Modeling Financial Risk Propagation and Systemic Contagion via Stability-Aware Graph Neural Networks

Abstract

The increasing complexity and interconnectivity of global financial systems have intensified the challenge of identifying and mitigating systemic risks. Traditional econometric and network-based approaches often fail to capture nonlinear dependencies and dynamic contagion effects that characterize modern financial networks. To address this issue, this paper proposes a Graph Neural Network (GNN)-based framework for financial risk propagation and systemic stability analysis. The model leverages both structural and temporal dependencies among financial entities-such as banks, corporations, and markets-represented as nodes and edges in a dynamic graph. Through message-passing mechanisms, the proposed model aggregates information across neighboring nodes to learn latent risk representations, enabling accurate modeling of cascading failures and contagion paths. A novel stability-aware loss function is introduced to penalize high-risk clustering and enhance robustness under volatile market conditions. Experiments conducted on real-world interbank transaction and equity exposure datasets demonstrate that the GNN-based approach achieves superior performance compared to baseline methods including LSTM, VAR, and classical contagion models. The results highlight that GNNs can effectively uncover hidden risk linkages, forecast systemic vulnerability, and provide valuable insights for regulatory supervision and stress testing. This study contributes to a growing body of research bridging graph deep learning and financial system analysis, offering a scalable and interpretable paradigm for risk-aware financial modeling.