DySTAGE: Dynamic Graph Representation Learning for Asset Pricing via Spatio-Temporal Attention and Graph Encodings

Current GNN-based asset price prediction models often focus on a fixed group of assets and their static relationships within the financial network. However, this approach overlooks the reality that the composition of asset pools and their interrelationships evolves over time, necessitating a flexible framework capable of adapting to this dynamism. Accordingly, we propose DySTAGE, a framework with a universal formulation that transforms asset pricing time series into dynamic graphs, accommodating asset addition, deletion, and changes in correlations. Our framework includes a graph learning model specifically designed for this purpose: assets at various historical time steps are structured as a sequence of dynamic graphs, where connections between assets reflect their long-term correlations. DySTAGE effectively captures both topological and temporal patterns via Asset Influence Attention, Asset-wise Importance Encoding, Pair-wise Spatial Encoding, and Edge-wise Correlation Encoding, together with a Temporal Module. We validate our approach through extensive experiments using three real-world stock pricing datasets, demonstrating that DySTAGE surpasses popular benchmarks in return prediction and offers profitable investment strategies.

Code: github.com/JunyiYe/DySTAGE

Recommended citation: Gu*, J., Ye*, J., Uddin, A., & Wang, G. (2024). "DySTAGE: Dynamic Graph Representation Learning for Asset Pricing via Spatio-Temporal Attention and Graph Encodings." Proceedings of the 5th ACM International Conference on AI in Finance (ICAIF '24), 388–396.
Download Paper