Volatility forecasting in emerging markets with application of stochastic volatility model

Abstract

The volatility of financial asset returns is a key variable in risk management and derivative pricing. The behaviours of emerging equity markets are now significant to global economies. This research examines the performance of five popular categories of volatility forecasting models on 31 emerging and developed stock indices with data series comprising recent 7 years. A modification in estimation processes of the Stochastic Volatility Model (SVM) is proposed. The empirical analysis shows that the equity markets of emerging markets are more volatile and difficult to model than those of developed countries. The SVM performs well in both settings, and has a clear advantage in developed markets.

Acknowledgements

This research was supported in part by the National Science Council of Taiwan (NSC96-2415-H-155-001). The author thanks Chi-Liang Me and Wei-Ling Chang for their excellent research assistance where part of data managements of this study are extracted from their Master theses.

Notes

¹ For a detailed discussion, please refer to Poon and Granger (2005).

² A detailed survey can be found in Ghysels et al. (1996).

³ Please see Carr et al. (2003) and Wu (2006) for details.

⁴ For detailed arguments on this topic, please see Jacquier et al. (1994) and Andersen and Sørensen (1996).

⁵ Please see Gallant and Tauchen (1996) for details.

⁶ The selection of emerging markets is based on the June 2007 MSCI Emerging Markets Index. Two countries, Jordan and Thailand, are excluded due to lack of data.

⁷ All data in this study are taken from the Thomson DataStream database.

⁸ The convergence criterion is then adjusted to 0.01 for the rest of forecasts being produced.

⁹ The daily trading volumes and open interests of selected futures contracts are also extracted from DataStream.

¹⁰ The R ² coefficient is the usual 1-SSR/SST statistic used in multivariate GARCH(1,1) frameworks.

¹¹ For simplicity, the table provides results only for four models: VOL20, MC, GARCH and SVM. The results of VOL30, VOL60, IGARCH and GJRGARCH are very similar to VOL20 and GARCH. The results of quantile regression models are also excluded due to their greatly inferior outcomes.