The Extent of Volatility Predictability Evaluation of forecasting accuracy dependent on time, distribution and model order

Abstract

This thesis focuses on the accuracy and ability of out-of-sample volatility forecasting over different time horizons. Using data at daily frequency we forecast the future volatility over multiple time horizons (1, 3, 6, 9 and 12 months) and evaluate the goodness of forecasting by comparing the Naïve, ARCH, GARCH, EGARCH and GJR-GARCH models using the MSE and the Predictive Power (P). We include different probability distributions for the error terms in an attempt to improve the models accuracy. The research is conducted using three indices: FTSE 100, S&P 500 and the Hang Seng. We find that the goodness of forecasting accuracy decreases dramatically after the 3 month horizon and the selection of a more representative error distribution improves the accuracy of the short term forecasts. The results also show that the higher order GARCH models, beyond (1,1), do not improve the forecasting accuracy.