Comparative Analysis of Bayesian and Classical Logistic Regression for Credit Scoring in Kenya Using Non-Informative Priors

Abstract

Credit scoring is essential for financial institutions, particularly in Kenya, where limited or nonexistent credit histories complicate the assessment of new clients. Classical logistic regression models, though widely used, can underperform in such data-scarce environments. This study compared classical logistic regression with Bayesian logistic regression models employing non-informative priors, based on several classification criteria, to assess whether non-informative priors can be successfully used in developing credit scoring models without compromising predictive accuracy. Data was collected retrospectively from one of the tier-one commercial banks in Kenya. The data included borrower demographic information, financial attributes, credit history, and loan default status. The performance of the models was compared mainly using AUC. The results showed that the Uniform prior model achieved the highest AUC (0.8642), Normal prior model (0.8639), Jeffrey’s prior model (0.8636), and Classical model (0.8635). The study provided a comprehensive overview of other metrics, such as residual deviance, precision, accuracy, recall, and specificity, in order to give room for the financier to choose their preferred model depending on the specific objectives and risk appetite. In addition, each model's strengths and limitations were highlighted in order to give room for further exploration of the models for better predictive performance.