Comparison of Logistic Regression and XGBoost Model Performance in Predicting Credit Scores

Abstract

Credit Scoring is a mathematical approach used to assess the creditworthiness of individuals or companies by classifying debtors into certain categories based on their risk profiles. This study aims to compare the performance of the Logistic Regression and XGBoost machine learning algorithms in predicting credit scores (credit scoring) to reduce the risk of Non-Performing Loan (NPL) risk at PT Graha Mazindo Mandiri. The secondary dataset used contains 1,533 car loan debtor data with 17 variables, including 1dependent variable and 16 independent variables. The research process includes data preprocessing (cleaning, handling outliers, encoding, normalization, and class balancing with SMOTE), modeling, and evaluation using the Accuracy, Precision, Recall, F1-score, and ROC-AUC metrics. The results show that XGBoost excels with 96% accuracy and ROC-AUC of 0.99 compared to Logistic Regression with an accuracy of 88% and ROC-AUC0.94, due to XGBoost ability to capture non-linear patterns and handle data imbalance. This study provides insights into credit risk factors and supports more accurate credit decision-making, with recommendations for hyperparameter optimization and model integration into operational systems.