Research On Influencing Factors Of Stock Market Time Series Prediction Accuracy Based On LSTM Deep Neural Network

Deep neural network is a major breakthrough in the field of machine learning in recent years.It has made a lot of achievements in machine learning,picture recognition,speech recognition and other related fields.Recurrent Neural Network is widely used in the study of time series.As a derived model of RNN,LSTM is good at discovering the nonlinear relationship between time series data and suitable for stock market series prediction.Whether the stock market series can be effectively predicted,it depends on whether the LSTM neural network learns the implicit essence from the training sample.From the analysis of the operational mechanism of RNN,the selection of training samples,the network structure,and the optimization algorithm all have the influence on the prediction accuracy of the model.Based on the LSTM deep neural network,this thesis makes a model to forecast the stock movement and make a study on the three factors which influence the prediction accuracy of the model.In the aspect of training samples,this thesis examines the selection of data characteristics,the sequence length of time series,and the number of samples on the prediction accuracy of the model.For the input characteristics,this thesis comprehensively considers the factors that affect the stock price fluctuation,and extracts the basic transaction class index,the technical index,the market index and the financial index.The various characteristics and their combination are taken as the input variables of the network.In addition,the principal component analysis method is used to reduce the dimension and establish the contrast model to study the influence of four kinds of input features and principal component analysis on the prediction accuracy of the model,then choose the best input characteristic.In the aspect of model structure,the model of different hidden neurons is trained and the training results are compared and analyzed to explain the influence of model structure on prediction accuracy.In the aspect of optimization algorithm,SGD,RMSprop and Adam algorithm are used to optimize the network training process.Finally,the optimal algorithm for LSTM depth neural network is selected.In this thesis,the data are further refined,the market status is divided into bull market and bear market,and the data of different market conditions are input into LSTM deep neural network respectively.The model is optimized by dropout and L2 regularization technique.Finally,the BP neural network,traditional RNN and LSTM deep neural network are compared and analyzed to illustrate the excellent performance of LSTM deep neural network.The thesis takes the stock data of Shanghai and Shenzhen 300 constituent stocks from 2006 to 2017 as samples.Based on the LSTM depth neural network,the three factors influencing the prediction accuracy of the model are systematically studied.The principal component analysis and the dropout and L2 regularization are used to optimize the model.The thesis has some reference significance for the training samples selection and parameter selection in the recurrent neural network training process,such as LSTM,and provides some theoretical and practical value for constructing the stock short-term forecasting model and the application of the deep neural network in the financial market.