PAPER: Building more accurate credit risk models with Machine Learning

Since the 1950s, credit risk modelling of risk ranking scorecards has utilised methods such as logistic regression and discriminant analysis, identifying generalised linear trends in the data and extrapolating from these simple trends to form predictions. This results in models which are relatively easy to interpret and understand. However, a recognised limitation of these traditional techniques has always been their inability to capture the more complex relationships in the data, resulting in sub-optimal model predictions for less-trivial cases. To model these features requires the linear statistical models to be overlaid with workarounds such as bespoke interaction variables and segmentation, often designed by portfolio experts, or manual underwriting.

Given the level of effort involved in building these workarounds there is a large window of opportunity for improving credit decisioning by adopting more sophisticated methods….click below to read how.