Improving Credit Decision with an Interpretable EBM-Based Model

Supply Chain Finance has been vulnerable to supplier default risks and opaque credit analysis. This paper proposes a credit risk model that evaluates CatBoost, Random Forest Variables, and Explainable Boosting Machine. Ablation tests confirm that CatBoost has 90% accuracy with AUC 0.90, followed by 89% for RF with AUC 0.90, while EBM has been found to be superior with 92% accuracy and AUC 0.93. EBM has been found superior to classical approaches of logistic regression with 82% and AUC 0.84, and decision tree with 85% and AUC 0.86. The model encompasses supplier payments, discrepancies, network dependencies, as well as macroeconomic variables. SHAP values reveal that late payments, liquidity, and order quantities are of prime significance. The model also ensures regulatory compliance with SR-11-7 model risk management practices, which are used to provide early warnings to sub-standard suppliers with algorithmic transparency.