Adaptive dynamic sub-hypergraph learning with gated temporal convolution for multi-factor financial trend forecasting

Abstract Financial markets exhibit complex high-order dependencies and dynamic temporal evolution, making accurate trend forecasting under multi-factor influences a challenging task. To alleviate this problem, we propose a dynamic sub-hypergraph network integrated with lightweight gated temporal convolution (DSHN-TC), which models evolving market structures through adaptive hypergraph learning while efficiently capturing temporal dependencies. Specifically, a temporal pattern encoding mechanism is introduced to enhance local temporal representations, and convolutional networks are employed to extract node-level features. An attention-enhanced hypergraph convolution with adaptive hyperedge refinement is designed to model dynamic high-order interactions. Meanwhile, a channel-specific aggregation strategy is adopted to learn expressive node embeddings from heterogeneous factors. Furthermore, a cross-channel fusion module integrates multi-source information into unified representations, followed by a lightweight gated temporal convolution to capture temporal dependencies with reduced computational cost. Experimental results on multiple financial markets demonstrate that DSHN-TC consistently outperforms state-of-the-art baselines in both classification and financial metrics, highlighting its effectiveness and potential for practical investment applications.