Research map / Graph learning / Knowledge and temporal graphs
MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MDGNN, a discrete dynamic multi-relational graph model that combines meta-path hierarchical graph attention on daily stock/bank/industry graphs with a Transformer (ALiBi bias, causal mask) over a look-back window to predict stock movement on CSI100/CSI300.
Problem
Predicting stock price movement is hard because stock relations are multifaceted (industries, investment banks, stock pairs) and evolve over time, which sequential and single-relation graph models fail to capture.
Approach
Build a daily multi-relational graph of stocks, investment banks and industries as discrete snapshots; aggregate per-meta-path with multi-head graph attention, fuse meta-paths adaptively, then encode the node's snapshot sequence with a causal Transformer using ALiBi relative position bias.
Key results
- On CSI300, MDGNN reports IC 0.0322, IR 0.2488, CR 0.9828 and Prec@30 0.5232, the best among compared methods (HTGNN IC 0.0192).
- On CSI100, MDGNN reports IC 0.0123, IR 0.0746, CR 0.2741 and Prec@30 0.5081, slightly above HTGNN and HGT on most metrics.
- Ablation on CSI300 shows removing the meta-path module causes the largest drop (IC 0.0322 to 0.0216).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add edge weight (ablation) | MDGNN · CSI300 | IC: 0.0268 -> 0.0322 | Helped |
| add meta-path module (ablation) | MDGNN · CSI300 | IC: 0.0216 -> 0.0322 | Helped |
| add hierarchical aggregation (ablation) | MDGNN · CSI300 | IC: 0.0303 -> 0.0322 | Helped |
| add temporal extraction layer (ablation) | MDGNN · CSI300 | IC: 0.0286 -> 0.0322 | Helped |
| add stock-investment-bank edges (SB) to the stock-stock graph (ablation) | MDGNN · CSI300 | IC: 0.0217 -> 0.0264 | Helped |
| add stock-industry edges (SI) to the stock-stock graph (ablation) | MDGNN · CSI300 | IC: 0.0217 -> 0.0210 | Hurt |
| add industry-industry edges (II) with SI edges (ablation) | MDGNN · CSI300 | IC: 0.0210 -> 0.0217 | Helped |
| add stock-investment-bank, stock-industry, and industry-industry relations (ablation) | MDGNN · CSI300 | IC: 0.0217 -> 0.0322 | Helped |
| increase window size | MDGNN · CSI300 | CR | Mixed |
| increase the number of GNN layers | MDGNN · CSI300 | CR | Mixed |
Surprising findings
- Homogeneous GAT and GCN outperform the Transformer baseline on the larger CSI300 dataset despite weaker performance on CSI100.
Limitations
- Evaluated only on Chinese stock indices (CSI100/CSI300) with a single backtest period.
- Relation construction relies on proprietary-style financial data from TuShare and lexicon/syntactic extraction, limiting reproducibility.
- Reported variance on several metrics is large relative to improvements, e.g. CSI100 CR and IR.
Experimental setup
Datasets: CSI100 constituents data CSI300 constituents data TuShare daily trading, macro reports, financial news, financial statements, research reports
Benchmarks: CSI100 CSI300
Metrics: IC IR CR Precision@30
Baselines: MLP LSTM Transformer GAT GCN RGCN HAN HGT EvolveGCN HTGNN
Builds on: HTGNN EvolveGCN ALiBi Graph Attention Networks
Closest papers by meaning
- ECHO-GL: Earnings Calls-Driven Heterogeneous Graph Learning for Stock Movement Prediction (AAAI 2024 · similarity 0.86)
- Temporal-Aware Evaluation and Learning for Temporal Graph Neural Networks (AAAI 2025 · similarity 0.86)
- Dynamic Graph Learning with Static Relations for Credit Risk Assessment (AAAI 2025 · similarity 0.85)
- MASTER: Market-Guided Stock Transformer for Stock Price Forecasting (AAAI 2024 · similarity 0.85)
- PRES: Toward Scalable Memory-Based Dynamic Graph Neural Networks (ICLR 2024 · similarity 0.85)
- CI-STHPAN: Pre-trained Attention Network for Stock Selection with Channel-Independent Spatio-Temporal Hypergraph (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.