VOLATILITY MODELING OF ASSET RETURNS

A. W. Babayemi1 and B. K. Asare2 Department of Mathematics, Kebbi State University of Science and Technology, Aliero, Nigeria 2 Department of Mathematics Usmanu Danfodiyo University, Sokoto, Nigeria Corresponding author: A. W. Babayemi; E-mail: [email protected]
1

ABSTRACT This research was carried out using the daily close share price of Nestle Nigeria Plc to identify and model its volatility of returns in the Nigerian Stock Exchange Market. The result of the study showed that basic Generalized Conditionally Heteroskedastic Model (GARCH (1,1)) model (with Gaussian Error Assumptions) best describes the volatility of the returns. The volatility was found to be quite persistent. The resultant model to some extent also exhibited other stylized facts about financial time series. Keywords: Volatility clustering, assets returns, Gaussian Error Assumptions, stylized facts INTRODUCTION The volatility modeling of price returns originated with Engle (1982), where autoregressive conditionally heteroskedastic model (ARCH model), was used to predict the uncertainty of UK inflation rate. Engle noted, large changes tend to be followed by large changes of either sign and small changes tend to be followed by small changes This phenomenon was named volatility clustering. He measured the clustering effect through the assumptions of constant conditional mean of returns. Bollerslev (1986), a former student of Engle, generalized the ARCH model to the Generalized Conditionally Heteroskedastic Model (GARCH Model). The model enormously extended the ability of the ARCH model to account for the stylized facts of volatility of returns. This led to a surge in research studies involving equity price returns. The GARCH model was subsequently extended into a family of models called the GARCH family of models and this marked a revolutionary turning point. The GARCH family of models includes EGARCH Model (Nelson, 1991) and TGARCH Model (Glosten et al., 1993). Among the workers that made use of GARCH models to study the stock market time series include Bollerslev (1990), who estimated conditional correlation of GARCH model for five European currency before and after the implementation of the European monetary system and found that there was an increase in the level of conditional correlation. Akgiray (1989) and Balaban (1995), showed in their respective studies of the day of the week effect on returns, that returns of stocks vary by the day of the week effect. Other studies in this area include research by Campbell and Hentschell (1992), Karolyi (1995), Kearny and Patton (2000), Engle and Patton (2001) and Poon and Granger (2003). The common point of all these studies is that the report on returns (in the stock market) is time varying and conditionally heteroskedastic. In a univariate setting, it should be noted that a good volatility model should be able to capture and reflect stylized facts about asset returns (Engle and Patton, 2001). In practice, researchers have uncovered many characteristics about volatility of financial time series and these are the so called stylized facts. Bollerslev et al. (1994) gave comprehensive account of these facts. Similar researchers include Nelson (1991), Glosten et al. (1993), Engle (2001), and Loriano (2005). Among the notable stylized facts, include volatility clustering, volatility persistence, volatility mean reversion, fat tails, asymmetric effect, influence of exogenous variable and co-movement of volatility. In Africa and most especially Nigeria, little has been done in this area due to lack of reliable sources of information in the past. Hence, the objective of the study is to identify and model volatility of asset returns of Nestle Nigeria Plc using its daily close share price for 10 years period. MATERIALS AND METHODS Stylized facts: Volatility Clustering: This is the clustering effect of large and small moves (of either sign) in the pricing process. It was one of the first documented features of volatility process of asset returns. (Bollerslve et al., 1994; Engle, 2001).

Volatility Persistence: The implication of volatility clustering is the volatility persistence. The volatility shocks today will influence the expectation of volatility many periods in the future (Engle and Patton, 2001) Volatility Mean Reversion: This is generally interpreted as meaning that there is a normal level of volatility, in the end, to which volatility eventually return. At this point, the current information has no effect on the forecast. Asymmetric Effect: It implies the asymmetric impact of positive and negative innovations/shocks on conditional volatility. Most of the volatility models impose the assumption of asymmetric effect to give better forecast of the model. Fat Tails: High frequency financial time series usually have fatter tails than a normal distribution. That is large changes are more often to occur than a normal distribution would imply. Typical kurtosis estimate range from 4 to 50, indicating very extreme non-normality (Hamilton, 1994) Influence of Exogenous variables: Aside from series history effect, some variables are expected to be contained in the volatility, hence influencing the series. (Engle, 2001) Co-movement of volatility: This is commonality in changes across the stock. When stock volatilities change, they tend to change in the same direction. Modeling approach: The volatility modeling approach adopted here is that of ARCH/GARCH approach. The properties and features of the models with regard to stylized facts were examined. The model design and statistical tests used in the course of the research were discussed. The model: Engle and Patton (2001) specified that a good volatility model should reflect and capture the stylized facts of asset returns. The simplest model for studying volatility in univariate time series is Autoregressive Conditionally Heteroskedastic model of order p, denoted ARCH(p). The model was originally introduced by Engle (1982). For the series { }, the ARCH(p) model specification is = + 2.0 , ,,) N (0, )
, = where is the innovation/shock at day t and follows heteroskesdastic error process Asset returns at day t = Conditional mean of { } Squared innovation at day t-i The time varying conditional variance is postulated to be a linear function of the past squared innovations. A sufficient condition for the conditional variance to be positive is that the parameters of the model satisfy the following constant constraint GARCH (p,q) Model and its properties: In practice, it is often found that a large number of lags p and a large number of parameters are required to obtain a good model fit of ARCH(p) model. To circumvent this problem Bollerslev (1986) proposed the generalized ARCH or GARCH(p,q) model with the following formulation: iiN(0,1) 2.1 2.2

2.3 where, is the volatility at day t-i >0 for i = 1, ..,p for i = 1, ..,q and are previously defined. Under the GARCH(p,q) model, the conditional variance of , depends on the squared innovations in the previous p periods and the conditional variance in the previous q periods. The GARCH models are to obtain a good volatility model. Rearranging the GARCH (p,q) model by defining , it follows that , 2.4 2

where, L is a backshift operator and

By standard argument, the model is covariance stationary if and only if all the roots of (1) lie outside the unit circle. For simplicity the GARCH (1,1) model is used to investigate some stylized facts about financial time series. 2.5 where, measures the extent to which a volatility shock today feeds through into the next periods volatility, while measures the rate at which this effect dies over time. is the volatility at day (t-1). Volatility clustering and persistence: It is obvious in equation 2.5 that large/small values of will be followed by large/small changes . The same reason can be obtained for the ARMA representation in equation 2.4, where large/small changes in will be followed by large/small changes in and persistence comes about as a result of volatility shocks today influencing the expectation of many periods in the future due to the clustering. Fat Tailed distribution: Provided the fourth moment exist, Bollerslev (1986) and Engle et al. (2001) showed that the kurtosis implied by a GARCH(1,1) process is greater than 3, the kurtosis of a normal distribution. Bollerslev (2001) extend results to general GARCH(p,q) models. Thus, GARCH models can replicate the fat tails usually observed in the financial time series (Hamilton, 1994). Volatility mean reversion: Financial markets experience excessive volatility from time to time, but it appears that volatility will eventually settle down to a long run level. Writing the GARCH(1,1) ARMA representation as follows: 2.6 If the above equation is iterated k times 2.7 Where, variance reverts to its long run level is a moving average process, since <1 for the stationary model, hence 0 as k . At times, there may be a large deviation between and the long run will approach zero on average, as k gets large, i.e. the volatility mean . In contrast, if >1 and the GARCH model are non

stationary, then the volatility will eventually explode to infinity as k . GARCH Model extensions: In most cases, the basic GARCH model provides a reasonable good model for analyzing financial time series and estimating conditional volatility. However, there are some aspects of the model, which can be improved so that it can better capture characteristics and dynamics of a time series. The impact of advantage effect can best captured with GARCH extensions such as EGARCH (Nelson, 1991) and TGARCH (Glosten et al., 1993). Statistical tests: ARCH LM test: Before estimating a full GARCH Model for a financial time series, it is usually good practice to test for the presence of ARCH effect in the residuals. This test is used to check if the error model is truly skedastic function. The regression is given thus: where, are the co-efficients of the regression and

is the intercept.

= 0 There are no ARCH effects in the residuals under the null, the LM statistic is distributed asymptotically as Test Statistic: n The test is from Engle, (1982) who derive a LM statistic for testing the null as n number of observation and critical value . 3 , where n is the is from the regression. Reject the null if the p-value is less than statistic.

Structural breaks test: It has been shown by that breaks in variance, which are not taken into account by statisticians/econometrics will have ARCH effects, when the whole sample is used. In other word, it might be that for a sub-sample, the unconditional variance changes from say to a level and then back to the previous level. In this case, to model the conditional as an ARCH model will be the wrong thing to do. It is recommended to divide the sample and test for ARCH for the sub-periods. If no ARCH effects are found for any of the sub-periods but are found for the whole sample that is a clear indication of a break in the unconditional variance and not ARCH effect. Many researchers wrongly estimated GARCH models in many situations, where there was only change in regime or change in the operational management procedure of the company/firm in transactions. Portmanteau test: A portmanteau test is a test used for investigating the presence of autocorrelation in time series. Number of lag u and h are predetermined. Hypoththesis ... = 0 ( i.e all lags correlation are zero) (i.e i=1,.h is tested and at least one lag with non zero correlations). The test statistic is given (it is known as Ljung box statistic for correlation.) where = If the are residual from an estimated ARMA(p,q) model, the test statistic have an approximate (h-p-q) distribution. However, if the null hypothesis holds we reject it if P-value is less asymptotic

the significance level. Jarque Bera test: Jarque and Bera (1987) proposed a test for non-normality based on skewness and kurtosis of a distribution. The test checks the pairs of hypotheses. : 0 and (The distribution is symmetry and hence normal). : and (The distribution is asymmetric and hence non normal). Test Statistic JB is an asymptotic distribution

if the null hypothesis is correct and T is the number of observation. Reject the null if the p-value is less than critical value. Model design: In order to explore the volatility of this asset returns we computed based on the formula: where, the asset price (i.e. the share price) at day (t) the asset price at day (t-1) = the continuously compounded returns on the asset over day (t-1) to day (t). Estimating GARCH models: There was a number of methods use in estimating the parameters of volatility models, among these were: a. Maximum Likelihood Estimation with Gaussian Error b. Kalman Filter Method c. Efficient Method of Moment d. Quasi-Newton Method Here, we implemented the maximum likelihood method with assumption of Gaussian errors. This helped us to provide a general framework for the issue of estimating GARCH type models and requires the need for regression model with conditionally heteroskedastic error to be investigated. The function was given thus: where, = Likelihood function value Vector that contains the parameters of the GARCH process Vector that contains the parameters of the regression function Explanatory variables, which determine the value of conditional, mean of the series. 4 2.8

The return at day t Conditional volatility at day t The function follows optimization routines, which suggest that maximum likelihood estimator is that specific parameter vector that maximizes the likelihood function or equivalently the loglikelihood function. For further information, see Berndt et al. (1974). Software for analysis: We utilize JMulti 3.11 for the main and residual analysis of the research. It is a free software designed to carter for time series analysis. We also used statistica 6.0 and s plus 4.5 to plot the graph of actual returns and autocorrelation function respectively. Analysis of data: The data set was from Nigerian stock exchange (NSE). It is a daily close share price for Nestle Nigerian Plc from 02\01\1995 to 13\05\2005, which amounted for a period of 10 years. RESULTS AND DISCUSSION Time plot: The focus is on the returns , of the time series as opposed to their levels returns are obtained by taking the first difference of the logarithms of the . The time plot of the series is shown below in figure I . The level:

Fig. I: Daily share price returns for Nestle Nigeria Plc (02/01/1995-13/05/2005): The plot of the returns revealed that some periods were riskier than others were. There was some degree of autocorrelation in the risk of financial returns. The amplitude of the returns varied over time as large (small changes) followed large (small) changes in returns. The phenomenon is called volatility clustering (Engle and Patton, 2001) and is one of the stylized facts of the financial time series. More so, the plot depicted that as time progressed, the shocks tended to persist giving long memory process, with high volatility half-life in asset returns. This phenomenon is called volatility persistence, which is another stylized fact of financial time series. ARCH effect test: Testing whole sample, the result shows that there is ARCH effect at 5% level of significance and similarly there are no structural break effects as the result depicted across the sub-periods is significant at 5% for ARCH effects (Tables 1 and 2). Hence, the series could be modeled as conditionally heteroskedastic model. Table 1: Result of ARCH LM Test (assuming there are no structural breaks) Variable Test statistic p-value 418.7568 0.000 Autocorrelation Function (ACF): Having discovered that the returns could be modeled as conditional heteroskedasticity variance, the next step was to examine autocorrelation function (ACF) to determine the degree of autocorrelation in the data pts of the series. The plot (Figure II) showed that autocorrelation was stronger in the squared of returns. This indicates that there is substantial dependence in the volatility of the returns. We formally confirm the presence of autocorrelation in the returns and squared returns by portmanteau test. Table 2: Results of ARCH LM test for structural breaks 5

Variable test p-value

Sub period 1, 1995-1998 Sub period 2, 1999-2002 Sub period 3, 2003-2005 10.0363 0.001 61.2894 0.000 76.6105 0.000

Fig. II: ACF of

(LEFT)

ACF of the Squared

(RIGHT)

Portmanteau test: The null hypothesis of no autocorrelation cannot be accepted in all cases but we concluded that there were stronger autocorrelations in the squared returns and hence we modeled the squared returns (Table 3). Jaque Bera test for normality: Examination of the characteristic of the unconditional distribution of the asset returns enabled us to explore and explain some stylized facts that were embedded in the financial returns. Jarque Bera normality test was used to demonstrate this and the results were given in table 4. Table 3: Results of the Portmanteau test Test } Ljung Box 116.9741 p-value 0.000 Table 4: Result of the Jaque Bera Test variable Mean Variance 0.001465 0.0004378

Squared{ } 3058.5845 0.000

Skweness 0.1757

kurtosis 4.8447

Test stat 378.2819

p-value 0000

The result in the table 4 showed that small positive average returns of about one thousandth of a percent per day will be recorded for share price returns in Nestle Nigeria Plc and daily variance is 0.004378. This result implies average annualized volatility of 0.33% for share price returns in Nestle Nigeria Plc. The skweness coefficient indicates that the distributions of the returns are substantially skewed i.e. a symmetric effect is a common feature of financial time series returns. Finally, the kurtosis co-efficient, which is a measure of the thickness of the tail of the distribution shows that the returns exhibit a high kurtosis (>3 of the normal distribution). This is one the stylized facts known in the early day of volatility modeling and named fat tail. Model identification: The particular optimization routine of the log likelihood function specified in equation 2.8 showed that among ARCH(p) and GARCH (p,q) classes for p , the optimal model for the returns was GARCH(1,1) model. ARCH(5) appears more stable, but from the practical point of view, higher order than 5 would be more suitable. The more parameter involved in ARCH models the better the results. The results for the estimate of the log-likelihood function values were presented in table 5. For further information, see Berndt et al. (1974). Table 5: Results of the estimate of log-likelihood function for ARCH and GARCH models

Model ARCH (1) ARCH (2) ARCH(3) ARCH(4) ARCH(5)

Log L 6.5185 6.627 6.6807 6.7346 6.7628

Model GARCH (1,1) GARCH(2,1) GARCH(1,2) GARCH(2,2)

Log L 6.844 6.7478 6.7949 6.7062

As the order of the ARCH model was increased, the optimization process improved and this implied that ARCH(p), with sufficient number of terms could capture every volatility. To avoid such problem of l number of terms sufficiency and for parsimony, It was resolved to implement GARCH(1,1) as if it was the optimum among GARCH(p,q) models examined. Estimation of the model: With assumption of the Gaussian error distribution for the innovation , we estimated the parameters of GARCH(1,1) model. The result in table 6 indicated that volatility of the returns was quite persistent with the sum of and being 0.9916. This implies volatility half-life of about 490 days although the returns volatility appears to have a very long memory, the volatility is still mean reverting, since the sum of the and is less than one. (i.e. ) Table 6: Result of the estimated GARCH(1,1) parameters Variable } 0.00006239 (19.73) 0.1203 (17.31) 0.8713 (168.60)

Model checking: The results in tables 7 and 8 showed that the null hypothesis that there is no ARCH effect remaining at every lag should be accepted since the p-value was greater than 0.05. The null hypothesis of no ARCH effect for ARCHLM test should also be accepted at 5% since all the p- values are greater than 0.05 and we conclude that the models is adequate. Table 7: Result of No Remaining ARCH Effect Test in GARCH(1,1) residuals variable p-value 0.9701 } Table 8: Results of ARCH LM test for GARCH (1,1) residuals Lag Length Test statistic 1 0.5432 2 0.5448 3 0.9123 4 1.4171

p-value 0.4600 0.7622 0.8225 0.8412

The results in figure III and table 9 demonstrated lack of serial correlation in the residuals of the returns. More elegantly, another way to choose among equal models is to select the one that its standard deviation process best mimics the returns. Note the behavior of graphical estimate of the returns and its volatility (or otherwise its standard deviations) in figure IV. Forecast Evaluation: In the table 10, were the next two-weeks forecast. The convergence of the results implies the mean reverting of the volatility of returns. Volatility will eventually return after sometimes to a normal level. The ability of the model to exhibit or demonstrate this stylized fact further testifies the validity of the model.

Fig. III: GARCH (1,1) residuals2 Table 9: Portmanteau test for GARCH(1,1)residuals Test Ljung Box

p-Value 0.9605

Fig. VI: Estimated Conditional Volatility of Nestle Nigeria Plc using GARCH(1,1) Model Table 4.0 Results of Forecast and validation for Nestle Nigeria Plc Time Forecast Lower C.I Upper C.I 2005 63\250 0.0014 -0.0390 0.0417 2005 64\250 2005 65\250 2005 66\250 2005 67/250 2005 68/250 2005 69/250 2005 70/250 2005 71/250 2005 72/250 2005 73/250 2005 74/250 2005 75/250 2005 76/250 CONCLUSION The daily close share price of Nestle Nigeria Plc was used to identify and model volatility of asset returns. The initial analysis showed that there was evidence of ARCH effects in the returns and that 8 -0.0001 0.0023 0.007 0.0018 0.0016 0.0015 0.0016 0.0014 0.0015 0.0014 0.0014 0.0015 0.0015 -0.0408 -0.0402 -0.0393 -0.0394 -0.0396 -0.0396 -0.0395 -0.0397 -0.0396 -0.0397 -0.0397 -0.0397 -0.0397 0.0407 0.0432 0.0416 0.0428 0.0427 0.0426 0.0427 0.0425 0.0426 0.0426 0.0426 0.0426 0.0426 +/_ 0.0404 0.0407 0.0409 0.0409 0.0410 0.0410 0.0411 0.0411 0.0411 0.0411 0.0411 0.0411 0.0411 0.0411 Actual returns 0.0405 0.0611 0.0387 0.0231 0.040 0.0134 0.0213 0.0129 0.0352 0.0417 0.0580 0.0541 0.0560 0.0574

the ARCH effects were not due to structural breaks as there was ARCH effects across the sub periods. Further investigation showed that basic GARCH (1,1) model (with Gaussian error assumption) best describes the volatility of the returns. The volatility was found to be quite persistence i.e. current volatility can be explained by past volatility that tends to persist overtime. The resultant model exhibited all other stylized fact such as volatility clustering, mean-reverting and fat tailed distribution which suggests large expectation in speculative prices. Finally, the result of the forecast values also shows that there is a normal level to which volatility will eventually return, as the values tends to converge to a point.

REFERENCES Akgiray, V. (1989). Conditional Heteroskedasticity in time series of stock returns: Evidence and Forecasts, Journal of Business, 62: 55-80. Balaban Ercan (1995). Day of the week effects: New evidence from an emerging market. Applied Economic Letter, 2: 139-143. Berndt, E., Hall, B., Hall, R. and Hausman, J. (1974). Estimation and inference in non-linear structural models. Annals of Economic and Social Measurements, 3/4: 653-665. Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics, 31: 307-327. Bollerslev, T. (1990). Modeling the coherence in short run nominal exchange rates: A multivariate generalized arch approach. Review of Economics and Statistics, 72: 498-505. Bollerslev, T., Engle, R. F. and Nelson, D. B. (1994). ARCH Models. In: The handbook of Econometrics, Amsterdam, North Holland. Pp: 2959-3038. Campbell, J, Y. and Hentschel, L. (1992). No news is good news: An asymmetric model of changing volatility in stock returns. Journal of Financial Economics, 16: 99-117. Engle, R. F. (1982). Autoregressive conditional heteroskedasticity with estimates of the variance of United Kingdom inflation, Econometrica, 50(4):987-1007. Engle, R. F. (2001). Financial econometrics - A new discipline with new methods. Journal of Econometrics, 100: 53- 56. Engle, R. F. and Patton, A. J. (2001). What good is a volatility models? IOP Publishing Limited, Britol. Glosten, L. R., Jagannathan, R. and Runkle D. E. (1993). On the relation between the expected value and the volatility of the nominal excess return on stocks. Journal of Finance, 48: 17791801. Hamilton, J. D. (1994). Time series analysis. Princeton University Press, Princeton. Jarque, C. M. and Bera, A. K. (1987). A test for normality of observations and regressive residuals, International Statistical Review, 52(2): 163-172. Karolyi, G. A. (1995). A multivariate GARCH modeling of exchange volatility transmission of stock returns and volatility. Business and Economic Statistics, 13(1): 11-12. Kearney, C. and Patton, A. J. (2000). Multivariate GARCH modeling of exchange rate volatility transmission in the European monetary system. Financial Review, 41(1): 29-48. Loriano, M. (2005). Introduction to GARCH type models. Retrieved on 21/09/05 from: [email protected];www.princeton.edu/~imacini. Nelson, D. B. (1991). Conditional heteroskedasticity in asset returns. Econometrica, 59: 347-370. Poon, S. and Granger, C. W. J. (2003). Forecasting volatility in financial markets, Journal of Economic Literature, 41: 478-539.

10