Stress Testing Russia's Systemically Important Banks Using Machine Learning Algorithms

Stress testing is an indispensable tool for forecasting stability of the banking system. However, stress tests that are carried out by the Russian mega-regulator frequently have extremely low accuracy. In this paper with the aim of stress tests predictive power increase, it was suggested to apply machine learning methods, namely: L1 regularization, random forest algorithm and feedforward neural network. Comparison of these methods forecasting quality with the methodology of the Central Bank based on the stress test of capital adequacy conducted in 2014 showed that machine learning algorithms essentially outperform mega-regulator models, therefore, their application seems rational.In this regard, the aforementioned algorithms were applied to conduct stress-test of Russia's twenty largest banks capital adequacy in 2017-2018. It was found that with the development of the domestic economy under a scenario other than negative, Russian banks will have a 3.5% -5% excess of the lower permissible limit of H1.0. Under the negative scenario, banks will also maintain an adequate level of equity capital counteract macroeconomic fluctuations; nevertheless, two banks will have to raise additional 10 billion rubles. Given the stability of the banking sector to any of the three scenarios, reverse stress testing was conducted to determine the values of the parameters that could lead to a significant deterioration of the banking system condition. According to the results of the reverse stress test, combination of oil price drop to $ 15 per barrel, an increase in the dollar rate to 150 rubles and an acceleration of inflation to 22% will make Russian banks raise 1.6 trillion rubles to meet H1.0 requirements.