in nature
and further analysis with more extensive plan to identify the type of
non-linearity, lag-structure, etc., may be interesting. Future empirical
research may also investigate the usefulness of STAR models in fore-
casting other economic variables.

Acknowledgement: The author is grateful to an anonymous referee for


offering constructive comments and suggestions on the earlier version
of the paper, without implicating him in any errors that may remain.
The views expressed in the paper are those of the author and not nec-
essarily of the organization he belongs to.

REFERENCES

Bollerslev, T. (1986): Generalized autoregressive conditional heteroskedas


ticity. J. Econometrics 3 307-327.
Box, G.E.P. and .Jenkins, G.M. (1976): Time Series Analysis, Fore-
casting and Control. San Francisco, Holden Day.
Bordley, Robert, F . (1982): The combination offorecasts, a Bayesian
approach. J. Operations Research 2 171-174.
Chan, K.S. and Tong, H. (1986): ·on estimating thresholds in auto-
regressive models. J. Time Series Analysis 7 no.3, 179-190.
Chan, Wai-Sum and Cheung, Siu-Huang (1994): On robust estimation
of threshold autoregressions, J. Forecasting 13 37-49.
Chatfield, C. (1989): Time Series Analysis - An Introduction. 4-th
ed. Chapman & Hall (First edition was published in 1975).
Dickey, D.A. and Fuller, W.A. (1979): Distribution for th€ estimators
for auto-regressive time series with unit root. J. Amer. Stat.
Assoc. 74 427-431.
Inflation Rate by STAR Model 287

Dickey, D.A. and Fuller, W.A. (1981) : Likelihood ratio statistics for
auto-regressive time series with unit root. Econometrika. 89
1057-1072.
Melinda, Deutsch, Clive, W.J. Granger and Terasvirta, Timo (1994):
The combinations of forecasts using changing weights. Int. J.
Forecasting. 10 47-57.
Engle, R.F. (1982) : Auto-regressive conditional heteroskedasticity with
estimates of the variance of the UK inflation. Econometriaka. 50
987-1008.
Government of India {1999): Revision of Index Numbers of Whole-
sale Prices in India - Report of the Working Group. Ministry of
Commerce & Industry, November, New Delhi.
Granger, C.W.J. and Anderson, A.P. (1978): Introduction to Bilinear
Time Series Models. Gottingen: Vandenhock and Ruprecht.
Granger, C.W.J. and Terasvirta, T. (1993): Modelling Nonlinear Eco-
nomic Relationships. Oxford University Press.
Haggan, V. and Ozaki, T. (1981): Modelling non-linear random vi-
brations using an amplitude dependent autoregressive time series
models. Biometrika 88 189-196.
Judge, George G., Carter Hall. R., Griffiths, William. E., Lutkepohl,
Helmut and Lee Tsoung-Chao (1988): Introduction to the Tehory
anrl Practice of Econometrics. 2nd ed. John Wiley & Sons. The
First edition was published in 1982.
Nachane, D.M. and Ray, D. (1993): Modelling exchange rate dynamics
- new perspectives from the frequency domain. J. Forecasting.
12 379-394.
Nachane, D.M. and Ray, D. (1997-98): Non-linear dynamics of the
money multiplier - selected case studies. The Ind. Econ. J. 45,
No.1, 36-53.
Petruccelli, Joseph D. (1992): On the approximation of time series
by threshold auto-regressive models. Sankhya, Ser-B, 54, No.1,
106-123.
Powell, M.J .D. (1964): An efficient method for finding the minimum
of a function of several variables without calculating derivatives.
Computer J. 7 155-162.
288 Calcutta Statistical Association Bulletin

Priestley, M.B. (1980) : State dependent models: a general approach


to non-linear time series analysis. J. Time Series Analysis. 1
No.1, 47-71.

Quandt, Richard E . (1983) : Computational problems and methods, in


Griliches Zvi and Michael D. Intriligator (eds.] (1990), Handbook
of Economics, 1 Third Impression, Elsevier Science Publishers
B.V., North Holland. The First Edition was published in 1983.

Ramsey, J .B. (1969) : Tests for specification errors in classical linear


least-squares regression analysis. J. Roy. Statist. Soc. B, 31
350-371.

Ray, D. (1988) : Comparison of forecasts : an empirical investigation.


Sankhya, Ser-B, 50, 258-277.

Samanta, G.P. (1999) : On forecast performance of SETAR model:


an empirical investigation. J. Quantitative Economics 15, No.1
89-114.

Subba Rao, T. (1981) : On the theory of bilinear time series models.


J. Roy. Statist. Soc. Ser-B, 244-255.

Terasvirta, Timo (1994) : Specification, estimation and evaluation of


smooth transition autoregressive models. J. Amer. Statist. As-
soc. 89 208-218.

Thrusby, J .G. and Schmidt, P. (1977): Some properties of tests for


specification error in a linear regression model. J. A mer. Statist.
Assoc. 72 635-641.

Tibilletti, Luisa (1994) : A nonlinear combinations of experts' fore-


casts: a Bayesian approach. J. Forecasting 13 21-27.

Tong, H. (1983): Threshold Models in Non-Linear Time Series Anal-


ysis. Springer-Verlag, New York.

Tong, H. (1990): Non-Linear Time Series - A Dynamical System


Approach, Oxford Statistical Science Series; 6, Oxford University
Press.
Tsay, Ruey, S. (1986) : Nonlinearity tests for time series. Biometrika
73 No.2 461-466.

You might also like