MODELING CREDIT SCORING USING DATA MINING METHODS

The paper substantiates a methodological approach to improving the credit scoring system for individuals based on data mining methods. A classification of the main types of scoring models is provided, and the conditions for their application for assessing the borrower’s creditworthiness are determined. The authors paid great attention to the formation of a list of characteristics-indicators that make it possible to assess the borrower. The feasibility of using both quantitative characteristics that reveal the client’s financial capabilities and qualitative ones that relate to demographic and social factors is proven. The methods of constructing scoring models were analyzed, from classical ones – logistic regression, decision tree and cluster analysis, which are the basic algorithms of intelligent data analysis, to machine learning and building neural network models, which allow obtaining a more holistic picture of the financial behavior of a potential client. The paper analyzes a database of borrowers to identify key factors affecting credit risk. The use of clustering allowed us to identify significant differentiation of credit behavior of different groups of bank clients. To improve the accuracy of creditworthiness forecasting, a neural network based on a multilayer perceptron was developed. The network architecture includes a hidden layer, for the formation of which the hyperbolic tangent function was used. The output layer was constructed using a sigmoid activation function, which provides an estimate of the probability of a borrower belonging to creditworthiness classes. The model showed high efficiency in identifying reliable borrowers, but requires optimization for better prediction of defaults due to the imbalance of training data. The proposed scoring model contributes to reducing the level of problem loans, reducing operating costs for processing applications, and improving customer segmentation. The feasibility of a combined approach is substantiated: the use of logistic regression and cluster analysis for preliminary selection of clients with the subsequent use of a neural network to make final decisions in complex cases.