Adaptive Credit Risk Prediction Under Concept Drift Using Graph Neural Networks and Gradient Boosting (Research Square)

Summary

• The proposed hybrid model combining a GraphSAGE encoder and XGBoost achieves an AUC of 0.852, outperforming standalone boosted trees on underwriting features (0.798) and a graph encoder alone (0.768).p.2
• The credit risk framework outputs loan-level probabilities of default aligned with the IFRS 9 staging requirements for expected credit loss calculations.p.2
• Under concept drift in a simulated panel, the recency-weighted adaptive model maintains an average AUC of 0.887 compared to 0.879 for a frozen static model.p.12
• Feature importance analysis shows that graph-embedding dimensions account for roughly half of total split importance in the fused downstream booster.p.11
• The adaptive scheme uses an eight-month rolling window with an exponential recency time constant of three months to dynamically retrain only the classifier.p.8

Key insights

1. Combining relational systemic borrower embeddings with fine-grained tabular underwriting variables overcomes the individual limitations of GNNs and tree ensembles.p.3
2. Constructing borrower graphs using both feature similarity edges and systemic-membership edges allows models to recover community-level risk signals.p.3
3. Separating the adaptation timescales by updating only the downstream classifier allows inexpensive tracking of concept drift without re-learning slow-moving graph structures.p.6
4. Isotonic regression effectively calibrates raw boosting scores to align predicted default probabilities with observed frequencies for regulatory IFRS 9 staging.p.7
5. Static credit scoring models suffer performance degradation during economic downturns compared to rolling-window models with recency weighting.p.12