Machine Learning in Small Business Finance: SBA Loan Default Prediction with XGBoost and Explainable AI

Abstract

This study advances small business credit risk assessment by comparing traditional statistical models with advanced machine learning (ML) techniques, using a publicly available data set of 89,621 U.S. Small Business Administration (SBA)-backed loans. The analysis compares logistic regression, decision tree and extreme gradient boosting (XGBoost) models, incorporating rigorous data cleaning, feature engineering, hyperparameter tuning and consistent evaluation metrics. XGBoost emerged as the best performer, achieving a receiver operating characteristic–area under the curve of 0.957, recall of 75% and F1 score of 78%, effectively identifying true defaulters, handling class imbalance and capturing complex non-linear relationships. Explainable AI (XAI) techniques were applied to interpret XGBoost’s predictions, offering transparent and interpretable insights that address regulatory concerns over black-box models. By combining high predictive accuracy with explainability, this research delivers a large-scale, open-data application of ML versus traditional models in small business lending, providing a replicable methodological framework and policy-relevant guidance for improving efficiency and decision-making transparency in SBA loan approvals.