Logo image
Back
HyperGAT: Hypergraph Attention Network for Stock Movement Prediction
Conference paper

HyperGAT: Hypergraph Attention Network for Stock Movement Prediction

Jianliang Gao, Shilin Xie, Shujin Wang and Xiaohua Tony Hu
IEEE International Conference on Big Data, (2025), pp 1389-1395
08 Dec 2025
DOI: https://doi.org/10.1109/BigData66926.2025.11400976

Abstract

Accuracy Big Data applications Prediction methods Probability distribution Profitability Market Research
As a typical financial big data application, stock movement prediction can be formulated as a classification problem to predict the future movement of stock prices, e.g., the rising, falling, or steady trends. Existing methods rely on the price movement signals of a single stock or a pair of stocks, ignoring the higher-order relationships among multiple stocks. In real markets, a group of stocks are linked by higher-order relationships, such as belonging to the same stock sector or being held by the same fund. This paper proposes the Hypergraph Attention Network for Stock Movement rediction (HyperGAT) to explore higher-order relationships among stocks. In HyperGAT, two heterogeneous hypergraphs are designed to characterize stocks' sector-belonging and fund-holding relationships. Hypergraph hierarchical attention spatially models the higherorder relationships of stocks by considering the importance of different stock nodes, hyperedges, and hypergraphs. Temporal attention captures the temporal dynamics of stock and sector sequences while distinguishing the impact of historical states. Sector rotation attention focuses on stocks that are more likely to rise. Finally, the sectoraware fine-tuning layer (SFL) fine-tunes the probability distribution of future stock movements to improve stock movement prediction. Experiments on the real-world dataset show that HyperGAT outperforms competing approaches in terms of accuracy and profitability.

Metrics

1 Record Views

Details

Logo image