Temporal Graph Learning for Contagion Risk Detection in Financial Networks

Abstract

Financial contagion is no longer adequately understood as a static property of balance-sheet exposure or a periodic stress-test outcome. It emerges from evolving relationships among banks, non-bank financial institutions, markets, payment systems, collateral networks, and cross-border funding channels. This paper examines temporal graph learning as a system-level capability for detecting contagion risk in financial networks. Rather than treating machine learning as a narrow predictive layer, the discussion positions temporal graph architectures within a broader socio-technical infrastructure involving data engineering, supervisory governance, model validation, privacy controls, operational resilience, and policy accountability. The paper first reviews classical network theories of financial contagion and their limits under regime shifts, liquidity spirals, and multiplex interdependence. It then analyzes temporal graph neural architectures, including memory-based, attention-based, and evolving convolution approaches, as mechanisms for representing dynamic exposure, latent channels, and time-varying centrality. The paper argues that detection quality depends less on any single model than on the institutional capacity to integrate heterogeneous data, calibrate trust across agents, preserve interpretability, and manage false positives during crises. It further considers fairness, robustness, and sustainability challenges, especially when models trained on stable periods are deployed during market dislocations. Policy implications include supervisory sandboxes, model risk standards, audit trails, and cross-border information-sharing arrangements. The conclusion emphasizes that contagion risk detection should be designed as a resilient public-private infrastructure rather than a proprietary alert system.