Stock Market Risk G-VaR Prediction Based On Deep Learning Model

With the accelerated development of economic globalization,the contagiousness of financial markets has increased accordingly,and as a core component of financial markets,the volatility of stock markets has also increased significantly.In view of the foregoing,how to ensure as much working capital as possible for investors while effectively resisting market risks has become an increasing concern for financial investment regulators.Compared with the traditional risk measurement indicator--Value at Risk(VaR),the risk measurement indicator under the framework of nonlinear expectation theory--VaR under the assumption of G-normal distribution(G-VaR),has higher data consistency and prediction validity in risk measurement because it takes the distribution uncertainty into account and meets the consistency requirements.Therefore,this paper aims to improve the prediction effect of G-VaR,a stock market risk measurement indicator,through the application of deep learning model.Therefore,this paper aims to improve the prediction effect of G-VaR,a stock market risk measurement indicator,through the application of deep learning model.In this paper,we use the Hang Seng Index(HSI)of Hong Kong stock market as an example to compare two methods for predicting G-VaR:the method based on deep learning model and the method based on the hybrid model of ICEEMDAN decomposition.Specifically,the method based on deep learning model is described as follows.The maximum mean estimation proposed by Peng and Yang is first used to obtain the estimated values of the three parameters of mean,upper volatility and lower volatility in the small-window case.Then eight deep learning models,including the LSTM,the GRU,the LSTM-Attention,the CNN-LSTM-Attention and the Informer,are used to predict parameters respectively.Finally,the predicted value of G-VaR is obtained.The method based on the hybrid model of ICEEMDAN decomposition first assumes that the decomposed components of the stock index log return series are independent under sublinear expectations,and then models the high frequency,low frequency and trend components that have proved to be independent respectively.For the high-frequency component,the ARMA-GJR-GARCH model is used to predict its mean and volatility,and the predicted values of upper volatility and lower volatility are calculated in combination with the sliding window.The maximum mean estimation in the small window case proposed by Peng and Yang is used to obtain the estimated values of the three parameters.The Informer model is then used to obtain the predicted values of these parameters.For the low-frequency component,given the length and number of sliding windows,the maximum mean estimation yields its parameter estimates,and the Informer model is used to predict the parameters.For the residual component,the parameters are estimated in the same way as for the low-frequency component,and then the ARMA model is used to predict the parameters.Finally,the predicted value of GVaR is calculated.In order to select the method with the best performance from the methods between using deep learning model and using hybrid model based on ICEEMDAN decomposition to predict G-VaR,we conduct comparative experiments on these methods,selecting binomial distribution test,Kupiec test,Christoffersen test,default rate and mean absolute deviation rate as evaluation indicators.The results show that among the G-VaR prediction methods used in this paper,the method using the Informer model has the best effect among the methods based on deep learning model.Additionally,the method using the Informer model for high frequency sequences has the best effect among the hybrid model methods based on ICEEMDAN decomposition,and its prediction effect has been further improved.