Abstract: Based on the theory of corrected realized threshold multipower variation(C_TMPV), the jump components of the realized volatility are estimated, and two newly developed realized volatility model allowing for jump, the AHAR-RV-CJ model and MIDAS-RV-CJ model, are proposed to predict realized volatility of Chinese Stock Markets. The forecast accuracies of several volatility models are also evaluated and compared. Our findings demonstrate that the jump components of the realized volatility estimated by C_TMPV have positive and significant impacts on daily, weekly and monthly volatility prediction, and the AHAR-RV-CJ model and MIDAS-RV-CJ models with the continuous and jump components of the volatility are the best models for future volatility prediction in different prediction horizons. These results hold up for both the in-sample and out-of-sample forecasts, especially the logarithmic models. It is also found that the out-of-sample forecasting performance of MIDAS model is better than HAR model with the same regressor and the out-of-sample predictive power of AHAR-RV-CJ and MIDAS-RV-CJ models is better than Jump-GARCH, SV-CJ and SV-IJ models in the medium and long prediction horizons.

Key words: volatility forecasting, realized volatility, C_TMPV, MIDAS model, SPA test