Dual-Graph Neural Network with Polarity-Aware Attention for Multi-stock Volatility Forecasting

Accurate stock volatility forecasting is essential for risk management, option pricing, and portfolio optimization. Most existing methods, from econometric models like GARCH and HAR-RV to deep learning architectures, treat each stock independently, neglecting inter-stock dependencies driven by industry co-movement, supply-chain linkages, and competitive dynamics. Recent GNN-based approaches attempt to model such dependencies but rely on static, predefined graph structures that cannot adapt to evolving market conditions. We propose a dual-graph neural network framework for multi-stock collaborative volatility prediction with three key components: (1) a HAR-inspired multi-scale GRU module that maintains hidden states updated at daily, weekly, and monthly frequencies to capture heterogeneous temporal patterns; (2) an adaptive graph learning mechanism that combines learnable static node embeddings with dynamic input features to generate directed, signed adjacency matrices, discovering asymmetric and inhibitory inter-stock relationships end-to-end; and (3) a polarity-aware graph attention network that decomposes attention weights into importance and polarity factors, enabling message passing over signed edges without distorting negative relationships. The framework integrates a hierarchical industry graph with the adaptively learned stock-level relation graph through complementary message-passing pathways. Experiments on CSI 300 constituents over eight years (2018–2025) demonstrate consistent improvements over econometric, machine learning, and deep learning baselines across multiple forecasting horizons. Ablation studies confirm the contribution of each proposed component.