Stock trend prediction based on industry relationships driven hypergraph attention networks

In financial research, accurately predicting the movement trends of stock prices has been a focus for many researchers. The interrelationships among stocks are important factors that influence stock prices. However, recent research has revealed several limitations of traditional deep learning models in capturing these interrelationships, including the inability to learn higher-order relationships among stocks, the inability to dynamically update the relationship graph, and the failure to model the impact of industry relationships on individual stocks. To address these limitations, we propose an industry relationship-driven hypergraph attention network (IRD-HGAT) for predicting stock price movement trends. A key aspect of our work is the use of a hypergraph structure to represent the higher-order relationships among stock industries. The hypergraph attention mechanism is used to dynamically update the relationships between stocks, and the properties of industry hyperedges are aggregated to analyze the impact of industry relationships on stock prices. By comparing the current state-of-the-art algorithms, IRD-HGAT achieves excellent predictive performance and profitability on both S &P500 and CSi500 datasets, with AUC and Sharpe ratios of 0.87 and 1.12, respectively. Ablation experiments and parameter sensitivity analyses also further validate the validity and predictive stability of the model components.