USING LOGISTIC REGRESSION TO PREDICT PAYMENT DEFAULT IN A TUNISIAN ISLAMIC BANK: DOES OVERSAMPLING IMPROVE MODEL PERFORMANCE?
Credit risk remains the greatest concern for both conventional and Islamic banks. However, there is a scarce literature on the accuracy of credit scoring models in predicting payment default in the case of Islamic banks. On the other hand, data on defaulting customers is scarce (Kiefer, 2009), which leads to a credit scoring model that favors creditworthy customers as the majority class. Thus, this paper aims to employ the logistic regression to predict the default of corporate clients of a Tunisian Islamic bank and to assess the impact of oversampling and undersampling techniques on the performance measures. We use accounting data for a sample of Tunisian companies whose credit applications have been accepted over the period 2012-2017. Findings reveal that logistic regression with oversampling turns out to be more accurate than logistic regression with undersampling and imbalanced data. Furthermore, using performance criteria covering both predictive power, discriminative power, and misclassification costs, we highlight that logistic regression performs better with oversampling.
Logistic regression; payment default; oversampling; undersampling; SMOTE