Stacking of Ensemble and Boosting Methods for Credit Risk Prediction

Abstract

A significant number of loan applicants may not be able to repay their loans, which poses a risk for banks. To help banks mitigate credit risk, this work proposes four machine learning models that provide a binary classification (good or bad) of credit payers. These models include two boosting algorithms, XGBoost (Extreme Gradient Boosting) and EBM (Explainable Boosting Machines), one bagging algorithm, RF (Random Forest), and a hybrid ensemble learning approach using a Stacking method. The latter is a meta-model which learns from the output probabilities of others algorithms to determine the optimal way to combine them, ensuring a more effective prediction. In fact, our stacking model, trained on these probabilities using Logistic Regression, outperformed the three individual models across various metrics, achieving a well-balanced and improved performance for both classes.We chose EBM because it has proven its performance in a many fields and, above all, its ability to provide transparent explanations.