Soloshenko O. Models and methods for creditworthiness assessment of private individuals

The study is aimed at improvement of systems methodology of building scoring models for lending process and its program implementation within decision support system for building scorecards, and it is aimed at creating concrete predictive models. Scientific novelty of the research includes proposed discretization method for continuous variables, proposed estimation method for key predictive power indicators of categorical variables using aggregated input data, improvement of estimation method for statistical significance levels, improvement of k-nearest neighbor method, generalizations of logistic regression, Weight of Evidence, the Gini index and the Kolmogorov-Smirnov statistic for a probabilistic target variable and improvements of the key reject inference methods. In particular, as an example, the application credit scoring model for retail consumer lending demonstrates the advantages of the improved iterative classification method for reject inference implementation using two-level generalizations for a probabilistic target variable case. The results obtained show, in particular, faster convergence of the proposed improvement using Chebyshev distance for difference estimation of predicted probabilities for set of rejected applications and better forecast performance using validation sample. Further complex methodological generalization, including generalization and domain extension of model fit measures which were previously applied for binary case only, allows the researchers to extend the application area of credit scoring methodology for the case of continuous probabilistic target variable that was not previously carried out. The information system for building scoring models is developed.