Non-linearity of scorecard log-odds

In this paper we illustrate this phenomenon of a misaligned log-odds relationship within a model built using a method which is standard in retail banking. The sequential elimination of correlated variable bins results in an improved estimation, at the cost of the model's ranking performance. A post-hoc quadratic transformation of the scores improves the model fit without affecting the ranking performance, but may prove difficult to apply in current operational systems.