The Predictive Power of Alternative Volatility Forecasting Models over Multiple Horizons

Abstract

This thesis paper examines the forecast accuracy and explanatory power of volatility models over multiple forecast horizon for three asset classes. Forecast horizon ranging from 1 month up to 12 subsequent months are investigated using Naïve, EWMA, GARCH, EGARCH, GJR-GARCH and APARCH model for S&P 500, DJIA, CBOE(^TNX ), CBOE(^FVX), USD/CHF and GBP/CHF. MSE and Predictive Power (𝑃�) are used to evaluate the forecast accuracy and predictive ability of the model over increasing horizon. Different distribution assumptions are also included with non-linear GARCH models in an attempt to improve forecast accuracy of the models. The in-sample estimation results revealed increased model fit for all assets considering the non-normal innovation but correspondingly didn’t always comply with out-of-sample forecast accuracy. Non-normal distribution provided best forecast accuracy at short forecast horizons for all asset classes except exchange rates. The result common to asset classes was that forecast accuracy and predictive power of the model are best at short horizon which gradually decreased with increasing forecast horizon. The predictive power suggested the longest forecastable horizon for Stock Indices, Interest Rates and Exchange rates are 4 months, 12 months and 2 months respectively. The results showed EGARCH model performed relatively well compared to other models and was able to increase the forecastable horizon. Further, it was concluded there is no best model for all asset classes over all horizons. The best model is largely dependent upon the type of asset and the horizon of interest.