A stock series prediction model based on variational mode decomposition and dual-channel attention network

Abstract

Due to the comprehensive impact of external factors (politics, economy, market, etc.) and internal factors (organizational structure, management ability, innovation capability, etc.), stock series exhibit strong volatility. Coupled with their inherent high liquidity, it poses great challenges for stock series prediction. However, the previous stock series forecasting methods often only pay attention to the long-term dependencies, and lack attention to the local features and short-term dependencies. In this regard a stock series prediction model based on variational mode decomposition and dual-channel attention network is proposed, which is called VMD-LSTMA+TCNA. To prevent information leakage, the stock series is divided into equal-length sub-windows by sliding window. To reduce the series volatility, each sub-window is decomposed into different frequency mode sub-windows through variational mode decomposition (VMD). To improve the prediction accuracy and robustness in different stock markets, we construct a dual-channel attention model called LSTMA+TCNA. The LSTMA channel is used to extract long-term dependencies and temporally correlated features, while the TCNA channel is used to extract local patterns and short-term dependencies, and self-attention is added to both channels to increase the weight of features at important times. Predict each frequency mode sub-window separately through the specific LSTMA and TCNA channels, and then obtain the predicted values by fusing the results of dual-channel. The final predicted stock series is obtained by superimposing the predicted values of each frequency mode sub-window. Through extensive experiments on the US and Hong Kong stock markets, it has been shown that the VMD-LSTMA+TCNA model exhibits better robustness and generalization compared to other state-of-the-art methods and has higher prediction accuracy.