Retail Credit Scoring models before and after the credit crunch
Andrássy, Árpád
There is extensive literature dedicated to comparing various scoring techniques in the retail banking business, but there are few papers analyzing the performance of scorecards after the credit crunch, and testing their resilience through distress periods . The first purpose of the Project was to analyze the predictive power of several parametric (simple linear regression, logistic regression) and nonparametric (Neural Networks and Support Vector Machines) statistical models for credit scoring, tool widely used in the lending activities of banks present on the retail credit market. Aside from estimating their predictive power, the project focuses on testing their functionality by assessing their stability and robustness in the pre- and post crisis environment. Relative comparisons of the four techniques are also provided. The findings shows that predictive power of pre-crisis models tend to break down in the post-crisis period, however there is “life after the credit crunch” for scoring tools, meaning that much of the model risk can be corrected by continuously maintaining(re-developing and testing) the predictive tools of a lending institution. Another attempt is to propose a simple model that would allow integrating scoring discriminating power, systematic (macro) risk into the management decision. This should be a starting point to provide clearer guidance to managers on how their decision would impact results in the short medium term.
