This study investigates the daily volatility forecasting performance of various GARCH models for China Securities Index–300 series from 2002 to 2010, which has not been empirically examined. The high-frequency realized volatility is employed as proxy for latent true volatility. M-Z regression and loss functions such as MSE, MAE and MAPE are used to evaluate the relative performance of competing GARCH-based forecasting models. GARCH models with asymmetric specifications and alternative distributional assumptions are examined to identify the source of performance improvements. Empirical results suggest that CGARCH model achieves the most accurate volatility forecasts. Such evidence, along with the results of sign bias tests, demonstrates that modeling long-term persistency is more important than specifying asymmetric components in GARCH models for improving volatility forecasts of financial returns. Furthermore, the empirical evidence indicates that the GARCH models with Gaussian distribution are consistently preferable to those with more sophisticated error distributions.
Keywords: realized volatility, GARCH, forecast evaluation, CSI-300.