Heterogeneous Dual-Dynamic Attention Network for Modeling Mutual Interplay of Stocks
Haotian Liu,Yadong Zhou,Yuxun Zhou,Bowen Hu
DOI: https://doi.org/10.1109/tai.2024.3374269
2024-01-01
IEEE Transactions on Artificial Intelligence
Abstract:Modern quantitative finance and portfolio-based investment hinge on the dependence structure among financial instruments (like stocks) for return prediction, risk management, and hedging. Several attempts have been made to leverage recent advancements in graph neural networks (GNNs) to achieve effective modeling to capture stock interactions. However, existing work suffers from the homogeneity and stationary of the predefined graph structure, and their reliance on hard-coded correlation-based relationships neglects the mutual intervention of different instruments. In this study, we propose the heterogeneous dual-dynamic attention network (HD 2 AT) to address these challenges in modeling the mutual interplay among stocks, across four key dimensions: heterogeneity, causality, structural dynamics, and lead-lag dynamics. More specifically, our method uncovers the causal interdependencies among instruments, dynamically inferring the heterogeneous network structure from real-time trading activities. Subsequently, we identify the cross-temporal propagation of interventions by integrating interactions at the node, time, and graph levels. Extensive experiments on real-world trading data demonstrate the superiority of the proposed framework over various baselines in terms of both return prediction and back-testing profitability.