Corporate default risk prediction model based on graph neural networks

Abstract

This study presents a graph neural network (GNN)-based framework for corporate default risk prediction by explicitly modeling inter-firm dependencies through supply-chain links, ownership ties, and board overlaps. We construct a heterogeneous information network (HIN) from longitudinal firm-level data (20142024), transforming relational structures into learned node embeddings via a three-layer relational graph convolutional network (R-GCN). These embeddings are further processed by a survival-analysis layer to estimate time-varying default probabilities. Compared to traditional financial ratio models and graph-agnostic machine learning baselines, the proposed GNN consistently outperforms in terms of AUC, F1-score, and Harrells C-index. Extensive ablation studies confirm the importance of incorporating economic relationships, particularly supply-chain dependencies, which exhibit the highest marginal contribution to predictive performance. Moreover, the model demonstrates robustness across economic cycles, including during the COVID-induced stress period. A case study further illustrates how GNN-based message passing reveals upstream contagion paths not captured by standalone financial metrics. The findings highlight the significant value of network-aware credit risk modeling in regulatory supervision and financial decision-making. Our approach suggests that integrating network topology into corporate risk assessment can offer timely, interpretable, and materially superior insights for early warning systems and financial resilience analysis.