Modern Approaches to the Modelling of Financial Risks and Stress-Testing in Financial Institutions

Choosing the best model for calculating potential credit risks remains a pressing issue for any financial institution. With the active development of machine learning models, internal models for risk assessment are also changing, becoming more complex and improving their accuracy. The paper analyses the portfolio of one of the largest Russian banks and the application of various methods of estimating expected losses from unpaid loan obligations to it. The most popular models in the field of risk assessment are used, such as IRB, CreditMetrics (standard approach and modified, using Monte Carlo simulation), and CreditRisk+. Each has its own modelling features. The data used in these approaches are also important: in case of significant imbalances, the accuracy of the forecast may significantly overestimate or underestimate the real values of credit losses. The most flexible tool turned out to be the IRB approach, which demonstrated high accuracy of the total portfolio risk assessment. The paper gives recommendations on possible adjustment of the risk strategy, with the choice of the model with the best predictive ability, which turned out to be CreditRisk+. The results of the study also allow us to conclude on the applicability of off-the-shelf credit risk modelling methodologies for a portfolio with low default frequency. In addition, the segment selected for the analysis is also characterised by a rather large variance between borrowers, which affected the accuracy of the developed PD model. The results of the study may potentially be of interest to bank management as well as the regulator.