Identifying Financial Risk Contagion with Large Language Models and Temporal Graph Learning

Fostering economic growth, particularly the development of small and medium-sized enterprises, is central to United Nations Sustainable Development Goal 8. Small and medium-sized enterprises can secure bank loans with the help of guarantors, forming interconnected networked-loans. While essential for access to credit, networked-loans can also amplify risk contagion and thus are vulnerable to epidemic financial crises. Previous research has failed to capture nuanced financial semantics and analyze the mechanisms of risk contagion. We designed FinDoctor to tackle these challenges. FinDoctor leverages large language models to extract context-aware features from unstructured financial text. It then combines temporal graph neural networks with the susceptible, exposed, infectious, recovered epidemiological model to identify financial risk contagion. This enables both accurate prediction and targeted risk mitigation. Evaluated on 2 real-world datasets, FinDoctor achieves state-of-the-art performance, with area under the curve scores of 88% and 89%. We further develop an interactive application that visualizes risk propagation and supports targeted intervention. Our work provides a full-cycle solution to identify financial risk contagion in networked-loans.