Stock trend prediction based on dynamic hypergraph spatio-temporal network

Predicting stock trends is conducive to optimize returns from stock investments, which gains great interest from investors and researchers. Relations between stocks can provide important information for stock trend prediction. However existing stock prediction approaches only consider pairwise linkages, and ignore complex higher-order relations among stocks. To address these limitations, this paper proposes a dynamic hypergraph spatio-temporal network (DHSTN). DHSTN utilizes GRU to learn the sequential embedding of stocks, and a dynamic hypergraph network is proposed to learn the spatio-temporal relations among stocks. In the dynamic hypergraph network, firstly, a novel dynamic hypergraph construction module based on graph attention network is designed to capture stock higher-order spatial relations which are dynamically changing over time. Secondly, an industry relations aggregator based on hypergraph is considered in hypergraph convolution. Finally, a multi-relation fusion module is designed to integrate static and dynamic stock relations. Experiments on CSI300 and NASDAQ100 datasets show that DHSTN outperforms representative stock prediction methods by at least 4.99% in terms of F1-score and at least 47.9% in terms of sharpe ratio. •Constructing a dynamic hypergraph module for spatio-temporal stock embedding.•Designing an industry relations aggregator to capture the impact of stock industry.•Fusing static and dynamic stock relations to predict stock trend.