Default prediction and contagion risk assessment in corporate guarantee networks via dynamic heterogeneous graph neural networks

Abstract

Corporate guarantee networks in China's bond market generate cascading defaults through interconnected guarantee chains, yet conventional credit risk models treat each borrower as an independent unit and cannot capture network-driven contagion. This paper proposes DHGN-Credit, a dynamic heterogeneous graph neural network framework that jointly addresses default prediction and contagion risk assessment in corporate guarantee networks. The framework constructs a directed, weighted heterogeneous graph from actual guarantee contracts and integrates three components: a metapath-aware heterogeneous graph attention network (H-GAT) that distinguishes guarantee, equity, and industry relation types; a continuous-time memory module that captures guarantee events between quarterly snapshots; and an LSTM temporal encoder with adaptive attention weighting. A dual-task decoder jointly optimizes default probability prediction and contagion impact assessment, with the contagion task supplying network-level regularization to the default prediction objective. Experiments on Chinese bond market data spanning 2015–2024 (3,842 firms, 152 defaults), conducted under a disclosure-aligned feature lag that restricts each snapshot to financials publicly available at the snapshot date, show that DHGN-Credit achieves an AUC of 0.913 and F1-score of 0.656, improving over the strongest dynamic baseline by 3.4 and 5.5 percentage points respectively, and attains a Spearman rank correlation of 0.865 on contagion impact prediction. Phase transition analysis confirms robust-yet-fragile behavior: targeted attacks on 50 core firms trigger contagion at 8.2 times the scale of equivalent random shocks, with a discontinuous inflection near a 30% leverage threshold.