Predicting the highest and lowest stock price indices: A combined BiLSTM-SAM-TCN deep learning model based on re-decomposition

Accurate prediction of stock price indices is crucial for market participants to obtain valuable information and mitigate risks. For more accurate forecasting of stock price indices, this study proposes a deep learning combined prediction model based on re-decomposition, utilizing the daily highest price series and the daily lowest price series of the Standard & Poor's 500 Index (S&P 500) and the Shanghai Stock Exchange Composite Index (SSEC) as experimental data. This model involves several key steps. Initially, an improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is used to decompose the stock's highest price sequence or lowest price sequence into multiple subsequences. Next, the highest frequency subsequence is re-decomposed using variational mode decomposition (VMD) optimized by particle swarm optimization (PSO). Then, bidirectional long short-term memory (BiLSTM), self-attention mechanism (SAM) and temporal convolutional network (TCN) are combined to construct BiLSTM-SAM-TCN for predicting each subsequence independently. Finally, the predicted values of the subsequences are linearly integrated to get the prediction results of the highest price sequence or the lowest price sequence of the stock. The empirical analyses demonstrate that the proposed ICEEMDAN-PSO-VMD-BiLSTM-SAM-TCN model exhibits excellent forecasting performance in both developed and developing country stock markets, surpassing other models. The modified Diebold-Mariano (MDM) test also statistically confirms the superiority of the proposed model. Furthermore, the model is also verified to have good robustness and generalization ability by changing the time scale of the original stock price series and then forecasting it, as well as forecasting the stock price under extreme market conditions. Overall, this study presents a reliable method for accurately predicting the highest and lowest stock price indices with the potential for practical application in diverse stock markets. •Proposing a combined model based on re-decomposition for stock index forecasting.•Predicting the highest and lowest stock indices to better manage market risk.•The model outperforms other models in both developed and developing markets.•Modified Diebold-Mariano test statistically proves the superiority of the model.•Robustness tests show the model has good adaptability under different conditions.