Multi-Stock Movements Classification: A Decoupled Graph Diffusion Neural Network Approach

Forecasting future stock status and trends is still challenging for academia and industry. It is attributed to the complex and stochastic interactions between stocks and the hierarchical dynamics within individual stocks. Recently, graph neural networks have shown significant promise in tackling these problems by modeling and learning multiple stocks with graph-structured data. However, many existing approaches rely on manually defined factors to construct static stock graphs. This results in failing to capture the rapidly evolving interdependencies among stocks adequately. In addition, these methods often overlook hierarchical intra-stock features during message-passing. In this work, we propose a novel graph-learning framework that does not require prior domain knowledge to address these challenges. Our approach first generates dynamic stock graphs through entropy-driven edge generation from a signal processing view. Then, the task-relevant inter-stock dependencies are further refined with a generalized graph diffusion process on the constructed graphs. Lastly, a decoupled representation learning scheme is adopted to capture distinctive hierarchical intra-stock features. Our experimental results demonstrate substantial improvements over competitive baselines on three real-world datasets. Furthermore, the ablation study and sensitivity study validate its effectiveness in modeling the evolving inter-stock and intra-stock dynamics.