GARCH Models have become a workhouse in volatility forecasting of financial and monetary market time series. In this article, we assess the small sample properties in estimation and the performance in volatility forecasting of four competing distribution free methods, including quasi-maximum likelihood and three regression based methods. The study is carried out by means of Monte Carlo simulations. To guarantee an utmost realistic framework, simulated time series are generated from a mixture of two symmetric generalized error distributions. This data generating process allow to reproduce the stylized facts of financial time series, in particular, peakedness and skewness. The results of the study suggest that regression based methods can be an asset in volatility forecasting, since model parameters are subject to structural change over time and the efficiency of the quasi- maximum likelihood method is confined to large sample sizes. Furthermore, the good performance of forecasts based on the historical volatility supports to use the variance targeting method for volatility forecasting.
GARCH volatility forecasting Monte Carlo simulation mixture of generalized error distributions , variance targeting