Application of Stacking Model Combined with SMOTE and Bayesian Optimization for Credit Risk Assessment

Abstract

    Credit risk prediction is a critical task for financial institutions to minimize the risk of default and optimize lending decisions. In the context of rapid advancements in machine learning techniques, many classification methods have been developed to improve credit risk prediction capabilities. This study applies a stacking model to assess credit risk, combining predictions from various machine learning models, including XGBoost, Random Forest, and CatBoost. A meta-model, logistic regression, is used to optimize predictions from base models to generate the final prediction. Data is processed using the SMOTE technique for balancing, and the hyperparameters of the base models are optimized through Bayesian optimization. The results show that the stacking model achieves an accuracy of 95.503 % and an ROC-AUC score of 98.15 %, demonstrating the high reliability of the predictions. These results highlight the applicability of machine learning models in credit risk assessment, supporting financial institutions in making individual credit decisions.