Leveraging Ensemble Learning Methods to Improve Credit Scoring Model Accuracy and Robustness

Abstract

Credit scoring is the term used to describe a statistical analysis that can be applied in finance institutions and banks to determine the creditworthiness of an individual. The bestowers normally manipulate it to decide whether to expand or withdraw credit. The score is also a major factor in evaluating an individual's creditworthiness and determining whether he/she approved for a loan. The proposed study uses an ensemble learning model with the XGBoost package to optimize credit scoring using the Credit Scoring Kaggle dataset, which consists of 31,219 customer records from 2023. The data preprocessing pipeline included handling missing data, outliers, and noise; label encoding; z-score normalization; and data balancing using the Synthetic Minority Oversampling Technique (SMOTE). Stratified sampling (80:20) was also employed to maintain the proportions of the different classes. The suggested XGBoost model was assessed using standard performance measures, including accuracy, precision, recall, F1-score, and ROC-AUC. The experimental findings showed increased prediction performance, with an F1-score of 95.9%, recall of 99.5%, precision of 92.6%, and an AUC of 0.99, which denotes practically perfect classification. Compared with more conventional models such as Decision Trees, Random Forests, and Gradient Boosting Machines, the XGBoost model proved more robust, efficient, and capable of strong generalization. The results underscore that the model is interpretable in addition to improving credit risk prediction accuracy and dependability, as demonstrated by the analysis of feature importance, making it an effective decision-support tool in current financial risk management.