Temporal Hypergraph Attentive Transformer for regional financial risk prediction: Integrating risk contagion dynamics and temporal pattern mining

The early prediction of regional financial systemic risk demands sophisticated modeling of cross-market contagion dynamics and multidimensional policy interactions. Traditional graph neural networks (GNNs) suffer from limited expressiveness in capturing higher-order dependencies among heterogeneous financial entities (banks, firms, regulators). This paper proposes a Temporal Hypergraph Attentive Transformer (T-HGAT) that synergizes risk contagion dynamics with temporal policy shock propagation. The key innovation lies in two aspects: (1) a hyperedge-aware message passing mechanism that models multi-body interactions through LSTM-enhanced hyperedge evolution, allowing adaptive risk diffusion across policy-sensitive channels; (2) a differential contagion intensity estimator that quantifies time-varying risk exposure via gradient flow analysis on the Lie algebra manifold. Validated on China’s interconnected regional financial datasets, our framework achieves 89.2 % F1-score in 3-month-ahead crisis alert, outperforming State-of-the-Art (SOTA) approaches like CoRiskNet (+18.6 %) and DHCF (+24.3 %). Crucially, the VaR-backtesting shows maximum drawdown reduction of 37.4 % compared to conventional stress testing models. T-HGAT has been deployed in Jiangsu Province’s Financial Stability Monitoring System, demonstrating capacity to detect emergent risks 12 weeks earlier than human analysts during the 2022 local government financing vehicle (LGFV) debt crisis.