Application Research Of Deep Learning In Financial Time Series Prediction

Today,with the globalization of the economy,financial markets,especially the stock market,have attracted more and more attention.The stock price index is an indicator compiled by financial institutions,who select certain sample stocks,to reflect the overall price changes of stock market.The stock price indexs have extremely complex nonlinear dynamic characteristics,and the characteristics of instability and chaos.Predicting financial time series such as stock price indexs can help enterprises quickly raise fund and expand the scale,guide investors to make correct decisions and bring higher returns,and help countries centralize idle funds,give full play to market mechanisms and improve the efficiency of resource allocation.Compared with econometrics and machine learning,deep learning has a stronger feature expression ability.Unsupervised learning is used for feature extraction.While improving the prediction accuracy,it can solve the fitting problem and has better generalization.The combination model proposed in this thesis is improved on the traditional EMD-LSTM model.Different components of frequencies are predicted and combined by using the deep neural network model,which improves the accuracy of prediction and has a stronger generalization ability than tradional networks.First of all,select the Shanghai industrial index’s five-dimensional index from April 17,2017 to April 15,2022 includes the closing price,opening price,the highest price,the lowest price and the transaction volume.The five-dimensional index is decomposed in an empirical mode decomposed way,and then we can obtain many IMF components and RES component.Then calculate the number of each component by the travel range judgment method.The IMF component with a range of greater than 100 is reconstructed into a high-frequency term,the IMF component with a range of less than 100 is reconstructed into a low-frequencuy term,and the RES component is a trend term.Then,the high-frequency terms,low-frequency terms and trend terms obtained after decomposition of the closing price of the Shanghai industrial index are put into the CNN-GRU model,CNN-RNN model and BP neural network model respectively.The quantitative evaluation indicators such as RMSE,MAPE,MAE and R~2,are uesd to adjust the number of windows and the hidden layers of the model.The results of component prediction which has the best fitting effect and the smallest error are added to obtain the predicted closing price.Compared with other models,the results show that the combined model can effectively improve the prediction accuracy.Finally,the combined model can be extended to the four constituent stocks of the Shanghai industrial index,such as Shanghai Petrochemical,ENN group,Sailun group and Dongfeng Motor,to prove that the combined model has a certain universality and generalization.