Answers are grounded in this paper and cite their sources.
Identifying and forecasting systemic risk, the threat that distress at individual financial institutions cascades into network-wide failure, is one of the central challenges in macroprudential supervision, with direct implications for regulatory intervention, capital-buffer policy, and financial stability. Graph-based models have emerged as a leading approach for this task by explicitly representing interbank exposure networks, yet they typically encode correlational rather than causal structure, which limits their reliability under distributional shift. To address these challenges, we present Causal Spatio-Temporal Graph Learning (CSTGL), a framework that integrates a PC-algorithm-based Causal Graph Builder, a Spatial Causal GCN (SC-GCN) that propagates risk signals over dual causal and temporal adjacencies, a Temporal Attention Transformer (TAT) for long-range dependencies, and a do-calculus-based Counterfactual Intervention Layer (CIL) for regulatory simulation. The SC-GCN, TAT, and CIL modules are jointly optimised end-to-end; the Causal Graph Builder runs as a pre-processing step and provides a fixed directed graph structure to the learnable components. On the FDIC BankFind financial network benchmark, CSTGL reaches 85.7% F1, 89.1% AUC-ROC, and 81.3% contagion-path accuracy, improving on the strongest baseline by 5.6, 6.2, and 7.4 points respectively, and retains 94.3% of its in-distribution performance under cross-market stress testing, compared with 77.5% for the best competing method (Transformer). Precision on rare regulatory-breach events is 72.4%, pointing to class imbalance as an open direction. Taken together, these results position CSTGL as a practical basis for real-time systemic-risk monitoring and for causally grounded intervention policy design.
No takes yet. Share an insight, caveat, or question.
sui et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: