22 April 2021
The seminar highlighted the significance of Credit scorecard modelling in banks and agencies, which has attained a high level of good practice over many years. Many of the early pioneers' and practitioners' beliefs are now well-tested and even required by the profession:
- Logistic regression
- Input variable classification
- Default over-weighting
- Reject inference
- Delta method for scorecard modification
- Use of AUC metrics and much more
These practices may come with warnings or limitations in their application, or they may be recognised as approximations or to have potential biases or errors, but guidance is frequently simply cautionary — "it's better to do it, but watch out in X circumstances!" — and in need of quantification — "add some conservatism to cover Y." Determining when to abandon one traditional technique or how to correct another is still a substantial concern in model management.
Alan offered a geometric description of scorecard adjustment and selection in this session to help address this issue. The seminar's goal was to unite and simplify a complicated menu of excellent practices as examples of geometric connection or misunderstanding in a high-dimensional model and data space.
This point of view is based on statistical information geometry, a lively area of mathematical statistics that began with Rao's differential geometric approach to Fisher information. By applying this basic theory to the situation of scorecards - regressions on contingency tables — the mathematics is considerably simplified, creating a structure that facilitates practical computation and insight.
To demonstrate this, Alan quantified three traditional scorecard practices: model selection, the delta approach in the presence of heavy correlation, and sample weighting. This provided new insights, often confirming the classical approach's correctness while also correcting or simplifying it. It also makes this powerful idea available to scoreboard creators for use in solving other comparable challenges.