Abstract
FactorGCL is a novel hypergraph-based model designed to enhance stock return prediction by integrating human expert knowledge with automated "hidden factor" discovery. Utilizing a cascading architecture to isolate residuals and a temporal contrastive learning objective, the model effectively filters market noise to identify robust, non-linear patterns often missed by traditional models.
Introduction: Addressing the Signal-to-Noise Challenge
Predicting stock returns is notoriously difficult due to the low signal-to-noise ratio in financial markets. Traditional factor models, such as the Fama-French model, rely on human-designed factors like company size or industry. While theoretically grounded, these linear models are often too simplistic for modern market complexities. Conversely, pure Machine Learning (ML) models excel at finding non-linear patterns but frequently overfit to transient market noise. FactorGCL addresses this by leveraging expert knowledge while discovering new, effective factors through a noise-resistant framework.
Methodology: The FactorGCL Framework
The FactorGCL (Factor model with Graph Contrastive Learning) framework is built upon three core innovations:
1. Cascading Residual Hypergraph Architecture
Rather than a direct prediction, the model decomposes returns into three distinct layers to isolate specific signals:
- Prior Beta: Uses a Hypergraph Convolutional Network (HyperGCN) to capture the influence of known, expert-designed factors (e.g., industry sectors).
- Hidden Beta: Analyzes the residual signal after subtracting prior factors to generate "soft" hyperedges representing data-driven hidden factors.
- Individual Alpha: Captures idiosyncratic, stock-specific information remaining after accounting for both prior and hidden factors.
2. Hypergraph Representation
FactorGCL utilizes hypergraphs, where "hyperedges" connect groups of nodes (stocks). This structure naturally models financial factors (e.g., "High Volatility") that affect multiple stocks simultaneously, capturing high-order, non-linear relationships.
3. Temporal Residual Contrastive Learning (TRCL)
To ensure hidden factors are meaningful, a self-supervised contrastive loss is applied. It encourages the "alpha" embeddings of a stock to remain similar across different time periods while distinguishing them from other stocks. This forces the model to ignore transient noise and focus on stable, stock-specific features.
Experimental Setup
The model was evaluated on the China A-shares market (January 2014 – June 2023) using a universe of 5,028 stocks. Input features included day-level price-volume data, with 83 secondary industry classifications serving as prior factors. Labels were cross-sectional standardized future returns for horizons of $\Delta t = 1, 5, 10, 20$ days.
Key Results and Performance
FactorGCL was benchmarked against standard RNNs, Transformers, and specialized models like FactorVAE and HIST.
- Predictive Accuracy: FactorGCL achieved the highest Information Coefficient (IC) across all horizons. For a 5-day horizon ($\Delta t=5$), it reached an IC of 0.0885, outperforming the ALSTM baseline (0.0813).
- Investment Profitability: In a "TopK" simulation (top 30 stocks), FactorGCL achieved an Annualized Return (AR) of 16.9% on the CSI300 index, nearly doubling many ML baselines.
- Ablation Studies: Removing either the contrastive learning or hidden factor modules led to significant performance drops, validating the integrated architecture.
Discussion and Future Directions
The "cascading" approach acts as a powerful regularizer by prioritizing known factors, forcing hidden factors to capture truly novel information. While effective, hypergraph operations can be computationally intensive. Future work may involve applying this framework to risk factor mining and more sophisticated portfolio optimization strategies.
Summary updated