Reproduce “Advanced User Credit Risk Prediction Model using LightGBM, XGBoost and Tabnet with SMOTEENN” on OpenResearch

About this paper

Chang Yu, Yixin Jin, Qianwen Xing, Ye Zhang, Shaobo Guo, Shuchen Meng

Researchers from Northeastern, Michigan, Chicago, Pittsburgh, Illinois Institute of Technology, and Central University of Finance and Economics developed an advanced credit risk prediction model. The model, leveraging LightGBM with PCA and SMOTEENN preprocessing, achieved an F1 score of 0.9989 and an AUC-ROC of 0.9999 for identifying defaulting customers.

Key insights

• Robust data preprocessing, specifically PCA for dimensionality reduction and SMOTEENN for class imbalance, is essential for achieving high accuracy in credit risk prediction.
• LightGBM consistently demonstrated superior and robust performance among all models tested, especially when integrated with the optimal data processing pipeline.
• While advanced machine learning models significantly outperform traditional methods when data is well-prepared, the benefits of SMOTEENN can vary among different model types, with some models like Logistic Regression showing decreased performance due to synthetic data sensitivity or aggressive balancing.
• The combination of PCA and SMOTEENN yielded near-perfect results, with LightGBM achieving an F1 score of 0.9989 and an AUC-ROC score of 0.9999, outperforming all other models in this configuration.

Summary by alphaXiv.