Interpretable Credit Scoring for Microfinance Borrowers with Gradient Boosting and SHAP

Abstract

Microfinance institutions increasingly use machine learning to assess loan applicants, but regulators and loan officers require explanations for each decision. This study builds a gradient-boosted credit scoring model on 26,000 historical microloans and uses SHAP values to explain individual predictions. The model achieved an area under the ROC curve of 0.81, compared with 0.74 for the logistic regression scorecard currently in use. Repayment history with the institution, business tenure, and group-lending membership were the most influential features. A usability study with twelve loan officers found that SHAP-based explanation cards made it easier to justify approvals and rejections to applicants. The findings suggest that interpretable boosting models can improve accuracy without sacrificing the transparency that microfinance lending requires.