Model for predicting the probability of event occurrence using logistic regression: case for a Kenyan commercial bank

Abstract

Credit Scoring has been a key undertaking of lenders over the last few decades. However, the most common use of credit scoring has been as a credit application appraisal tool rather than as a credit monitoring tool for existing credit holdings. This study, in the literature, investigated the different sets of sophisticated and classical credit scoring techniques used in the areas of classification and prediction of customers defaulting on credit repayments. The main aim of this study was to build a behavioral credit scoring model from a provided dataset using Logistic Regression. The study also aimed to use data mining concepts to develop a prototype to automate the modeling process and calibrate predictions so as to formulate score cards. Our review of literature indicated the non-existence of an overall superior method for credit scoring since modeling objectives always differ and are thus suited for diverse methods. Model validation and assessments methods, variable selection and interpretation of validation results were investigated. The results of the study indicated the possibility of use of validation methods from related fields, the formulation of model deployment frameworks and formulation of a guide to trade-off between model hit-rate and false alarms. Transformation variables were also found to often offer better predictive power over raw data variables whilst the Gini-Index was shown to be a good indicator of model over-fitting. Finally, this study suggests possible future research endeavors in the effort to eliminate the requirements of statistical knowledge in the Credit Scoring process.