Dynamic graph neural networks for credit risk modeling : a multi-modal approach to SME default prediction

Abstract

Accurate credit risk assessment is crucial for financial stability, particularly for SMEs, which are often underrepresented in traditional models. This thesis presents a multi-modal framework integrating Graph Neural Networks (GNNs) with Long Short-Term Memory (LSTM) networks to improve credit rating predictions. By integrating temporal loan data, financial reports, and dynamic intercompany relationships for 14,710 Icelandic SMEs, the model outperforms traditional benchmarks such as Logistic Regression and XGBoost and achieves superior or comparable performance compared to predictions reported from Icelandiccommercial banks that report data to the Central Bank of Iceland, particularly outperforming in recall and precision.The GNN component captures financial contagion effects, while the LSTMmodels temporal trends, offering actionable insights into credit risk. These advances enable more precise early warning systems, targeted interventions, and improved macroprudential oversight. The framework sets the stage for future research to incorporate richer data sources, increase interpretability, and adapt to macroeconomic shifts, advancing credit risk modeling for SMEs.