Dynamic asymmetric relational learning for stock price movement prediction

Stock movement forecasting remains a highly challenging task because existing deep learning models often struggle to capture the complex, dynamic, and asymmetric influence relationships among assets, leading to performance degradation when confronted with structural market changes. To overcome this limitation, this paper proposes AsymAlpha, an end-to-end, market-state-adaptive asset relationship learning framework. AsymAlpha is comprised of two core modules called Dynamic Asymmetric Relationship Module (DARM) and Market-Gated Predictor (MGP). The DARM employs an asymmetric attention mechanism, with a differentiable Directed Acyclic Graph (DAG) constraint providing structural guidance, to model intricate market information flow graph. Moreover, the MGP leverages context states derived from market snapshot data to help the forecasting process adaptively discern the nuances between various market environments. In this way, AsymAlpha can effectively capture the complex influence dynamics in evolving online financial markets. Extensive experiments conducted on benchmark datasets covering four major international markets demonstrate that AsymAlpha significantly outperforms state-of-the-art baselines in both prediction accuracy and simulated trading performance.