Explainable Graph Neural Networks for Interbank Contagion Surveillance: A Regulatory-Aligned Framework for the U.S. Banking Sector

Researcher Mohammad Nasir Uddin has introduced a novel AI framework, the Spatial-Temporal Graph Attention Network (ST-GAT), designed to provide explainable surveillance of contagion risk within the U.S. banking system. The model constructs a dynamic, directed graph of the entire interbank network by reconstructing bilateral exposures between 8,103 FDIC-insured institutions using publicly available Call Report data from 2010 to 2024. It uniquely combines a Graph Attention Network (GAT) to model spatial connections between banks with a Bidirectional LSTM (BiLSTM) component to capture temporal trends, achieving a high Area Under the Precision-Recall Curve (AUPRC) score of 0.939, just shy of top-performing XGBoost models.

Crucially, the framework is built for regulatory alignment and explainability. An ablation study confirmed the temporal component adds a significant +0.020 boost to performance. The model's 'attention' mechanisms allow regulators to see which banks and which time periods the model focuses on when assessing risk. Furthermore, a permutation importance analysis pinpointed Return on Assets (ROA) and the Non-Performing Loan (NPL) Ratio as the dominant predictive features, a finding that aligns with post-mortem analyses of the 2023 regional banking crisis, validating the model's real-world relevance.

The entire project is a significant step toward transparent, AI-powered financial oversight. All data used are publicly available from the FDIC and FRED, and the author has released all code and results, promoting reproducibility and trust. This moves beyond a 'black box' approach, giving regulators a tool that not only predicts potential bank distress with high accuracy but also explains why, highlighting specific vulnerable institutions and the financial health indicators driving the concern. It represents a practical blueprint for integrating advanced, explainable machine learning into systemic risk monitoring.