Two-Stage Deep Ensemble Paradigm Based on Optimal Multi-scale Decomposition and Multi-factor Analysis for Stock Price Prediction

Stock price forecasting is important for financial risk management and investment decisions. However, traditional forecasting techniques are challenged and under pressure due to the complex characteristics of stock prices and the impact of quantitative trading. Therefore, to address these issues and produce a more accurate stock price forecasting method, this study proposes a two-stage deep integration paradigm based on optimal multi-scale decomposition and multi-factor analysis. This paradigm will also serve as a scientific support and reference for investors’ actual investment decisions. (1) Optimal multi-scale decomposition methods are proposed to achieve decomposition of the closing price. Noise suppression and optimal decomposition modes are achieved by singular spectrum analysis (SSA) and variable mode decomposition (VMD) to reconstruct the high- and low-frequency sub-series. (2) The approximate entropy is used to measure the complexity of the time series, and the approximate entropy of the decomposed sequence is judged to reconstruct the characteristic subsequence. (3) The multi-factor analysis method uses a joint feature selection technique, the spearman correlation coefficient test, and the mutual information (MI) index to find the optimal feature influence factor and extract the optimal influence factor using feature compression. (4) Two-stage deep integration was performed using bidirectional gate recurrent unit (BIGRU) to obtain the final predictions. The two major indices of the Chinese A-share market, namely the Shenzhen Stock Index (SZI) and Shanghai Stock Exchange (SSEC), with data sourced from the Flush Finance website, were used to verify the validity of the model proposed in this paper. The MAE, RMSE, and MAPE values of the proposed model in this paper are 171.1673, 221.2086, and 1.2796% in SZI market and 29.9973, 39.3508, and 0.8946% in SSEC market, respectively, which are significantly better than other comparative models. Based on the experiments, it can be seen that the optimal multi-scale decomposition method performs better in capturing the hidden information of the original data, while the multi-factor analysis provides more relevant data for the prediction, which helps to improve the accuracy of the prediction. After integrating the predictions, the decomposition-integrated hybrid model proposed in this paper predicts significantly better than with other prediction models.