Improved estimation methods of the Kolmogorov-Smirnov statistic, weight of evidence and information value indicators in the credit scoring

Abstract

The improved evaluation methods of the Kolmogorov-Smirnov statistic, Weight of Evidence and Information Value indicators are developed without explicit splitting of the original sample into two subsets with developing corresponding formulas for the predictive (forecasting) power analysis of categorical variables in the credit scoring tasks and other fields of practical application of binary classification methods. The generalization of the classical formulas for the Kolmogorov-Smirnov statistic, Weight of Evidence and Information Value indicators have been performed by means of the aggregate expressions transformation for discrete distributions and cumulative distribution functions applying the inner product of two vectors, projection operators, and also a conditional substitution operator. The improved estimation formulas for the Kolmogorov-Smirnov statistic, Weight of Evidence and Information Value indices are proposed and generally described in terms of the discrete unconditional distribution of the input variable and the conditional distribution of the binary target variable.