Comparative Analysis of Machine Learning Models for Credit Default Prediction Using Credit Score Dataset

Abstract

Accurately predicting credit defaults is a key requirement for the proper financial risk management of a modern banking system. In this research, the approach suggested is calculating credit card default prediction through machine learning in the case of the dataset of 30,000 records with 25 features named Default of credit card clients. Exploratory data analysis was also done to know how variables relate to each other and what patterns exist that could be attributed to default behavior. Techniques of data preprocessing, such as treatment of missing values, irrelevant attributes, and outliers, were used to enhance data quality. Feature engineering was carried out to produce more attributes like average payment, average bill amount, and total repayment delay. In order to deal with the imbalance caused by classes, the Synthetic Minority Over-sampling Technique (SMOTE) was used. Some of the machine learning models were implemented and tested with accuracy, precision, recall, F1-score, and ROC-AUC values, such as Random Forest, CatBoost, XGBoost, and LightGBM. The LightGBM model presented results of the best performance of 86.59, 87.83, 84.95, 86.37, and 93.43, indicating its usefulness in credit default prediction.