Prediction Of CSI 300 Index Based On Deep Learning

The stock market plays an important role in the development of national economy and is called "the barometer of economy".If we can predict the future trend of the stock market more accurately,it will not only help the government to make better macro-control policies,but also help investors to obtain wealth.The nonlinearity and complexity of stock market lead to the traditional linear time series model can not achieve good results.Due to the ability of deep learning model to capture nonlinear characteristics,predicting stock price index has become an important application field of machine learning research.In this thesis,a novel deep learning model(CNN-LSTM-AM)is proposed to predict the short term CSI 300 index price by combining convolutional neural network,long and short term memory neural network and attention mechanism.The research work of this thesis is mainly reflected in the following aspects:(1)This thesis collected a total of 4,288 daily level data of CSI 300 index from April 8,2005 to November 28,2022 through the financial data platform of Flush i Find.After a series of data preprocessing,27 kinds of characteristic data are obtained as experimental data sets.(2)In this thesis,Radam is selected as the optimization algorithm for back propagation,and particle swarm optimization(PSO)is used to determine the superparameters of CLA model.This article also conducted a manual search for the size of the sliding window and finally determined it to be 15 days.In order to avoid overfitting problems,this thesis not only adds the Dropout layer to the model,but also sets an early stop during training.(3)In this thesis,the training set is used to train the CLA model,and the verification set is used to judge whether the model meets the condition of early stop.After the training,the CLA model predicted the closing price of CSI 300 index in the next 1,2,3 and 5 days on the test set.The results showed that the model was the best predictor of the next day’s closing price.(4)In this study,the same data set was used to compare the predictive power of LSTM,CNN-LSTM,LSTM-AM and CLA models.The results show that the addition of attention mechanism is crucial to the improvement of the prediction ability of the network model,and the R square of the CLA model reaches 0.954,showing the best prediction ability in the four groups of experiments.(5)In this thesis,a generalization experiment is carried out.The SSE index,SSE 50 and GEM index are respectively used as experimental data for prediction experiment of the above four network models.The results show that,compared with the other three models,CLA consistently shows better prediction ability when facing different data sets.(6)In this thesis,a simple quantitative investment strategy is designed based on the predicted value of the CLA model,and the backtest analysis is carried out on the CSI 300 index test set.The results show that in 857 trading days,the excess return of this strategy is 20.62% and the Sharpe ratio is 0.192,which further proves the validity of CLA model prediction.