Abstract
The South African banking sector has been faced with many different challenges in recent years, driven by the slow growth of the economy. This among other factors has forced the South African banking sector to explore growth opportunities in other countries on the African continent. The drive for growth on the African continent has brought different changes for banks looking to provide loans in these countries. One particular challenge during the credit scoring process has been the banks' ability to distinguish between good and bad customers who require a loan. Considering that a key income for retail banking results mainly in their ability to provide loans, this dissertation explores an alternative to credit scoring for banks that are looking to move into an environment with data limitations. The objective of this dissertation is to ascertain if the Bayesian approach can improve banks’ ability to properly distinguish between bad and good customers applying for credit when working with data limitations. This approach is compared to the logistic regression approach currently used by the bank under study. Data was obtained from a South African Bank with exposure to Botswana overdraft accounts for retail lending from 2014 to 2017, with only 964 accounts. The MCMC procedure in SAS was used to build the Bayesian model and was compared to the bank's logistic regression model. The RMSE and graphical representation of the actuals vs predicted defaults were used as performance measures to compare the two models. The logistic regression model was found to be better at predicting default than the Bayesian model, when based on RMSE. If we consider the graphical representation, we can identify that the Bayesian model is more stable than the logistic regression model. However, the Bayesian logistic regression model did not outperform the bank's logistic regression model.
M.Com. (Financial Economics)