Default Contagion Detection in Industrial Enterprise Networks Using Multi-Relational Graph Neural Inference
TL;DR
The results demonstrate that multi-relational graph neural inference can capture default transmission mechanisms that are difficult to observe through firm-level indicators alone.
Abstract
Default contagion in industrial enterprise networks may arise through trade credit, cross-guarantees, shared shareholders, interlocking directors, regional industry clusters, and financing dependencies. A single enterprise default can therefore increase risk for connected firms, even when their own financial indicators appear stable. This study proposes a multi-relational graph neural inference model for default contagion detection in industrial enterprise networks. The model constructs a multi-layer enterprise graph containing transaction relations, guarantee relations, ownership relations, executive relations, financing relations, and regional industry association links. A relational graph neural network is used to aggregate risk signals across different relation types, while a path-based inference module estimates whether default risk is transmitted through direct or indirect enterprise connections. The evaluation dataset contains 118,000 industrial enterprises, 2.36 million multi-relational edges, 96,000 financing contracts, 58,000 cross-guarantee records, 31,400 ownership-change events, and 12,800 confirmed default cases over 48 months. The proposed model identifies 18,600 contagion-sensitive enterprise pairs and 4,270 high-risk default transmission communities. Compared with a standalone financial-statement model, it reduces median missed-warning interval from 74 days to 29 days for firms affected by upstream default events. The graph inference module discovers 9,350 hidden contagion paths longer than two hops, many of which are caused by shared guarantors and repeated financing intermediaries. Portfolio simulation shows that prioritizing firms by inferred contagion exposure reduces expected overdue loan amount by 184 million RMB across the validation period. The results demonstrate that multi-relational graph neural inference can capture default transmission mechanisms that are difficult to observe through firm-level indicators alone
