CREDIT SCORING MODELLING USING MACHINE LEARNING METHODS

Abstract

Abstract. Credit scoring is considered a successful fi eld of application and a methodological subset of intelligent data analysis. The problem of credit scoring within the framework of thoughtful data analysis can be attributed to the class of classifi cation methods of machine learning. A classifi cation task is used to fi nd diff erent representatives of certain predefi ned classes. The main area of application for scoring modeling is risk management, but in general, scoring models are used for various tasks of binary classifi cation, diagnostics, forecasting of the probability of occurrence of a certain unexpected event, detection of hidden signs through the prism of observed characters with a certain probability, etc. The article discusses credit scoring models built using machine learning methods. The article builds 6 models, among which two with the best characteristics are highlighted: the «random forest» model and the TPOT model. Using the classifi cation report and the ROC curve, it was determined that both models achieve an F1-score -- 0.71 for the studied sample, which is quite acceptable for the credit scoring task. The built models can be used when assessing the creditworthiness of a bank’s clients. The credit scoring model eliminates subjective judgments of bank employees and analyzes data that a person may not notice, making it more eff ective for those clients who can be given credit. The use of machine learning methods also saves time in analyzing data provided by the client for credit approval. In future research, it is advisable to create ensembles of the analyzed algorithms or to consider Bayesian models, and certain types of neural networks or boosters. It is also necessary to carry out detailed feature engineering and model interpretation which is important for fi nal decision-making. Keywords: machine learning, credit scoring, «random forests», TPOT model, classifi cation report, ROC curve