VMD-MSANet: A multi-scale attention network for stock series prediction with Variational Mode Decomposition

Abstract

The stock market is a crucial component of the financial system. Accurate prediction of its price is essential for effective risk management and informed investment decision-making. However, the complex dynamics of the stock market, including multi-scale non-stationarity and complex stock-market interactions, pose significant challenges for prediction. To address these challenges, we introduce VMD-MSANet, a stock price prediction model that combines Variational Mode Decomposition (VMD) with a multi-scale attention mechanism. We employ VMD to decompose the stock price series into sub-components with distinct frequency components, and use the multi-scale attention mechanism to capture both short-term and long-term temporal patterns effectively. By incorporating external market factors, the model enhances its comprehensive understanding and adaptability to the market environment. Extensive experiments on the Chinese market demonstrate that VMD-MSANet achieves higher predictive accuracy and exhibits enhanced robustness and generalization compared to existing state-of-the-art methods.