March 10, 2014 Abstract This paper tries to analyze nancial time series data using various econometric techniques. 1 Data Description The study looks at 2 data series. The series are the National Stock Exchange (NSE) Nifty Index and the Nifty Junior index. The period of study is for 10 years from 01/01/2004 to 31/12/2013. There are 2493 observation per series totalling to 4986 data points The software packages Stata 11 and Eviews 6 have been used to analyze the data. 2 Descriptive Statistics The following table gives the descriptive statistics of the raw data. 1 Particulars Nifty Junior Nifty Mean 4241.702 8168.192 Median 4676.950 8597.700 Maximum 6363.900 13555.15 Minimum 1388.750 2685.100 Standard Deviation 1435.469 3026.478 Skewness -0.468699 -0.189799 Kurtosis 1.870381 1.669637 Jarque Bera 223.8249 198.8127 Probability 0.000000 0.000000 Both the indices are negatively skewed and have high volatility. The line graphs show a general upward trend with two signcant drops, one in the period 2008-09 and the second around the year 2012. The Junior Nifty seems to be more volatile than the Nifty. 3 Stationarity tests 3.1 Testing for unit root in the raw data The corellogram of both data series show that the series are non-stationary. The following set of tables presents the results of the stationarity tests on the raw data (levels):- The Augmented Dickey Fuller (ADF) Test for a unit root The null hypothesis is that there is a unit root (the series is non station- ary) 2 Series p values Nifty 0.6149 Nifty Junior 0.6426 The Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test of station- arity The null hypothesis is that the series is stationary. Series p values Nifty 0.0000 Nifty Junior 0.0000 From both the ADF and KPSS tests it can be concluded that the series are non-stationary. 3.2 Transformation of data In order to make the series stationary, the data has been converted to its natural log and the rst dierence has been taken. The line graphs of the transformed series and the correlograms show that the series now seem to be stationary. No trend can be discerned from the line graphs. 4 Testing for unit root in tranformed data The Augmented Dickey Fuller (ADF) Test for a unit root The null hypothesis is that there is a unit root (the series is non station- ary) 3 Series p values Nifty 0.0001 Nifty Junior 0.0000 The Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test of station- arity The null hypothesis is that the series is stationary. Series p values Nifty 0.1466 Nifty Junior 0.1432 It can be observed that both the series are now stationary i.e., there is no unit root. 5 Autoregressive/Moving average models Once the data is stationary, the univariate series could be modelled. In order to study the process, the correlogram is used. From the autocorrelation and partial autocorrelation plots, it is not pos- sible to comment on the choice of the model nor the lag-length. The process is neither pure autoregressive (AR) or a moving average (MA.). Therefore the only way is to use the Akaike Information Criterion (AIC), Hannan Quinn Information Criterion (HQIC) or the Schwarz Bayesian Infor- mation Criterion. The model which gives the lowest value of the criterion will be chosen. It is important to remember that the data has been dierenced already. The model that t the best for the Nifty returns series was an ARMA(2,2) model. An ARMA process is one where the current value of some series de- 4 pends linearly on its own past values plus a combination of current and pre- vious values of a white noise error term. The results show that the two lags of both the AR process and the MA process are signicant. An ARMA(3,3) model was also signicant but the ARMA(2,2) was better based on the in- formation criterion. In the case of the Nifty Junior returns series, an ARMA(1,2) showed best t. The MA(2) coecient however is signicant at 10% only. 6 Vector Autoregression Vector Autoregression (VAR) is a systems regression model. All variables in a VAR model are endogenous. The lag length was selected using the information criterion. The model was run with lag lengths of 1, 2 and 3. It was observed that the rst lag with Nifty returns as the independent variable and the Junior Nifty returns as the dependent variable was signicant and not vice-versa. It can be inferred that the Junior Nifty can be predicted using the lagged values of the Nifty or that the Nifty leads the Junior Nifty. The strange behavior however is that the lag of the Junior Nifty returns does not explain itself i.e., the coecient of its own lags are not statistically signicant. The residuals have been tested for serial correlation in the VAR model with two lags. It is seen from the VAR Lagrange Multiplier (LM) test that there is serial correlation in the residuals. 5 7 Impulse response The impulse response tests the responsiveness of the dependent variables in the VAR to shocks in each of the other variables. A unit shock is applied to the error and the eects on the VAR are noted over time. It is observed that the shocks die down after a short period of time. The VAR result is conrmed by the impulse response graphs as there is an impact on the Junior Nifty when a shock is given to the Nifty. 8 Granger Causality The Granger causality test is performed in a VAR framework. The VAR model as specied above is run. The VAR Granger Causality/Block Exogeneity Wald Tests have been used to determine Granger causality. It shows that the Nifty series seems to Granger cause the Junior Nifty series. Granger causality does not imply causality in the true sense. It simply implies a chronological ordering of the series. Hence the Nifty returns series seems to lead the Junior Nifty returns series. 9 Cointegration The cointegration is done when there is a presence of unit root. It shows the long term relationship between time series. In this study, the raw series are used. Cointegration is tested using the Johanssen approach. The tests suggest that there is 1 cointegrating relationship between the two series. The relationship shows a quadratic trend. The equation has a trend and intercept. The VAR however has a linear trend. 6 10 Modelling volatility Financial data have some peculiar characteristics. This include leptkurtosis (fat tails), volatility clustering and leverage eects (a large rise in volatil- ity following a rise in price and a small fall for a fall in price of the same magnitude). In order to model these characteristics various non-linear mod- els are used. Popular among these are the Autoregressive Conditional Het- eroscedasticity (ARCH) model, the Generalized ARCH (GARCH) model and the Exponential GARCH (EGARCH) model. The GARCH model is parsimonious and avoids overtting. Firstly the series was tested for ARCH eects. It was found that both the series had ARCH eects at 1 lag. The GARCH model was then tted to tha data. It is observed that coecients on both the lagged squared residual and lagged conditional variance terms in the conditional variance equation are highly statistically signicant. The sum of the coecients on the lagged squared residual and the lagged conditional variance of the Nifty and the Junior Nifty returns series is close to unity, dominated by the coecient of the conditional variance. In both models, the coecients are positive. Therefore there is no risk of forecasting negative variance. 11 Conclusion This paper looks at two dierent nancial time series and tries to analyze them and help build models for forecasting. 7