Empirical Research On Stock Index Volatility Prediction And VaR Measurement Based On Deep Learning

Financial risk management is an important part of preventing risks,resolving risks,maintaining market stability and rapid development,and volatility forecasting is an important tool and core variable of risk management.Therefore,in terms of volatility forecasting and risk management,in dealing with and fitting the complex nonlinear structure of the current financial system,this paper hopes to find new methods based on traditional econometric models.To solve this problem,this article aims to combine deep learning with organic standard model algorithms.In this case,the volatility of VaR prediction is modeled and analyzed based on deep learning.Improve prediction accuracy.Firstly,this paper summarizes the history and research status of financial market volatility prediction and VaR measurement;Then,it introduces the relevant theoretical basis and model construction;Secondly,we use the single GARCH model,LSTM model and their seven fusion models to model the volatility,evaluate the volatility prediction effect through the loss function,and take the model prediction result with the best comprehensive performance as the input variable of the subsequent VaR measurement model;Finally,on the basis of the Quantile Regression Neural Network(QRNN)model,considering the lag endogenous variables,the Quantile Autoregression Neural Network(QARNN)model is established to measure VaR,and the performance of the model is compared and evaluated through the return test.Main research results:(1)In the volatility prediction,we compare the volatility fitting prediction ability of GARCH model under different residual distribution assumptions.Based on the results of multiple loss functions,it is proved that the accuracy of the model prediction can be improved when the residual is assumed to be subject to GED;Compared with the performance of single model through loss function,the prediction accuracy of LSTM model is higher than that of GARCH model;After using GARCH model to obtain model parameters through rolling prediction,GARCH model parameters are added on the basis of volume and price data as input of LSTM model,and the performance of the fusion model obtained is generally better than that of single model;Among the fusion models,the fusion model that integrates three GARCH models and LSTM models has the best performance.(2)In the VaR measure,the Quantile Regression(QR)model and Quantile Autoregressive(QAR)model representing the traditional econometric model are underestimated in the test set by returning the accuracy of the VaR measure of the comparison model;The model effect of QRNN model is better than that of QR model and QAR model,the failure rate is close to the expected failure rate,and the VaR measure is more reasonable;The QARNN model has the best overall performance,and the failure rate is lower than the expected failure rate.This model not only takes into account the key impact of volatility on VaR,but also takes into account the nonlinear effect of lag endogenous variables,which significantly improves the accuracy of the model.In summary,in the volatility prediction of the Shanghai Composite Index,the LSTMGARCH model combining traditional econometric model and deep learning algorithm has better performance than a single model,and provides more accurate volatility input variables for VaR measurement.In terms of VaR measures,the QARNN model,which comprehensively considers lagging endogenous variables and integrates traditional econometric models,greatly improves the prediction performance of traditional quantile regression models on VaR measures.