Forecasting Commercial Customers Credit Risk Through Early Warning Signals Data: A Machine Learning based Approach

This study introduces an innovative Early Warning System (EWS) for credit portfolio monitoring for commercial customers at a large international bank headquartered in The Netherlands. Traditional EWS methods use backward-looking indicators like default probability or loss given default, which restricts their prediction accuracy. The study investigates the efficiency of using a Watchlist (WL) trigger to improve customer credit risk forecasting. We evaluate migration sensitivity, trigger precision, and time lag for different client statuses to determine this trigger's usefulness. Machine Learning (ML) models, including Linear Discriminant Analysis (LDA), Logistic Regression (LR), Decision Trees (DT), Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting Machine (GBM), Extreme GradientBoost (XGBoost), and Artificial Neural Networks (ANN), are implemented to predict financial strain and negative client migration using internal and external data. We also explore the use of SHAP values to enable explainability. Our results show that the best-performing model is the RF algorithm, with high F1 scores, trigger precision, and migration sensitivity. This model predicted 12.7\% of negative client migration and prevented 67.6\% of cases leading to bank losses. The study suggests the use of EWS for credit risk monitoring of commercial customers and supports bank managers to make strategic decisions.