Evaluation of stacking for predicting credit risk scores

Over the past few years, credit scoring has got an increasing attention for the financial institutions and has become a popular research topic. The purpose of this study is to construct an ensemble classification model based on the machine learning techniques to increase the prediction performance of credit scoring. First; Logistic Regression, Multivariate Adaptive Regression Splines, Support Vector Machines, Random Forest, Gradient Boosting are selected as base classifiers and fitted on the data sets. Second, a stacked generalization ensemble model is integrated through these base classifiers. The model is evaluated on four real-life credit scoring data sets to test its prediction performance and effectiveness. Four performance metrics are chosen to evaluate performance of single base models and stacking. The results demonstrated that the model has slightly better performance than the single base model classifiers in terms of different performance criteria. This study shows that this method can be an alternative and provide an effective decision support for the financial institutions. Keywords: Credit Scoring, Machine Learning, Stacking, Classification, Probability of Default