Machine learning vs logistic regression in credit scoring: A trade-off between accuracy and interpretability?

Abstract

In this thesis, I compare logistic regression to the machine learning models k-nearest neighbor, decision trees, random forest, and gradient booster by creating different credit models. By using data from an anonymous Norwegian bank for consumer loan borrowers, I compare the models when continuous variables are split into intervals by using weight of evidence, and when they are kept in their raw form. By using Area under Receiver Operating Characteristic (AUROC) and Brier score as performance measures, I find that logistic regression and gradient booster are the most accurate models for this dataset, and logistic regression is recommended because of its interpretability.