Graph Neural Networks for Corporate Credit Risk: A Supply-Chain-Aware Default Prediction Model

Traditional credit-scoring models treat firms as independent observations, ignoring contagion effects propagating through supply-chain and ownership networks. We construct a heterogeneous corporate graph for ~5,800 Chinese listed firms, integrating supplier-customer linkages, equity holdings, and bank-loan co-exposure. A relational graph attention network (R-GAT) predicts 12-month default probability with an AUC of 0.872, outperforming firm-only logistic regression (0.781) and a feedforward neural network (0.812). Attention weights identify systemically connected firms whose distress most strongly predicts cascading defaults.