Φ- и Ψ-преобразования, повышающие дискриминирующую силу немонотонных переменных скоринговой модели

Управление финансовыми рисками. 2022. Т. 70. № 2. С. 108–120.

Pomazanov M. V.

As an alternative to various methods of binning variables of the scoring model, a new method of continuous transformation of variables is proposed (ΦΨ  - transformations), which is used when the probability of default (rating) is non-monotonically dependent on these variables. Two types of one-parameter families of transformations and strict conditions indicating the need for "treatment" and what kind of it to use are presented (Φ or  Ψ). The method of calculation of transformation parameters is proposed and justified, practical examples are presented. Based on numerous experience in developing scoring models with the introduction of the proposed transformation, it was found that the use of ΦΨ-transformations for the "treatment" of non-monotonic variables (increasing their local Gini index) gives a noticeable increase in discriminating power at the level of 5-10% of the total Gini index for the revised models.

Keywords: стандартизация, скоринг, standardization, индекс Джини, scoring, Gini index, binning, discriminating power, биннинг, дискриминирующая сила