International Journal of Economic Sciences and Applied Research 3 (2): 21-39

A Time-Varying Parameter Vector Autoregression Model for

Forecasting Emerging Market Exchange Rates

Manish Kumar1

Abstract

In this study, a vector autoregression (VAR) model with time-varying parameters (TVP) to
predict the daily Indian rupee (INR)/US dollar (USD) exchange rates for the Indian economy
is developed. The method is based on characterization of the TVP as an optimal control
problem. The methodology is a blend of the flexible least squares and Kalman filter
techniques. The out-of-sample forecasting performance of the TVP-VAR model is evaluated
against the simple VAR and ARIMA models, by employing a cross-validation process and
metrics such as mean absolute error, root mean square error, and directional accuracy. Out-
of-sample results in terms of conventional forecast evaluation statistics and directional
accuracy show TVP-VAR model consistently outperforms the simple VAR and ARIMA models.

Keywords: Stock Prices, Exchange Rates, Bivariate Causality, Forecasting

JEL Classification: C22, C52, C53, F31, G10

1. Introduction

Various significant structural transformations between 1960 and early 1970s led to
the dramatic end of the Breton-Woods system of pegged exchange rates. Numerous
efforts to bring back the fixed exchange rate system proved futile and by March 1973, the
regime of floating currencies began. Collapse of Breton-Woods and rapid expansion of
global trading markets has altered the dynamics of foreign exchange market dramatically.
This market is considered the largest and most liquid of the financial markets, with an
estimated $1 trillion traded every day.
So, in such an environment where exchange rate fluctuates, policymakers strive to
understand the exchange rate movements and their implications on interest rates and
inflation. The interest rate is set on the basis of an overall assessment of the inflation
outlook. Moreover, exchange rate movements affect consumer price inflation through

1
IREVNA, A Division of CRISIL, Chennai: 600016, India. He is also a PhD Research Scholar,
Department of Management Studies, IIT Madras, Chennai: 600036, India.

21

Volume 3 issue 2.indd 21 9/12/2010 10:13:29 πμ

 Manish Kumar

various means. Therefore, fluctuation in the rates is considered a source of economic

disturbance and helps determine future inflation.
Multinational companies (MNCs) thrive on most of their sales coming from
overseas. And for this, they have to dabble with many tasks such as decisions on hedging,
short-term financing, short-term investment, capital budgeting, earning assessment, and
long-term financing. This makes MNCs vulnerable to exchange rate movements. Thus
understanding the exchange rate movements can help MNCs take better decisions. Large
volume and value of currencies transacted around the world have also led to speculation
in these markets, as huge benefits are netted from the differences in exchange rates.
Moreover, post Breton-Woods, many countries including India resorted to
managed floating exchange rate system. In this system, the central bank checks the free
movement of exchange rates and intervenes in the market, as a buyer or seller of
currencies to influence adverse market conditions or exchange rate movements. What is
more, it is needless to say that policymakers need to have a thorough knowledge of
currency movements.
As exchange rates are considered an important input in various decision-making
processes, numerous significant contributions (Frankel, 1976; Meese and Rogoff, 1983a,
1983b; Alexander and Thomas, 1987; Wolff, 1987; Baillie and McMahon, 1989; Mussa,
1990; Gallant et al.,1991; Meese and Rogoff, 1991; Frankel and Rose, 1995) have been
made theoretically and empirically to characterize and understand exchange rates. However,
results of most studies (Meese and Rogoff, 1983a; 1983b; Alexander and Thomas, 1987;
Somanath, 1986; Boothe and Glassman, 1987; Wolff, 1987; Wolff, 1988) suggest that
econometric forecasting models of exchange rates failed to outperform the random walk
model, in forecasting variations in exchange rates. These results also point out that the
exchange rate is an asset price and is determined efficiently in the foreign exchange market.
Results of few studies (Woo, 1985; Schinasi and Swami, 1989; Kuan and Liu,
1995; Brooks, 1997; Gencay, 1999; Kumar and Thenmozhi, 2004, 2005; Hongxing et al.,
2007; An-Pin Chen et al., 2008) contradict the earlier findings and show that their models
outperform the naïve random walk, for certain time periods and currencies. These studies
conclude that the presence of nonlinearity, volatility, chaos, nonstationarity, etc in the
exchange rates was not handled properly in previous empirical models. This resulted in
the random walk scoring over other models.
Most studies (Brooks, 1997; Panda and Narsimhan, 2003; Corte et al., 2007; Rime
et al., 2007, An-Pin Chen et al., 2008 etc.) on exchange rate forecasting have either used
macroeconomic variables such as inflation rates, money supplies, interest rates, trade
balance, and crude oil price; or technical indicators such as past returns of exchange rates,
short- and long-term moving averages etc as an input variables to the forecasting models.
Moreover, studies (Lisi and Schiavo, 1999; Qi and Wu, 2003; Kumar and
Thenmozhi, 2003, 2004 and 2005; Chen and Leung, 2004; Cao et al., 2005; Hongxing et
al., 2007; An-Pin Chen et al., 2008; Carriero et al., 2009) in the domain of exchange rate
forecasting have used a wide range of linear and nonlinear models such as autoregressive
models, ARIMA, linear regression, VAR, ARCH/GARCH, artificial neural networks

22

Volume 3 issue 2.indd 22 9/12/2010 10:13:29 πμ

 A Time-Varying Parameter Vector Autoregression Model for
Forecasting Emerging Market Exchange Rates

(ANN), and support vector machines (SVM). Earlier studies also used a wide range of
statistical metrics such as root mean square error (RMSE), mean absolute error (MAE),
mean absolute percentage error (MAPE), and Theil’s coefficient—to evaluate the
performance of forecast.
During the last decade, the use of various nonlinear models such as ANN and SVM
in forecasting various financial time series has considerably been increased. However,
having a forecasting model which is consistent with economic theory and also forecast
well is very appealing. So, in this study, we use the VAR time series approach to forecast
the daily exchange rates of INR vs. USD. We select INR/USD rates because India is the
world’s sixteenth largest foreign market, in terms of daily turnover ($34 billion in 2007).
India’s contribution to global foreign market turnover has grown to 0.9% in 2008, a three-
fold jump from just 0.3% in 2004. It also recorded the second-highest growth in the daily
average foreign exchange market turnover after China. In 2006–07, India's annual gross
foreign exchange market turnover grew to $6.5 trillion from $1.4 trillion, six years earlier.
In August and October 2008, the Security and Exchange Board of India (SEBI) permitted
National Stock Exchange (NSE), Bombay Stock Exchange (BSE) and Multi Commodity
Exchange of India (MCX) to set up a currency derivative segment as well, where only
currency futures (INR/USD) contracts are traded. Total turnover in this segment
increased to $19.52 billion in March 2009 from $3.38 billion in October 2008. Open
interest in the segment also grew 96% to 4, 51,819 contracts in March 2009 from 2,
30,257 contracts in October 2008.
Moreover, India has attracted unprecedented foreign investment (touching a
phenomenal $10 billion) and is poised to become a major hub in the Asian economy.
With the growing interest and research on emerging markets, India remains in the focus
due to its rapid growth and potential investor opportunities. Moreover, volatility in
emerging markets seems to be higher than in the developed markets—often making
prediction difficult. This background makes the study more worthwhile: whether the
dynamic linkages between INR/USD and stock market indices in India can be deployed
to build a superior and accurate forecasting model.
This specific study improves upon the existing studies in several ways. To develop
a successful model, it is necessary to consider the structural change in data. For this,
‘regime-switching’ and ‘time varying parameter’ (TVP) are two popularly used
techniques; The TVP approach is used in the study given. This approach is beneficial in
the sense that it makes use of all available data points and retains the long-term
relationship inherent in the old data points (Hongxing et al., 2007). The study
characterizes the TVP of VAR model based on the optimal control theory as proposed by
Rao (2000). This methodology is a blend of flexible least squares and Kalman filter
technique. Rao (2000) estimated TVP for linear regression models without an intercept;
however, this study extends Rao’s approach. Our method updates the coefficient of the
VAR model and its covariance matrices in each time unit. Thus, the model can also be
called a Bayesian VAR model. There is also a comparison between the results of the
time-varying VAR model given with those of a linear VAR and ARIMA models.

23

Volume 3 issue 2.indd 23 9/12/2010 10:13:29 πμ

 Manish Kumar

Earlier literature used macroeconomic variables, technical indicators, etc to

develop forecasting models; but not stock prices data. Several recent studies (Vygodina,
2006; Pan et al., 2007; Ai-Yee Ooi, 2009; Aydemir and Demirhan, 2009 etc.) have
reported causality from stock prices to exchange rates. Their results support the goods
market or portfolio approach. Many firms in India that are on the stock index, NSE’s S&P
CNX Nifty (thereafter, CNX) have American Depository Receipts (ADRs) or General
Depository Receipts (GDRs) that are traded on the NYSE, NASDAQ or on non-
American exchanges. Over the years, INR has gradually moved towards full
convertibility. The two-way fungibility of ADRs/GDRs allowed by India’s central bank,
Reserve Bank of India (RBI) has also possibly enhanced the linkages between the stock
and foreign exchange markets in India. So, the dynamic relations between a stock index
and exchange rates are examined, using linear granger causality tests. In addition, there is
also usage of unit root and cointegration tests to analyze the long-run equilibrium
between the two variables. The study exploits the dynamic linkage between stock price
and exchange rate, and uses the granger causality test results to select important inputs for
forecasting INR/USD.
Most previous studies arbitrarily split the available data into training (in-sample)
and test (out-of-sample) set for model construction and validation, respectively.
Performance of the forecasting model is likely to vary with the variation in time. So,
consistency and robustness of the several competing models on various test data should
also be evaluated. So in our study, we employ a cross-validation methodology (in-sample
and out-of-sample data sets) to investigate the performance of various forecasting models.
Besides, earlier studies measured the degree of accuracy and acceptability of
forecasting models, by the estimate’s deviations from the observed values; they have not
considered turning-point forecast capability using sign and direction test. In this study,
there is a rigorous evaluation of the performance of forecasting models using MAE,
RMSE, and directional accuracy. Directional accuracy measures the degree to which the
forecast correctly predicts the direction of change in the actual INR/USD exchange rate
returns.
It is expected that the outcome of the study will help enhance the existing
literature, and offer some meaningful insights for policy makers, MNCs, traders, and
individuals. The paper is set out as follows. In Section 2, daily exchange rates and the
concept of unit root tests, cointegration tests, linear granger causality, VAR, TVP-VAR
framework and ARIMA models are described. In Section 3, the empirical results are
presented. Finally, Section 4 concludes with some discussion on future research.

2. Data and Methodology

We examined the daily closing price of CNX and INR/USD, obtained from NSE
and RBI Web sites, respectively. The time period used is from January 4, 1999 to August
31, 2009. The original series was transformed into a continuously compounded rate of
return, computed as the first difference of the natural logarithm of CNX and INR/USD.

24

Volume 3 issue 2.indd 24 9/12/2010 10:13:29 πμ

 A Time-Varying Parameter Vector Autoregression Model for
Forecasting Emerging Market Exchange Rates

2.1 Unit Root Tests

To test the unit roots (stationarity) in CNX and INR/USD exchange rates,
Augmented Dickey and Fuller (ADF) and Kwiatkowski, Philips, Schmidt and Shin
(KPSS) tests are used. An ADF is a test for a unit root in a time series sample. KPSS tests
are used for testing a null hypothesis that an observable time series is stationary around a
deterministic trend. KPSS type tests are intended to complement unit root tests, such as
the Dickey–Fuller tests. Table 1, contains the results.

2.2 Engle and Granger Cointegration Test

To investigate the existence of long-run relationship between two variables (CNX

and INR/USD), the methodology suggested by Engle and Granger (1987) is employed.
We prefer to use this method over Johansen Cointegration test, as Engle and Granger test
is simple and helps in having at most one cointegrating vector (because we are examining
two variables).
In the first step, we examine the order of integration of each variable. Cointegration
between CNX and INR/USD requires that both series should be non-stationary of the
same order of integration. In the second step, we run the following cointegration
regression:

ln St J 0  J 1 ln ERt  H t (1)

Where ln St and ln ERt are logarithms of CNX and INR/USD, respectively.

The third step is to obtain the error terms and run the ADF and KPSS tests on them
in order to examine whether error terms have unit root and are stationary or not. If the
error series is stationary, then the null hypothesis of no-cointegrating vectors is rejected.
Table 2, lists the results of Engle and Granger cointegration test.

2.3 Granger Causality Test

In literature, various tests of Granger causality have been proposed and used. These
tests are mainly based on the context of VAR models. So we employ the VAR framework
to examine the presence of linear Granger causality between CNX and INR/USD. Let Vt
denote the vector of endogenous variables and p, the number of lags. Then the VAR
model can be represented as:
p
Vt ¦< V
i 1
i t s  Ht (2)

Where Vt=(V1t,….,Vpt), the p x 1 vector of endogenous variables, < i the p x p coefficient

matrices, and H t is a zero-mean vector of white-noise processes.

25

Volume 3 issue 2.indd 25 9/12/2010 10:13:29 πμ

 Manish Kumar

A VAR model including CNX returns denoted as X and INR/USD returns as Y can
be expressed as:

Xt D ( p) X t  E ( p )Yt  H xy ,t t 1, 2,...., N (3)

Yt F ( p) X t  G ( p)Yt  H yx ,t
t 1, 2,...., N (4)

where D (p), E (p), F (p) and G (p) are all polynomial in the lag operator with all roots
outside the unit circle
The error terms are identically and independently distributed (i.i.d) with zero-mean
and constant variance.
If cointegration exists between CNX and INR/USD series, then the Granger
representation theorem states that there is a corresponding error correction model. This
model for CNX and INR/USD series can be represented as:

Xt D ( p) X t  E ( p )Yt  I Z t 1  H xy ,t t 1, 2,...., N (5)

Yt F ( p) X t  G ( p )Yt  M Z t 1  H yx ,t t 1, 2,...., N (6)

Where Z ln St  J 0  J 1 ln ERt are the residuals from the cointegration regression of the
log levels and ¨ln St and ¨ln ERt is the first log difference of CNX and INR/USD,
respectively (or simple exchange rate and CNX returns). Optimal lag length is selected
based on the Akaike Information Criteria (AIC).
Within the context of this VAR/VECM model, linear Granger causality restrictions
can be defined as follows. If the null hypothesis that E ’s jointly equal zero is rejected, it
is argued that INR/USD returns (Y) Granger-cause CNX returns (X). Similarly, if the null
hypothesis that F ’s jointly equal zero is rejected, CNX returns (X) Granger-cause
exchange rate returns (Y). If both the null hypotheses are rejected, a bi-directional
Granger causality, or a feedback relation, is said to exist between variables. Different test
statistics have been proposed to test for linear Granger causality restrictions. To test for
strict Granger causality for pairs of (X,Y) in this linear framework, Chi-Square statistics
is used to determine whether lagged value of one time series has significant linear
predictive power for the current value of another series. Table 3 lists the results.

2.4 ARIMA Model

Popularly known as Box-Jenkins (BJ) methodology, but technically known as

Autoregressive Integrated Moving Average (ARIMA) model, it is of the following form:

26

Volume 3 issue 2.indd 26 9/12/2010 10:13:29 πμ

 A Time-Varying Parameter Vector Autoregression Model for
Forecasting Emerging Market Exchange Rates

p q
Yt a0  ¦ D iYt i  ¦ E i et i  [t (7)
i 1 i 0

Where Yt is INR/USD returns, [t is an uncorrelated random error term with zero mean
and constant variance, and a0 is a constant term.
The correlogram, simply a plot of Autocorrelation Functions (ACFs) and Partial
Autocorrelation Functions (PACFs) against the lag length, is used in identifying the
significant ACFs and PACFs. Lags of ACFs and PACFs, whose probability is less than
5%, are significant and are identified. Possible models are developed from these plots for
CNX returns series. The best model for forecasting is picked by considering information
criteria such as AIC and Schwarz Bayesian Information Criterion (SBIC).

2.5 The VAR Methodology

In this study, a VAR model is employed to forecast the INR/USD returns. Given Vt
the vector of variables (CNX (Xt) and INR/USD (Yt) returns), the classical first-order bi-
variate VAR model can be represented as:

Vt <Vt 1  H t t 1,....N (8)

where Vt (Yt , X t )' , < is a 2 x 2 coefficient matrix and H t is a zero-mean vector of

white-noise processes, with positive definite contemporaneous covariance matrix Ȉ and
zero covariance matrices at all other lags.
To estimate the parameters, our study uses the least squares method. Optimal lag
length is selected using information criteria (AIC and SBIC).
Optimal one-period ahead forecast of Vt using the VAR model can be obtained
using the below equation:

Vt 1,t <Vt (9)

2.6 VAR Model with TVP

Traditional time series model with time-invariant coefficients cannot be used to

examine the relationship among economic variables, as they alter with changes in
economic policies. So time variation of these parameters should go hand in hand with
policy changes.
Thus, several estimation methods (switching regression model (Quandt, 1958;
Goldfled and Quandt, 1973), pure random coefficient model (Rao, 1965), adaptive and
varying parameters model (Cooley and Prescott, 1973a; 1973b), Kalman filter model
(Athan, 1974), flexible least squares (Kalaba and Tesfatsion, 1988), recursive model

27

Volume 3 issue 2.indd 27 9/12/2010 10:13:30 πμ

 Manish Kumar

(Rao, 1991), optimal control model (Rao and Nachane, 1988) have been proposed in the
literature to deal with the problem is estimating TVP models.
Rao (2000), proposed a methodology to estimate the TVP of a linear regression
model without an intercept term. The methodology is based on the characterization of the
TVP problem as a near-neighborhood search problem, with an explicit allowance for
welfare loss considerations. This leads to an updating algorithm, capable of predicting the
optimal values of TVP as well as their covariance matrices. This methodology is a blend
of the flexible least squares and Kalman filter techniques.
In the present study, there is an attempt to extend Rao (2000), to estimate TVP of
VAR model. As the method updates the VAR coefficients and their covariance matrices
for every time period, it could be considered as a Bayesian method. As discussed in the
earlier section, the classical linear first-order bivariate VAR model is represented as:

Vt <Vt 1  H t , t 1,....N (10)

In the optimal control theory, the TVP of VAR model using near-neighborhood
search method can be represented as:

Vt * < *t Vt 1 t 1,....N (11)

A control problem comprises a cost function, which indeed is a function of state

and control variables. The linear quadratic cost functional, which has to be minimized for
the TVP VAR model, can be stated as:

1 N 1 N
W ¦ t t t 2¦
2t1
Q (V *
 V ) 2

t 1
(< *t  < *t 1 )' Rt (< *t  < *t 1 ) (12)

Where Vt , Vt * are states, < *t is control, and t is the time. In the finite-horizon case, the
matrices Qt and Rt are positive semi-definite and positive definite, respectively. The
solution of the optimal control problem may not be unique. Most often, the solutions of
such problems is locally minimizing. However, the main advantage of the cost function is
that the constraints < *t lie within the neighborhood of < *t 1 (Rao, 2000).
Equation 12, can be solved using the method of Lagrange multipliers. It is
converted into an equivalent cost function, using certain unspecified parameters known as
Lagrange multipliers (Ȝt). The new cost function called as Lagrange function can be
defined as:

1 N 1 N N
J ¦
2t1
Qt (Vt *  Vt ) 2  ¦ (< *t  < *t 1 )' Rt (< *t  < *t 1 )  ¦ Ot (Vt *  < *t Vt 1 )
2t1 t 1
(13)

28

Volume 3 issue 2.indd 28 9/12/2010 10:13:30 πμ

 A Time-Varying Parameter Vector Autoregression Model for
Forecasting Emerging Market Exchange Rates

To obtain the solution of the new cost function, we take the partial derivative of
J with respect to three variables Vt , < *t and Ot ; and set the partial derivative to zero. We
get:
GJ
Qt (Vt*  Vt )  Ot 0 (t 1,....., N ) (14)
G Vt *
GJ
Rt (< *t  < *t 1 )  Vt 1Ot 0 (t 1,....., N ) (15)
G< *t
GJ
(Vt*  < *t Vt 1 ) 0 (t 1,....., N ) (16)
GOt

Solving Equation 14 for Ot gives:

Ot QtVt *  QtVt (17)

Solving for Vt * using Equation 16, we get:

Vt* < *t Vt 1 (18)

Substituting the value of Vt* in Equation 17 yields:

Ot Qt < *t Vt 1  QtVt (19)

Substituting Equation 19, in Equation 15, results in:

< *t ( Rt  Vt 1QtVt 1 ) 1 Rt < *t 1  ( Rt  Vt 1QtVt 1 ) 1Vt 1QtVt (20)

At this point, we have all the variables in terms of < *t

In Equation 20, if the variable Qt 0 , then it becomes:

< *t < *t 1 (21)

Equation 21, suggests that the coefficient of the VAR model is time-invariant.
However, in Equation 15, if we set Rt 0 , then it becomes:

Vt 1Ot 0 (22)

From equation 19 and 22, we get:

Vt 1QtVt 1< *t Vt 1QtVt (23)

29

Volume 3 issue 2.indd 29 9/12/2010 10:13:30 πμ

 Manish Kumar

Equation 23, results in Vt 1< *t Vt . This suggests that the dependent variables of
VAR model can be tracked. Moreover, the adaptive nature of control system is evident, if
Equation 20 is transformed. To do this, let us assume that:

Kt ( Rt  Vt 1QtVt 1 ) 1Vt 1Q (24)

Using Equation 24, Equation 20, can be rewritten as:

( Rt  Vt 1QtVt 1 ) 1[( Rt  Vt 1QtVt 1 )  Vt 1QtVt 1 ] I  K tVt 1 (25)

Substituting Equations 24 and 25, in Equation 20, we get:

< *t < *t 1  K t (Vt  < *t 1Vt 1 ) (26)

Where Kt is the correction factor

Equation 26, can be used to compute the < *t from its previous estimate.
Subsequently, < *t can be used to compute the predicted state of Vt* .
The above three methodologies have been used to forecast the daily returns of
INR/USD.

2.7 Estimation and Prediction

Examining the robustness of the forecast models is an interesting topic as well as a

meaningful trial. This helps access the forecast performance vis-à-vis different sample
data sets. To investigate the performance of the competiting forecasting models, four
validation sets are used. In the first set, daily data of CNX and INR/USD from January 4,
1999–December 31, 2006 is used. The data is divided into an estimation period (in-
sample data) from January 4, 1999–December 31, 2005, and a forecast period (out-of-
sample data), from January 1, 2006–December 31, 2006. In the second validation set,
daily data is considered from January 4, 1999–December 31, 2007. There are estimations
conducted over period from January 4, 1999–December 31, 2006 and data from January
1, 2007–December 31, 2007 is reserved for the forecasting exercise. The third validation
set covers a daily period from January 4, 1999 – December 31, 2008. The data is divided
into an estimation sample from January 4, 1999–December 31, 2007, and a forecast
sample from January 1, 2008– December 31, 2008. In the last validation set, there is daily
data from January 4, 1999–August 31, 2009. The data is divided into two periods:
January 4, 1999–December 31, 2008, for model estimation and is classified as in-sample
and a period from January 1, 2009–August 31, 2009 is reserved for out-of-sample
forecasting and evaluation.

30

Volume 3 issue 2.indd 30 9/12/2010 10:13:30 πμ

 A Time-Varying Parameter Vector Autoregression Model for
Forecasting Emerging Market Exchange Rates

3. Results

3.1 Unit Root Test

Results of ADF and KPSS tests for the two time series: CNX and INR/USD returns
and for the entire time period are shown in Table 1.

Table 1: Unit Root Test

ADF Test KPSS Test

Variable t-statistics t-statistics
ln St (Log level) -0.6553 5.3562
¨ lnSt (First Diff) -36.7465 0.0890
ln ERt (Log Level) -1.3609 0.7149
¨ lnERt (First Diff) -52.5394 0.1765
Source: Author estimation
Note: The critical values for ADF and KPSS test at 1 % level are -3.4327 and 0.739 respectively.

These results indicate that the log level of CNX and exchange rates series have a
unit root. However, ADF and KPSS tests test on the first log order difference for the two
series: ¨ln St and ¨ln ERt; and this confirm the stationarity of the two series.

3.2 Engle and Granger Cointegration Test

After testing for the unit root in the two series, the two-step Engle and Granger
cointegration test was applied to examine whether the logarithms of INR/USD exchange
rate and CNX are cointegrated. Table 2, reports the results of the cointegration regression.

Table 2: Engle and Granger Cointegration Test

Cointegrating Regression
Coefficient Coefficient Value t-statistic Probability
Jo 25.3729 40.3541 0.0000
J1 -4.6699 -28.2934 0.0000
Unit Root Test of Cointegrating Errors
ADF Test KPSS Test
t-statistics Critical Value (1%) t-statistics Critical Value (1%)
-0.5415 -3.4327 5.4933 0.739
Source: Author estimation

31

Volume 3 issue 2.indd 31 9/12/2010 10:13:30 πμ

 Manish Kumar

To find out whether the variables are actually cointegrated, the cointegration error
terms are tested for stationarity. ADF and KPSS test results clearly indicate that the error
terms are nonstationary. So, the results suggest that there is no cointegrating vector and
eventually no long-run relationship between exchange rate and stock indices for India.
Hence, an error correction term need not be included in the Granger causality test
equations. Findings of Engle and Granger Cointegration tests are consistent with the
findings of previous studies for developed markets such as US, UK, and Japan as well as
for Asian markets like India, Malaysia, and Pakistan.

3.4 Linear Granger Causality Test

The dynamic (causal) relationship between CNX and INR/USD returns is investigated
using bi-variate VAR framework, without the error correction term. The appropriate lag
length for VAR models is selected using SBIC. Granger causality test results are in Table 3.
Panel A of Table 3, reports the results of linear Granger causal test, while panel B
reports the Granger causality results between volatility-filtered CNX and INR/USD returns.

Table 3: Linear Granger Causality Test

Null Hypothesis Chi-Sq-Statistics P-Value

Nifty Returns does not Granger cause 8.2422 0.0162**
INR/USD
INR/USD does not Granger cause 9.6352 0.0081*
Nifty Returns
Source: Author estimation
* Represent the relationship being significant at 1 %
** Represent the relationship being significant at 5 %
The optimal lag length is 2 which are selected based on the SBIC criteria.

The results reported in Panel A, indicate that both the null hypotheses ‘Nifty
Returns does not Granger-cause INR/USD returns’ and ‘INR/USD returns does not
Granger-cause Nifty returns’ are rejected. Chi-Square statistics are significant and
provide strong evidence for the argument that there is a Bi-directional linear Granger
causality between CNX and INR/USD returns.
In general, the results suggest that exchange rates do help explain changes in the
stock index and vice versa. So, the results of our study do not support the ‘Efficient
Market Hypothesis’ for the Indian market. Moreover, the findings strongly support the
portfolio approach on the relationship between exchange rates and stock prices. Thus, we
could use stock price as an initial attribute to forecast exchange rates and vice versa.

32

Volume 3 issue 2.indd 32 9/12/2010 10:13:30 πμ

 A Time-Varying Parameter Vector Autoregression Model for
Forecasting Emerging Market Exchange Rates

3.5 ARIMA Model

Plots of ACFs and PACFs are used to identify the significant lag length. Various
orders of ARIMA models are developed using ACF and PACF plots. Information criteria
(AIC and SIC) help identify the best forecasting model (results available upon request).
After considering all possible models and looking at AIC and value of each model, it was
decided that ARIMA (2,1,1) is best model for forecasting daily returns of INR/USD
series for the first validation set (January 1, 1999–December 31, 2006). Moreover, for the
subsequent validation data sets, ARIMA (1,1,2) is the best. Further diagnostic tests are
performed to check the model’s adequacy.
We use a popular diagnostic test known as Breusch-Godfrey LM test to examine
the presence of serial correlation in the residuals of the developed ARIMA model. This
test helps examine the relationship between residuals and several of its lagged values at
the same time. The null hypothesis is that ‘there is no serial correlation’. If the
predictability value is greater than 5%, then we accept the hypothesis (at 95% confidence
levels); so there is no serial correlation in the series. The LM test for serial correlation of
residuals suggests that the ARIMA (2,1,1) and ARIMA(1,1,2) models capture the entire
serial correlation; and the residuals do not exhibit any serial correlation (results available
upon request). It suggests that the residuals, estimated by the two ARIMA models, are
purely random. So, another ARIMA model may not be searched (Gujrati, 1995).

3.6 VAR Model

This model generally uses equal lag length for all its variables. One of its
drawbacks is that many parameters need to be estimated, some of which may be
insignificant. This problem of over-parameterization, resulting in multicollinearity and a
loss of degrees of freedom, leads to inefficient estimates and possibly large out-of-sample
forecasting errors (Litterman, 1986; Spencer, 1993). One solution, often adopted, is
simply to exclude the insignificant lags based on statistical tests. Another approach is to
use a near VAR, which specifies an unequal number of lags for the different equations.
In our study, while examining the causality test in the VAR framework, we
selected two lags of CNX and INR/USD, based on SBIC criteria. However, when the
parameters in VAR model of Equation 3 are estimated, it was found that the second lag of
CNX and INR/USD seems to be insignificant. So we excluded these lags from the VAR
model and re-estimated the model, using ordinary least squares criteria. So the forecast is
done using a bi-variate first-order VAR model.

3.7 TVP-VAR Model

This model has been used to forecast the daily returns of INR/USD. First, an initial
coefficient of TVP-VAR model is chosen that minimizes the welfare loss function.
Estimates of the simple VAR model based on the in-sample data are used as an initial

33

Volume 3 issue 2.indd 33 9/12/2010 10:13:30 πμ

 Manish Kumar

coefficient of the model for every cross-validation data set. The near-neighborhood
approach using grid search is then employed on the initial coefficient to identify the
coefficient values that further minimizes the loss function. This new coefficient as
identified by the grid search is used to estimate the consecutive sequence of the
coefficient and the corresponding predictions of the INR/USD returns, for each validation
set of the model.
Forecasting performance of various models and for the four out-of-sample periods
are summarized in Table 4.

Table 4: Prediction Accuracy

First Validation Test Set (January 1, 2006 –December 31, 2006)

Model Performance Metrics
MAE RMSE Directional Accuracy
TVP-VAR 0.002065 0.002982 54.48%
VAR 0.002074 0.002986 53.13%
ARIMA 0.002081 0.002987 52.30%
Second Validation Test Set (January 1, 2007–December 31, 2007)
Model Performance Metrics
MAE RMSE Directional Accuracy
TVP-VAR 0.002550 0.003856 59.91%
VAR 0.002552 0.003861 56.61%
ARIMA 0.002555 0.003864 54.95%
Third Validation Test Set (January 1, 2008–December 31, 2008)
Model Performance Metrics
MAE RMSE Directional Accuracy
TVP-VAR 0.004687 0.006841 56.06%
VAR 0.004705 0.006885 53.13%
ARIMA 0.004759 0.006897 49.70%
Fourth Validation Test Set (January 1, 2009–August 31, 2009)
Model Performance Metrics
MAE RMSE Directional Accuracy
TVP-VAR 0.004725 0.006466 56.96%
VAR 0.004772 0.006493 55.69%
ARIMA 0.004872 0.006673 48.10%
Source: Author estimation

The results display the out-of-sample results of the various forecasting models. For
the INR/USD returns, we find that the TVP-VAR model yields better forecast (smaller
RMSE and MAE) than the simple VAR and ARIMA models, for all the validation sets.

34

Volume 3 issue 2.indd 34 9/12/2010 10:13:30 πμ

 A Time-Varying Parameter Vector Autoregression Model for
Forecasting Emerging Market Exchange Rates

Moreover, between VAR and ARIMA model, the VAR forecasts on the four out-of-
sample data have smaller RMSE and MAE.
The TVP-VAR model also exhibits good market-timing ability as indicated by the
results of directional accuracy. The directional accuracy of the model is 54–60% over the
four test samples. This means the forecasts are comparatively better than the chances in
tossing a coin. Compared to the ARIMA models, the simple VAR forecast have higher
directional accuracy values.
As for the forecasting stability, two observations can be made from Table 4. First,
the time series models (TVP-VAR) are robust across the cross-validation test and the
results seem to be more stable. Second, no matter what method is used, there are no
consistent patterns in MAE and RMSE across all out-of-sample periods. There is a
difference in the values of various performance measures such as RMSE and MAE of
TVP-VAR, VAR and ARIMA models for all out-of-sample periods. This result is
expected since the structure of the exchange rate time series varies from one time period to
the other. If in-sample and out-of-sample date generally increase or decrease or vice versa,
then it is clear that no method can predict well particularly in the short run—leading to large
variations in prediction. So, it may be concluded that the predictive accuracy of all models
changes across time, for different forecasting horizons.
Overall, results suggest that the TVP-VAR model contains added information for
INR/USD and strongly outperforms the other two models. They are consistent with our
expectations that allowing TVP using optimal control theory enhances the model’s
forecasting performance.

4. Conclusion

A Bayesian VAR model was developed based on the optimal control theory, which
updates the coefficient and their covariance matrices in each time period. Our TVP-VAR
model helps forecast the daily returns of INR/USD. The results of the TVP-VAR model
are compared with the simple VAR and a linear ARIMA model. A cross-validation
scheme is employed to examine the robustness of the three models with regard to
sampling variation in time series. Out-of-sample performances of the three models were
evaluated along performance metrics like MAE, RMSE and Directional Accuracy.
Results from the study indicate that the TVP-VAR model achieves high rate of accuracy,
in terms of MAE, RMSE and Directional Accuracy for the four validation sets. The
results in general supports the study of Carriero et al. (2009), Sarantis (2006), Kumar and
Thenmozhi (2003, 2004 and 2005), and Chen and Leung (2004) etc. The forecast gains
are due primarily to the time-variation of coefficients. Moreover, the results also suggest
that, informational content of indicators like stock index can be exploited to improve the
exchange rate forecasts. Thus, the results reject the efficient market hypothesis and lend
support to the technical analysis.
The findings of the study would be of great interest to traders, MNC’s, regulators
and others. The better forecasting or understanding of the movements of exchange rate

35

Volume 3 issue 2.indd 35 9/12/2010 10:13:31 πμ

 Manish Kumar

for the developing economy of Asia would help traders to devise more effective business
and trading strategies and a proper decision on asset allocation. Moreover, based on the
forecast, they can also take precautionary measure to reduce potential currency risk.
Corporate and MNC’s can effectively use such models for their foreign exchange
risk management plan/policy/programme. Such models would help them to reduce the
volatility in profits after tax, cash flows, and to reduce the cost of capital and thus
increase the value of the firm on one side of the pole and to reduce the risks faced by the
management on the other side of the pole.
The TVP-VAR model would also help policy makers in India to intervene
successfully and at the right time in the market in order to prevent overshooting and
decisively break the momentum in currency movement. Thus, the policy makers can
conduct a suitable monetary policy which will in turn achieve its desired objectives of
price stability and higher economic activity. Moreover, the dynamic bi-directional causal
relationship between exchange rates and stock index also suggests that, the SEBI and the
central bank i.e. RBI in India should be very careful in conducting exchange rate policies
or capital market polices as it may impact on the development of the financial markets.
The study can be extended by taking into account the set of potential
macroeconomic input variables such as interest rates, consumer price index and industrial
production, as well as technical indicators. Moreover, various trading strategies can be
used to examine if trading profits can be obtained from the forecasting model used in the
study. The other logical extension could be to combine the TVP-VAR model with some
nonlinear models used in the financial time series literature.

Acknowledgement

The author would like to thank the anonymous referees for going through the
manuscript patiently and critically and providing their constructive suggestions in
improving it. The views expressed here are those of the author, and do not necessarily
represent those of their employers.

References

Ai-Yee Ooi, Syed Azizi Wafa Syed Khalid Wafa, Nelson Lajuni, Mohd Fahmi Ghazali,
2009, ‘Causality between Exchange Rates and Stock Prices: Evidence from
Malaysia and Thailand’, International Journal of Business and Management, 4, 3,
pp. 86-98.
Alexander, D., Thomas, L.R., 1987, ‘Monetary/Asset models of Exchange Rate
Determination: How well have they Performed in the 1980's?’, International
Journal of Forecasting, 3, pp. 53-64.

36

Volume 3 issue 2.indd 36 9/12/2010 10:13:31 πμ

 A Time-Varying Parameter Vector Autoregression Model for
Forecasting Emerging Market Exchange Rates

An-Pin Chen, Yu-Chia Hsu, Ko-Fei Hu. 2008, ‘A Hybrid Forecasting Model for Foreign
Exchange Rate Based on a Multi-neural Network’, Fourth International
Conference on Natural Computation, 5, pp. 293-298
Athans, M., 1974, ‘The Importance of Kalman Filtering Methods for Economic Systems’,
Annals of Economic and Social Measurement, 3, pp. 49–64.
Aydemir, O. Demirhan, ve E., 2009, ‘The Relationship between Stock Prices and
Exchange Rates: Evidence from Turkey’, International Research Journal of
Finance and Economics, 23, pp. 207-215.
Baillie, N., McMahon, R. P., 1989, The Foreign Exchange Market: Theory and
Econometric Evidence, Cambridge University Press, Cambridge.
Boothe, P. and Glassman, D., 1987, ‘The Statistical Distribution of Exchange Rates’,
Journal of International Economics, 22, pp. 297–319.
Brooks, C., 1997, ‘Linear and Nonlinear (Non-) Forecastability of Daily Sterling
Exchange Rates’, Journal of Forecasting, 16, pp. 125-145.
Cao, D. Z., Pang, S. L. and Y. H. Bai, 2005, ‘Forecasting Exchange Rate Using Support
Vector Machines’, Proceedings of the 4th international conference on machine
learning and cybernetics, Guangzhou, China.
Carriero, A., Kapetanios, G. and Marcellino, M., 2009, ‘Forecasting Exchange Rates with
a Large Bayesian VAR’, International Journal of Forecasting, 25, 2, pp. 400-417.
Chen, A.S., Leung, M.T., 2004, ‘Regression Neural Network for Error Correction in
Foreign Exchange Forecasting and Trading’, Computers and Operations, 31, 7, pp.
1049-1068.
Cooley, T.F. and Prescott, E.C., 1973, ‘An Adaptive Regression Model’, International
Economic Review, 14, pp. 364–371.
Corte, D. P., Sarno, L. and Tsiakas, I., 2007, ‘An Economic Evaluation of Empirical
Exchange Rate Models: Robust Evidence of Predictability and Volatility Timing’,
Working paper, University of Warwick.
Engle, R.F. and Granger, C.W.J., 1987, ‘Cointegration and Error-Correction:
Representation, Estimation, and Testing’, Econometrica, 55, pp. 251-276
Frankel, J.A. and Rose, A.K., 1995, ‘Empirical Research on Nominal Exchange Rates’,
Grossman, G. M. and Roggoff K. (ed.), Handbook of International Economics,
Elsevier, 1, 3, chapter 33, pp. 1689-1729.
Frankel, J.A., 1976, A Monetary Approach to the Exchange Rate: Doctrinal Aspects and
Empirical Evidence, Scandinavian Journal of Economics, 78, pp. 169-91.
Gallant, A., Hsieh, D. and Tauchen, G., 1991, ‘On Fitting a Recalcitrant Series: The
Pound/Dollar Exchange Rate, 1974-1983’, Barnett, W.A., Powell J. and Tauchen
G. (eds.), Nonparametric and Semiparametric Methods in Econometrics and
Statistics, Cambridge University Press, Cambridge, pp. 199-204.
Gençay, R., 1999, ‘Linear, Non-linear and Essential Foreign Exchange Rate Prediction
with Simple Technical Trading Rules’, Journal of International Economics, 47, pp.
91-107.

37

Volume 3 issue 2.indd 37 9/12/2010 10:13:31 πμ

 Manish Kumar

Goldfeld, S.M. and Quandt, R.E., 1973, ‘The Estimation of Structural Shifts by Switching
Regressions’, Annals of Economic and Social Measurement, 2, pp. 475–485.
Hongxing L.I, Fang, Z. and Zhao, D., 2007, ‘GBP/USD Currency Exchange Rate Time
Series Forecasting Using Regularized Least-Squares Regression Method’,
Proceedings of the World Congress on Engineering, Vol II, London U.K.
Kalaba, R. and Tesfatsion, L., 1980, ‘A Least-squares Model Specification Test for a
Class of Dynamic Nonlinear Economic Models with Systematically Varying
Parameters’, Journal of Optimization Theory and Applications, 32, pp. 538–567.
Kuan, C.M. and Liu, T., 1995, ‘Forecasting Exchange Rates using Feedforward and
Recurrent Neural Networks’, Journal of Applied Econometrics, 10, 4, pp. 347-364.
Kumar, M. and Thenmozhi, M., 2005, ‘Static and Dynamic Neural Network in
Forecasting Exchange Rate Returns’, Proceedings of the Conference on Research
in Finance and Accounting, Indian Institute of Management, Lucknow, India.
Kumar, M. and Thenmozhi, M., 2004, ‘Forecasting Daily Returns of Exchange Rates
Using Artificial Neural Network and ARIMA Model’, ICFAI Journal of Applied
Finance, 10, 11, pp. 16-36.
Kumar, M., Thenmozhi, M. and Rajendran, C., 2003, ‘Neural Network-Based Forecasting
of Daily Returns of Exchange Rate’, Proceedings of the Seventh Annual Capital
Market Conference, IICM-Mumbai, India.
Meese, R., Rogoff, K., 1983a, ‘Exchange rate Model of the Seventies: Do They Fit Out of
the Sample?’, Journal of International Economics, 14, pp. 3-24.
Meese, R., Rogoff, K., 1983b, ‘The Out of sample Failure of Empirical Exchange Rate
Model: Sampling Error of Misspecification?’ Freintel J., Ed., Exchange Rate and
International Macroeconomics, Chicago University Press.
Mussa, M., 1990, Exchange Rates in Theory and Reality, Princeton Studies in
International Economics, No. 179, International Economic Section, Department of
Economics, Princeton University, Princeton, N.J.
Pan, M.S., Fok, R.C.W. and Liu, Y.A., 2001, ‘Dynamic Linkages Exchange Rates and
Stock Prices: Evidence from Pacific Rim Countries’, Working Paper at College of
Business Shippensburg University mimeo.
Panda, C., Narasimhan, V., 2003, ‘Forecasting Daily Foreign Exchange Rate in India
with Artificial Neural Network’, The Singapore Economic Review, 48, 2, pp. 181-
199.
Qi, Min, Yangru, Wu., 2003, ‘Nonlinear Prediction of Exchange Rates with Monetary
Fundamentals’, Journal of Empirical Finance, 10, 5, pp. 623-640.
Rao, C.R., 1965, ‘The Theory of Least Squares when the Parameters are Stochastic’,
Biometrika, 52, pp. 447–458.
Rao, M.J.M. and Nachane, D.M., 1988, ‘Varying-parameter models: An Optimal Control
Formulation’, Journal of Quantitative Economics, 4, pp. 59–79.
Rao, M.J.M., 1991, ‘The Recursive Estimation of Time-varying Parameters: A Variation
on the Near-neighborhood Search Problem’, Journal of Quantitative Economics, 7,
pp. 303–311.

38

Volume 3 issue 2.indd 38 9/12/2010 10:13:31 πμ

 A Time-Varying Parameter Vector Autoregression Model for
Forecasting Emerging Market Exchange Rates

Rao, M.J.M., 2000, ‘Estimating Time-varying Parameters in Linear Regression Models

Using a Two-part Decomposition of the Optimal Control Formulation’, Sankhya,
62, Series B, Pt. 3, pp. 433–447
Rime, D., Sarno L. and Sojli, E., 2007, ‘Exchange Rate Forecasting, Order Flow and
Macroeconomic Information’, Norges Bank Working Paper Series ANO 2007/2.
Schinasi, G. J, Swamy, P. A., 1989, ‘The Out-of-Sample Forecasting Performance of
Exchange Rate Models When Coefficients are Allowed to Change’, Journal of
International Money and Finance, 8, pp. 373-390.
Somanath, S.V., 1986, ‘Efficient Exchange Rate Forecasts’, Journal of International
Money and Finance, 5, 2, pp. 195–220.
Vygodina, Anna V., 2006, ‘Effects of Size and International Exposure of the US Firms on
the Relationship between Stock Prices and Exchange Rates’, Global Finance
Journal, 17, pp. 214– 223.
Wolff, C.C.P., 1987, ‘Time Varying Parameters and the Out-of-Sample Forecasting
Performance of Structural Exchange-Rate Models’, Journal of Business Economics
and Statistics, 5, pp. 87–103.
Woo, Wing, T., 1985, ‘The Monetary Approach to Exchange rate determination Under
rational Expectations’, Journal of International Economics, 18, pp. 1-16.

39

Volume 3 issue 2.indd 39 9/12/2010 10:13:31 πμ