The paper introduces MDGNN, a novel framework for stock investment prediction that addresses two key challenges: the multifaceted nature of stock relations and their temporal evolution. MDGNN constructs a discrete dynamic graph where each daily snapshot is a multi-relational graph containing nodes for stocks, industries, and investment banks, with edges representing diverse relations such as industry membership, investment bank activities, and stock-stock correlations. A hierarchical multi-relational graph embedding layer uses attention mechanisms to aggregate information from multiple meta-paths, capturing both local and global dependencies. To model temporal dynamics, a Transformer-based layer with relative position bias (ALIBI) encodes the evolution of node representations across time, emphasizing recent events. The model is trained end-to-end to predict stock returns relative to a benchmark index.

Extensive experiments on CSI100 and CSI300 datasets demonstrate that MDGNN consistently outperforms state-of-the-art methods, including traditional time series models (MLP, LSTM, Transformer), static graph neural networks (GCN, GAT), heterogeneous graph models (RGCN, HAN, HGT), and dynamic graph models (EvolveGCN, HTGNN). The improvements are particularly notable on the larger CSI300 dataset, where MDGNN achieves an IC of 0.0322, IR of 0.2488, CR of 0.9828, and Precision@30 of 0.5232. Ablation studies confirm the importance of each relation type and the temporal component. The work provides a comprehensive solution for modeling complex, evolving stock market relationships, offering both predictive performance and interpretability through attention weights.