КАЛИБРОВКА СКОРИНГОВОЙ МОДЕЛИС УЧЕТОМ ЦЕНЗУРИРОВАННЫХ ДАННЫХ

Abstract

In connection with the introduction of the Basel II Agreement and IFRS 9, the question of a more accurate assessment of bank credit risk is becoming increasingly important. In accordance with this provision, banks independently calculate the credit risk assessment, which is most often based on a historical sample in the form of a scoring model. The problem that arises when building a model is the evaluation of credit agreements that they stop functioning before the date on which the model forecast was built, that is, these loans are eliminated from the observation before the end date of the study. These loans are called censored, and in the context of the study there is censorship on the right. At the same time, the influence of such credit agreements on the level of bank default is significant, and, therefore, the value that serves as the basis for calibrating the scoring model also influences the value of the calibration coefficient. The purpose of this study is to solve the problem of accounting for censored data when calibrating a scoring model at the validation stage. The article discusses various ways of accounting for censored data, namely, 1) accounting for censored loans as “good”, 2) excluding censored loans from the sample, 3) Kaplan-Meier method, 4) weighting method. At the same time, attention is paid to the currently relevant issue of several estimates of the share of defaults itself, taking into account censored contracts, and the use of censored data when adjusting the model risk assessment during the validation of the scoring model. In this article, for each of the censored data accounting methods, the influence of the calibration coefficient on the ratio of the model number of defaults to the actual is analyzed, for which three methods of model calibration are considered: Linear calibration values from the probabilities, Linear calibration values from the odds, Logarithmic calibration values from the odds. According to the results of the research, the conclusion is drawn on the dependence of the method of accounting for censored data on the policy of a credit institution. A regional retail bank experience is taken to bring an example for calculation. According to the results of the study, it is concluded that the method of accounting for censored data depends on the policy of a credit institution: for organizations with a low-risk appetite, it is necessary to use the method of eliminating censored loans, for organizations with a high-risk appetite, consider censored loans as “good”, and to obtain more accurate forecast with adequate risk appetite use the methods of weighing censored data and Kaplan-Meyer. Further studies in the field will consider censored data not only at the stage of validation of the scoring model, but also at the initial stage of its construction. Keywords credit risk, commercial bank, probability of default, a scoring model, validation of a scoring model, censored data, methods for the accounting of censored data, calibration coefficient, calibration methods, risk appetite.