Spatiotemporal Fusion for Stock Prediction via Hypergraph Attention Gated Recurrent Units.

Area of Science:

• Computational Finance
• Data Science
• Machine Learning

Background:

• Stock prediction necessitates modeling both temporal dynamics and inter-stock relationships.
• Current methods often separate spatial and temporal analysis, limiting interaction.
• Hypergraph methods capture higher-order dependencies but may not fully integrate with temporal models.

Purpose of the Study:

• To propose a novel Recurrent Spatiotemporal Hypergraph Attention Gated Recurrent Unit (RST-HGA-GRU) model.
• To integrate higher-order financial dependencies and temporal dynamics within a single recurrent update.
• To enhance stock forecasting accuracy by fusing spatiotemporal information.

Main Methods:

• Constructing a hypergraph offline using Tucker decomposition, similarity estimation, and Top-K sparsification on heterogeneous financial features.
• Integrating the hypergraph as a structured relational prior into a Gated Recurrent Unit (GRU) framework.
• Employing a spatiotemporal attention mechanism within the recurrent updates.

Main Results:

• The RST-HGA-GRU model demonstrated superior performance on CSI 300 constituent stocks across various metrics and forecasting horizons (1-6 days).
• Ablation studies, sensitivity analyses, and backtesting confirmed the model's effectiveness and robustness.
• Multi-horizon Diebold-Mariano tests validated the framework's predictive superiority.

Conclusions:

• Recurrent spatiotemporal fusion combined with hypergraph-based higher-order relation modeling significantly improves stock price forecasting.
• The proposed RST-HGA-GRU framework offers a robust and effective approach for financial time-series prediction.
• Integrating relational priors directly into recurrent updates enhances the modeling of complex financial market dynamics.