Research map / Label-efficient and robust learning / Recognition and long-tail
FactorGCL: A Hypergraph-Based Factor Model with Temporal Residual Contrastive Learning for Stock Returns 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 FactorGCL, a hypergraph-based nonlinear factor model that mines hidden factors from residuals after prior industry factors and regularizes them with temporal residual contrastive learning, evaluated on China A-share stock return prediction.
Problem
Predicting stock returns with factor models is hard due to low signal-to-noise ratio; data-driven factors overfit noise and fail to supplement human-designed prior factors comprehensively.
Approach
A cascading residual hypergraph architecture (prior beta, then hidden beta from residuals via learnable soft hyperedge prototypes, then individual alpha) trained with an InfoNCE loss that contrasts each stock's alpha embeddings across past and future time windows.
Key results
- FactorGCL achieves IC 0.0684 / ICIR 0.7487 at Δt=1 and IC 0.0929 / ICIR 1.0350 at Δt=20 on China A-shares test set, the best among 12 compared methods.
- In the TopK=30 investment simulation on CSI300, FactorGCL reaches AR 0.169, IR 4.208, RoMaD 2.978, versus best baseline SFM AR 0.095, IR 2.913.
- Ablation removing the contrastive loss lowers IC at Δt=1 from 0.0684 to 0.0661; removing the prior module drops it to 0.0654.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FactorGCL instead of the strongest baseline | China A-shares stock return prediction · China A-shares test set, Δt=1 | IC: 0.0646 -> 0.0684 | Helped |
| Use FactorGCL instead of the strongest baseline | China A-shares stock return prediction · China A-shares test set, Δt=5 | IC: 0.0813 -> 0.0885 | Helped |
| Use FactorGCL instead of the strongest baseline | China A-shares stock return prediction · China A-shares test set, Δt=10 | IC: 0.0829 -> 0.0915 | Helped |
| Use FactorGCL instead of the strongest baseline | China A-shares stock return prediction · China A-shares test set, Δt=20 | IC: 0.0861 -> 0.0929 | Helped |
| Add the prior beta module (ablation) | FactorGCL · China A-shares test set, Δt=10 | IC: 0.0904 -> 0.0915 | Helped |
| Add the hidden beta module (ablation) | FactorGCL · China A-shares test set, Δt=10 | IC: 0.0828 -> 0.0915 | Helped |
| Add the individual alpha module and contrastive learning loss (ablation) | FactorGCL · China A-shares test set, Δt=10 | IC: 0.0857 -> 0.0915 | Helped |
| Add the contrastive learning loss (ablation) | FactorGCL · China A-shares test set, Δt=10 | IC: 0.0899 -> 0.0915 | Helped |
| Increase the number of hidden factors, then continue increasing beyond the best-performing point (ablation) | FactorGCL hidden beta module · China A-shares stock return prediction | IC | Mixed |
Surprising findings
- Removing the prior beta module barely changes IC at longer horizons (Δt=20: 0.0906 vs 0.0929 full), and some baselines like FactorVAE have higher IC than FactorGCL at Δt=5 range is not the case; FactorVAE's ICIR falls to 0.6526 at Δt=20.
- Many hypergraph and graph baselines (GAT, HyperGCN, CI-STHPAN) underperform simple GRU/ALSTM baselines on CSI500 investment simulation.
Limitations
- Relies on given human-designed prior factors (secondary industry exposures) as the only prior knowledge.
- Evaluated only on China A-share market data with one train/valid/test split scheme.
- Performance advantage in IC is modest (e.g., 0.0684 vs 0.0646 for FactorVAE/ALSTM at Δt=1).
Experimental setup
Datasets: China A-shares daily price-volume data (2014-2023) CSI300 CSI500
Benchmarks: China A-shares stock return prediction (IC, ICIR) CSI300 investment simulation CSI500 investment simulation
Metrics: IC ICIR AR IR RoMaD
Baselines: MLP GRU TCN Transformer ALSTM SFM GAT HIST HyperGCN STHAN-SR FactorVAE CI-STHPAN
Builds on: HyperGCN InfoNCE FactorVAE HIST
Closest papers by meaning
- CI-STHPAN: Pre-trained Attention Network for Stock Selection with Channel-Independent Spatio-Temporal Hypergraph (AAAI 2024 · similarity 0.86)
- AlphaForge: A Framework to Mine and Dynamically Combine Formulaic Alpha Factors (AAAI 2025 · similarity 0.85)
- AlphaPROBE: Alpha Mining via principled retriveval and on-graph biased evolution (NeurIPS 2026 · similarity 0.84)
- MASTER: Market-Guided Stock Transformer for Stock Price Forecasting (AAAI 2024 · similarity 0.83)
- DHMoE: Diffusion Generated Hierarchical Multi-Granular Expertise for Stock Prediction (AAAI 2025 · similarity 0.83)
- MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction (AAAI 2024 · similarity 0.83)
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.