Random Forest vs. XGBoost for Peer-to-Peer Loan Default Prediction: Evidence from LendingClub

Abstract

This study compares Random Forest (RF) and Extreme Gradient Boosting (XGBoost) for predicting loan defaults on 25,000 LendingClub loans from 2015–2016, using 14 financial features and 5-fold cross-validation. Both models perform similarly: RF reaches an AUC-ROC of 0.717, XGBoost reaches 0.715. RF does better on precision (0.572 vs. 0.504), while XGBoost leads on recall (0.164 vs. 0.116) and F1-score (0.247 vs. 0.193). The gap is small enough that the choice should rest on operational priorities rather than accuracy: Random Forest for interpretability, XGBoost for recall.