Binary classification in an imbalanced dataset: An application on credit customers of a Turkish bank

Credit risk assessment is of great importance for credit institutions, and many academic studies aim to predict default-prone customers. In our study, we developed binary classification models with common classification algorithms that predict customers likely to experience delays within a month rather than focusing on the long term. Due to the highly imbalanced structure of our dataset, the SMOTE oversampling technique is employed before the machine learning algorithms. We compared the models in three different approaches. In our advanced approach using both oversampling and cross-validation, all models except NB achieved accuracy higher than our benchmark value of 90%. In addition, their AUC values ranged from 84% to 95%. However, when no oversampling was performed, the results showed that the models except NB proved to be ineffective. In the other approach, omitting crossvalidation led to remarkable differences in the results, highlighting the importance of cross-validation.