Credit risk stress testing for EU15 banks: a model combination approach

In bank stress tests, the role of a satellite model is to tie bank-specific risk variables to macroeconomic variables that can generate stress. For valid stress tests it is important to develop a comprehensive satellite model that both preserves the sense of known economic relationships and also exhibits high predictive ability. However, it is often difficult to achieve these desiderata in a single satellite model. Multicollinearity of key macro variables and limited data may militate against inclusion of all important stress variables, thus limiting the range of stress scenarios. In order to address this problem we depart from the custom of using a single model as the "true" satellite. Instead, we generate a full space of candidate models that we then screen for reasonable candidates that remain sufficiently rich to cover a wide range of stress scenarios. We then develop composite models by combining the surviving candidate models through weighting. The result is a composite satellite model that includes all the desired macroeconomic variables, reflects the expected relationships with the dependent variable (NPL growth) and exhibits more than 20% lower RMSE compared to a commonly used benchmark model. An illustrative stress testing application shows that this approach can provide policy makers with prudent estimates of credit risk.