СОПОСТАВИТЕЛЬНЫЙ ТЕСТ ДЛЯ РАССЧИТАННЫХ ЗНАЧЕНИЙ ВЕРОЯТНОСТЕЙ ДЕФОЛТА, ПОЛУЧЕННЫХ В РЕЗУЛЬТАТЕ ПРИМЕНЕНИЯ РЕЙТИНГОВЫХ МОДЕЛЕЙ
Importance Validation of the consistency of rating model forecasts.
Objectives To provide rating model developers and validators with a practical fundamental test for benchmarking the calculated default probabilities resulting from the application of the models used in the rating system.
Methods The classical interval approach of testing statistical hypotheses, focused on the subject area of calibration of rating systems.
Results In addition to the generally accepted tests for the correspondence of the predicted probabilities of default of credit risk objects to the historically realized values, a new statistical test is proposed that corrects the shortcomings of the generally accepted ones, focused on "diagnosing" the consistency of the implemented discrimination of objects by the rating model. Examples of recognizing the reasons for a negative test result and negative consequences for lending are given while maintaining the current settings of the rating model. The proposed method, in addition to the bias in the assessment of the total frequency of defaults in the loan portfolio, makes it possible to objectively reveal the inadequacy of discrimination against borrowers with a calibrated rating model, to diagnose the “disease” of the rating model. Moreover, this does not require the completeness of statistics in each rating category, which expands the scope of applicability of comparative analysis on historical data with a small number of defaults that occurred during the validation period.
The scope of the results is the process of internal validation by the bank of its own rating models, which is required by the Bank of Russia for approaches based on internal ratings.
Conclusions and Relevance It is concluded that the new practical benchmark test allows, at a given level of confidence and available historical data, to reject the hypothesis about the consistency of assessing the probability of default by the rating model, and the test has the advantage of practical interpretability, based on its results, it is possible to draw a conclusion about the direction of the model correction.