CSI 300 Stock Index Futures Price Prediction Study Based On ND-Haar Decomposition And DSALSTM Combined Model

Since its inception,financial futures have been an important part of the financial market.Its trading and price reflect the market’s expectations for relevant financial assets,which is beneficial for market participants to conduct risk management and investment decisions.Stock index futures are a type of financial futures,and their prices reflect the market’s future expectations for the relevant stock market,which is one of the important reference indicators for the stock market.With the development of the economy,stock index futures have become not only a tool for risk management,but also an investment and speculative attribute.China’s futures market has gone from nothing to being gradually becoming an important means of helping the development of the real economy and effectively managing price risks.It is precisely because of the leverage effect and two-way trading mechanism of stock index futures,and its constantly increasing status in China’s financial market,that more and more scholars have begun to study the price trends of futures.Based on this,this paper selects the closing price of the CSI 300 stock index futures for predictive analysis,with a data period from January 2011 to December 2022.Firstly,this paper combines previous research content,starting from linear and nonlinear methods,and selects data and indicators related to the closing price of stock index futures for data and indicator screening,leaving only the opening price,highest price and lowest price as related data for predicting the closing price sequence.Secondly,this paper introduces the LSTM(DSA-LSTM)based on the dual-attention mechanism model structure based on the encoder-decoder structure.The model improves on the original temporal attention by adding a spatial dimension attention mechanism.The structure of the dual-attention mechanism is conducive to the model adaptively adjusting the data weight in different time and space during the prediction process to obtain better predictive results.Afterwards,in order to further optimize and improve the model,this paper adopts the undecimated wavelet decomposition method to decompose the time series data to capture different cycles and more detailed signals in the closing price sequence.Based on the different data features obtained from the decomposition of low-frequency and high-frequency data,the ARIMA model is used to predict the linear and steady-state low-frequency data,and the DSA-LSTM model is used to predict the nonlinear high-frequency data with large fluctuations.By constructing the non-decomposition model and the decomposition-combination model,this paper further discusses the role of wavelet decomposition in prediction and the effectiveness of the decomposition-combination method in prediction.In empirical testing,the models are divided into non-decomposition models,decomposition non-combination models,and decomposition-combination models for comparison,and MAE,MAPE and RMSE are used as model performance evaluation indicators.The model results show that whether in the non-decomposition model,the decomposition non-combination model or the decomposition-combination model,the DSA-LSTM has the best performance in terms of RMSE,MAPE and MAE indicators;the decomposition model can achieve better performance compared to the non-decomposition model;and the decomposition-combination model,because it considers the characteristics of the high-frequency and low-frequency time series data after decomposition and selects different prediction methods for different time sequences,is the best predictive model among all comparison models,with RMSE,MAPE and MAE of 43.6,9922 and 0.53%,respectively.After that,in order to further consolidate the research conclusions of this paper,this paper first conducted rolling predictions on the CSI 300 stock index futures data starting from 2015,with a period of 5 years for each prediction.Secondly,the Zhongzheng 500 stock index futures data was used instead of the CSI 300 stock index futures data for prediction.The prediction results in both aspects confirm the effectiveness of the wavelet decomposition and DSA-LSTM model,and it is also demonstrated that the combination model with ARIMA performs best.