?
Continuous transformation of nonmonotonic factors of the scoring model increasing its discriminating power
When developing scoring models, Weight of Evidence (WOE) transformation is actively used, which is in demand to improve the discriminating power of the factors included in the model. For WOE transformation, various binning methods are used, the dimension of which is limited by the available statistics. As an alternative, the method of continuous transformation of model factors is proposed, which is applied when the target binary variable depends nonmonotonically on these factors. Two types of one-parameter families of transformations and strict conditions are proposed that determine the need for "healing" of factors, as well as determining the appropriate type of transformation (Φ or Ψ). The method for calculating transformation parameters is substantiated, and practical examples are presented. The use of the proposed transformation in the practice of developing scoring models in the banking sector confirms that the use of ΦΨ - transformations to correct the nonmonotonicity of the model factors gives a significant increase in the discriminating power at the level up to 10% of the Gini index of applied models for binary classification
In book
