GRU-PFG: Extract Inter-Stock Correlation from Stock Factors with Graph Neural Network
GRU-PFG: Extract Inter-Stock Correlation from Stock Factors with Graph Neural Network
The paper introduces GRU-PFG, a stock trend prediction model that leverages graph neural networks to extract inter-stock correlations purely from Alpha360 stock factors, without requiring additional data sources like industry information or financial reports. The model consists of three stages: preliminary extraction via GRU, primary relationship extraction using softmax and Pearson correlation to capture explicit stock correlations, and secondary relationship extraction to uncover latent correlations. The final prediction combines these features with learnable weights. Experiments on CSI100 and CSI300 datasets show that GRU-PFG achieves an IC of 0.134 on CSI300, outperforming both single-source models (e.g., GRU, ALSTM) and multi-source models (e.g., HIST with IC 0.131). The model demonstrates that inter-stock correlations can be effectively inferred from stock factors alone, offering a more standardized and generalizable approach to stock prediction.
Highlights
- 1Proposes GRU-PFG, a novel stock prediction model that extracts inter-stock correlations solely from Alpha360 stock factors using graph neural networks.
- 2Achieves superior performance compared to models relying only on stock factors (e.g., GRU, ALSTM) and matches multi-source models (e.g., HIST) on CSI300 dataset.
- 3Introduces a two-stage relationship extraction mechanism (primary and secondary) to capture both explicit and latent correlations between stocks.
- 4Demonstrates greater generalization potential by avoiding subjective and hard-to-standardize additional data sources.
Methods
- MGRU for preliminary information extraction from Alpha360 factors (60 time steps, each with 6 factors).
- MPrimary relationship extraction using softmax and Pearson correlation to compute inter-stock and intra-stock correlation matrices.
- MSecondary relationship extraction by subtracting primary effects and applying similar correlation operations to capture residual latent relationships.
- MFinal stock feature combines original GRU output, primary, and secondary features via learnable weights, followed by linear prediction.
Results
- ROn CSI300 dataset, GRU-PFG achieves IC of 0.134, outperforming HIST (0.131) and significantly surpassing GRU and Transformer.
- ROutperforms all single-source models (MLP, ALSTM, GRU, SFM, GATs, Transformer) on both CSI100 and CSI300 across IC, Rank IC, and Precision@N metrics.
- RMatches or exceeds multi-source models (HIST, TRA) while using only stock factors, indicating strong generalization.
- RThe two-stage relationship extraction consistently improves prediction accuracy over single-stage or no graph components.
Analyze Paper
Generate insights from "GRU-PFG: Extract Inter-Stock Correlation from Stock Factors ...".