APPLICATION OF MACHINE LEARNING METHODS TO THE CALCULATION OF BANK'S RETURN ON EQUITY UNDER A LOAN AGREEMENT

Abstract

The requirements for a more accurate assessment of the individual risk of a borrower became more complicated with the introduction of Basel II and IFRS 9. Such risk assessment is more and more often carried out using the construction of scoring models, however, as a rule, the Gini coefficient acts as a quality criterion for the constructed models, and the influence of modeling on financial component, namely on the return on equity, which acts as the basis for doing business in the field of lending, is not investigated at all. In this regard, the article proposes a methodology for assessing the return on equity without taking into account risk and its complication by taking into account the individual risk of a borrower. The construction of a dynamic model for assessing credit risk in the article is considered on the basis of survival models constructed by machine learning methods. The problem of accounting for censored data is solved using specific construction of variables for the model and methods that take into account censorship: logistic regression, Cox proportional risk model, random survival forest model. On the example of the data of a regional commercial bank, the return on equity is estimated and compared, depending on the choice of a risk assessment model. The result of the study is the conclusion that it is necessary to apply the methodology for calculating the return on equity taking into account risk assessed by the machine learning method.