Adaptive Credit Risk Prediction Under Concept Drift Using Graph Neural Networks and Gradient Boosting

Abstract Credit-risk models in retail banking face two persistent difficulties. Default risk depends not only on a borrower’s own attributes but also on systemic factors shared across related obligors—branch, sector, region, or relationship group—that conventional tabular models cannot represent. The relationship between borrower features and default is also non-stationary, drifting with macroeconomic conditions. We propose a hybrid framework that integrates a GraphSAGE encoder, learning borrower embeddings over a graph that links obligors by feature similarity and shared systemic membership, with a gradient-boosted classifier (XGBoost) that consumes both the embeddings and the underwriting features. A recency-weighted rolling-window mechanism keeps the classifier aligned with the prevailing regime, providing adaptivity under concept drift. Outputs are loan-level probabilities of default consistent with the IFRS 9 staging framework. On a sample, IFRS 9-structured benchmark constructed to contain both risk channels, the hybrid attains an AUC of 0.852, against 0.798 for boosted trees on underwriting features alone and 0.768 for the graph encoder alone, with corresponding gains in Kolmogorov–Smirnov separation and precision–recall. The adaptive variant sustains discrimination through a simulated downturn better than a frozen model. A complete, reproducible implementation accompanies the paper. JEL Classification: C45; C53; G21; G32