Spatiotemporal Fusion for Stock Prediction via Hypergraph Attention Gated Recurrent Units

Stock prediction requires the joint modeling of temporal dynamics and cross-stock dependence. Existing graph-based and hypergraph-based forecasting methods often process spatial relation modeling and temporal evolution in separate stages, which may weaken the interaction between relational information and recurrent state updating. This study proposes a Recurrent Spatiotemporal Hypergraph Attention Gated Recurrent Unit model for stock forecasting, in which hypergraph-based higher order dependence and temporal dynamics are integrated within each recurrent update. The hypergraph is constructed offline from heterogeneous financial features through Tucker decomposition, similarity estimation, and Top-K sparsification, and is then used as a structured relational prior during forecasting. Experiments on CSI 300 constituent stocks from January 2014 to October 2024 show that RST-HGA-GRU achieves the best overall performance across multiple evaluation metrics and forecasting horizons from 1 to 6 days. Ablation, sensitivity, back testing, and multi-horizon Diebold-Mariano tests further support the effectiveness and robustness of the proposed framework. These results demonstrate that recurrent spatiotemporal fusion with hypergraph-based higher-order relation modeling is effective for stock price forecasting.