Systemic Risk Minimization via Risk-Aware Graph Neural Network Reconstruction

We develop a risk-aware graph representation learning framework for financial exposure network reconstruction that incorporates systemic risk directly into the learning objective. The model combines a GraphSAGE encoder with a two-head decoder for link existence and exposure magnitude, and augments reconstruction loss with a differentiable DebtRank penalty. On the EBA 2016 bank–sovereign network (N = 49), the reconstruction-only base line achieves AUC-ROC 0.966. The risk-aware model reduces DebtRank by 50.3%, mainly through lower exposure magnitudes while preserving connectivity, and outperforms uni formly rescaled networks matched to the same total exposure mass. On a synthetic bank–asset system with N = 1,000 and K = 20,000, reconstruction is harder (AUC ROC 0.743±0.022) because sparse topology is only partially recoverable from node-level features. Across five seeds, the risk penalty reduces DebtRank by 42.5% ± 8.0% relative to seed-matched baselines. Since these baselines overestimate observed-network risk, the synthetic results primarily demonstrate computational scalability rather than achievable risk reduction. Overall, the framework provides a scalable alternative to repeated instance specific constrained optimization at inference time.