Evaluation of Gradient Boosting Algorithms on Balanced Home Credit Default Risk

Default risk divides customers into two groups-those who pay on time (non-defaulters) and those who don’t (defaulters). Recently, banks and data scientists, have employed different machine learning classification algorithms to categorize default risks. Among many algorithms, using imbalanced credit risk datasets, gradient boosting methods such as XGBoost, LightGBM, and CatBoost have been used to determine credit defaulters. However, this research aims to examine and compare the effectiveness of the three algorithms in terms of ROC AUC, and F1 performance scores on a balanced dataset using SMOTE and ADASYN oversampling techniques providing higher reliability in real-time with decreased wait time. Additionally, this research extracted the top 40 features shortlisted by each model resulting in lesser convergence time and decreased computational overhead during prediction. This study confirms that XGBoost outperforms both LightGBM and CatBoost in terms of AUC score and F1 score which is significantly higher compared to results from earlier studies on the same dataset.