Evaluating Machine Learning Models for Personal Credit Risk: A Comparison of Logistic Regression, Random Forest, and XGBoost

Traditional methods for predicting personal credit risk have historically lacked accuracy and comprehensiveness, failing to effectively analyze a large number of nonlinear user characteristics. Large-scale modeling algorithms have advanced quickly in recent years, and more and more financial institutions are adopting them to forecast the likelihood of personal credit default. However, challenges persist in understanding the distinct features and applicability of different models. In this research, a publicly accessible personal credit dataset from LendingClub covering the years 20072010 is empirically analyzed using three well-known algorithms: Logistic Regression (LR), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost). The Synthetic Minority Over-sampling Technique (SMOTE) is used to generate synthetic samples for the minority class to tackle the notable class imbalance in the dataset. The performance of these models is evaluated and contrasted by four pivotal criteria and Receiver Operating Characteristic (ROC) curves. The outcomes illustrate that XGBoost outperforms the other models across all evaluated metrics. Based on these findings, this paper recommends XGBoost as the preferred algorithm for personal credit risk prediction.