Research On Realized Volatility Prediction Based On Window Average Hybrid Model From The Perspective Of High-order Moments

Volatility refers to the degree of volatility of financial asset prices,is a measure of the uncertainty of asset returns,and is used to reflect the risk level of financial assets.Volatility plays a very important role in asset pricing and allocation,risk management,and portfolio decision making.Based on the CSI 300 stock index market and stock index futures market,this paper constructs the realized volatility,realized skewness and realized kurtosis of the two markets,to study the predictive ability of the realized high-order moments of two markets on the realized volatility of the index market.Adopting the HAR-RV model,and considering the directionality of realized fluctuations and the importance of jumps,the HAR-RV-RS model and the HAR-RV-CJ model are introduced,taking the realized high-order moments of the two markets as the exogenous variable to add to the HAR family model.According to the OLS empirical results,it is found that the HAR-RV-RS model has the best performance among the three models,and the realized high-order moments of the two markets can effectively improve the prediction accuracy of the realized volatility of the index.The improvement ability of the realized high-order moments of the index market is better than that of the stock index futures market’s.However,the specific improvement of different variables is related to the selected HAR model,among which the optimal model is the HARRV-RS model which added the realized kurtosis of the index market.Considering the limitations of OLS,in addition to OLS method,non-linear Support Vector Regression(SVR)method and Long Short-Term Memory network(LSTM)were adopted in this paper.Besides,the sample included the stock market crash of 2015,resulting in model uncertainty and parameter instability.In order to overcome the realized volatility forecast accuracy degradation problems caused by systemic structure mutation,on the basis of a single method,the data-driven windows average prediction method is proposed.The empirical results show that the LSTM can capture the sequence of contact for a long time,although make some realized higher-order moments after adding model can’t improve the prediction accuracy,its performance is still better than that of non-linear SVR method than OLS method.In addition,the windows average prediction method based on these three methods has obvious improvement over the single method.The optimal method is the windows average prediction method based on LSTM.In addition,the windows average prediction method is used as the benchmark,and the windows median average prediction method is proposed to compare with it.Further,based on the windows average prediction method,considering the performance of the model in different windows,the windows MAE weighted average prediction method and the windows Lasso regression prediction method are proposed.It can be seen from the empirical results that the windows median average prediction cannot be stably improved;the windows MAE weighted average prediction method based on OLS and SVR is not significantly improved compared with the windows average prediction method,and only slightly improved based on LSTM;the windows Lasso regression prediction method has a significant improvement on these three methods.Finally,the robustness of the model is tested by changing the length of training samples,and consistent conclusions are obtained,which further increases the reliability of the conclusions in this paper.