ST-GAT Achieves 0.939 AUPRC in FDIC-Based US Interbank Contagion Model
Regulatory-aligned ST-GAT delivers leading GNN performance on public FDIC data for US interbank distress detection while supplying temporal attention explanations consistent with 2023 crisis drivers.
ST-GAT models 8103 FDIC-insured institutions over 58 quarters (2010Q1-2024Q2) as dynamic directed weighted graphs with bilateral exposures reconstructed via maximum entropy from public Call Reports (https://arxiv.org/abs/2604.14232). The model records AUPRC 0.939 +/- 0.010, exceeding all tested GNN variants and trailing XGBoost by 0.005. Ablation isolates +0.020 AUPRC from the BiLSTM component; temporal attention weights decline monotonically, assigning higher importance to long-run structure.
Permutation importance ranks ROA (0.309) and NPL ratio (0.252) highest, matching FDIC post-mortem statistics on 2023 regional bank failures (https://www.fdic.gov/news/press-releases/2023/pr23033.html). Original abstract omits explicit linkage to Eisenberg-Noe clearing mechanisms used in pre-2010 contagion literature (Upper, BIS Quarterly Review, 2011, https://www.bis.org/publ/qtrpdf/r_qt1109e.pdf). Framework synthesizes GAT attention (Veličković et al., https://arxiv.org/abs/1710.10903) with FSB-noted requirements for explainable AI in systemic risk monitoring (FSB, 2021, https://www.fsb.org/wp-content/uploads/P011121.pdf).
Temporal component and public-data constraint address documented regulatory barriers to black-box deployment cited in both FSB and BIS reports on AI for financial stability, patterns visible across post-2008 maximum-entropy network reconstructions (Mistrulli, Journal of Banking & Finance, 2011).
AXIOM: ST-GAT shows regulators can obtain high AUPRC contagion forecasts from public Call Report graphs while using attention weights and permutation ranks to satisfy explainability mandates.
Sources (3)
- [1]Explainable Graph Neural Networks for Interbank Contagion Surveillance(https://arxiv.org/abs/2604.14232)
- [2]Contagion in financial networks(https://doi.org/10.1016/j.jfs.2010.02.001)
- [3]Artificial intelligence and machine learning in financial services(https://www.fsb.org/wp-content/uploads/P011121.pdf)