Research Of Quantitative Transaction Based On Deep Learning

At present,the domestic quantitative investment market is in a steady stage of development and machine learning technology has achieved considerable success in the application of various fields.Therefore,the use of machine learning to conduct quantitative trading has gradually become a hot point in interdisciplinary research,which will make quantitative investment not a portfolio strategy constructed by a simple combination of technical analysis,but a prediction model build by relevant algorithms.The predicted results can provide reference for investors to make investment portfolio.In this paper,five models are established by Recurrent Neural Network to predict the closing prices of the CSI 300 stock index and 100 weighted stocks in CSI 300 stock index.The main work of this paper is as follows:(1)Model 1 selects the closing price of CSI 300 stock index as the input features.The closing price of each trading day is intercepted through the sliding window technology and the price of current window is compared with that of previous window,which can calculate the relative change rate of closing price.Therefore,LSTM model is build by the relative rate to predict the closing price of the next period.But the predicted result is not ideal.(2)Model 2 selects six kinds of basic data including opening price,closing price,ceiling price,bottom price,trading volume and the value of deals of CSI 300 stock index for each trading day as the input features to construct LSTM model to predict the closing price of the next trading day.Super parameter is tuned on the validation set,so the model has excellent generalization capability on the test set.(3)Based on Model 2,Model 3 changes the neural network structure to RNN.After the iterative training is completed,Model 3 is compared with Model 2,which finds that the prediction accuracy of LSTM model on the time series trading data of the stock is higher than that of RNN model.(4)Model 4 selects the technical indicators of the stock market as the input features to construct LSTM model,which finds that the technical indicators cannot improve the prediction accuracy of the model.(5)Model 5 selects historical trading data of 100 stocks in CSI 300 stock index weighting stocks as input features to build LSTM model.Through the analysis of the experimental results,it is found that LSTM network has a good result in fitting the closing price curve of a single stock.(6)According to the prediction results of models M2 and M5,2019 is taken as the back testing period for quantitative investment.When the number of fund segmentation is 20,30 and 40 respectively,the annualized return of the algorithm is 17.2%,9.12%and 5.26%respectively,and the investment success rate of all investment periods is 74%,66.7%and 62.96%respectively.