Model and estimation risk in credit risk stress tests

This paper deals with stress tests for credit risk and shows how exploiting the discretion when setting up and implementing a model can drive the results of a quantitative stress test for default probabilities. For this purpose, we employ several variations of a CreditPortfolioView-style model using US data ranging from 2004 to 2016. We show that seemingly only slightly differing specifications can lead to entirely different stress test results—in relative and absolute terms. That said, our findings reveal that the conversion of a shock (i.e., stress event) increases the (non-stress) default probability by 20–80%—depending on the stress test model selected. Interestingly, forecasts for non-stress default probabilities are less exposed to model and estimation risk. In addition, the risk horizon over which the stress default probabilities are forecasted and whether we consider mean stress default probabilities or quantiles seem to play only a minor role for the dispersion between the results of the different model specifications. Our findings emphasize the importance of extensive robustness checks for model-based credit risk stress tests.