Research On Stock Price Index Prediction Based On Time Series And Deep Learning Model

Stock price index is an important economic indicator reflecting the overall condition of the stock market.Accurate prediction of stock price index fluctuations is an important tool to prevent risks and guarantee the stable development of the financial market.China’s stock industry is a strong noise market,and its fluctuations are susceptible to the combined effects of rumors,domestic and international economic environment,etc.The stock data show the characteristics of co-existence of linear and non-linear relationships.In this paper,we adopt the combination of time series model and deep learning to carry out the research of stock price index prediction.The two models complement each other,the time series model can well capture the statistical characteristics of the data and fit the linear part of the data,while the deep learning model can well deal with the nonlinear part,and the two models are integrated in order to obtain more accurate results.The main innovations and research results are as follows.(1)Comparison model experiments.In this paper,we further improve the considerations in the process of stock forecasting,introduce the differences in forecasting methods,and comprehensively compare the forecasting effects of time series models and deep learning models on different forecasting methods.From the perspective of dynamic forecasting and static forecasting,ARIMA model,ARIMA-GARCH model,LSTM model,and GAN model are constructed to compare the forecasting effects of the two types of models based on SSE Composite Index stock price data.The experimental results show that the ARIMA-GARCH model has the best average prediction accuracy in static forecasting with an RMSE value of 21.66%,while the deep learning model represented by LSTM has the most accurate prediction effect in dynamic forecasting with an RMSE value of only 6.32%,which indicates that the forecasting method affects the prediction effect of the model.(2)Deep learning model optimization.In the comparison experiment,it is found that the deep learning model is not stable in static prediction,based on this finding,this paper optimizes the deep learning model,and selects the equal-weight combination method and the dominant combination method in the linear combination method to optimize the GAN and LSTM models,and it is found that the prediction accuracy is significantly improved after optimization,and the model optimization method is effective.(3)Constructing a combined model ARIMA-Bi LSTM to forecast stock price index.based on the advantages and disadvantages of the two types of models,this paper proposes a new combined model ARIMA-Bi LSTM,which is mainly based on the idea of residual optimization and makes full use of the information contained in the data.The results show that the combined model can avoid the defects brought by a single model,and the ARIMA-Bi LSTM model has higher prediction accuracy than other advanced methods.Finally,ablation experiments are conducted on the constructed models to study the rationality of model parameter settings,and the effectiveness of the combined model is further verified based on different data sets.