Systematic stress tests with entropic plausibility constraints

Stress tests with handpicked scenarios might misrepresent risks either because dangerous scenarios are not considered or because the scenarios considered are too implausible. To overcome these two pitfalls we propose a systematic search for the worst case within a relative entropy ball of sufficiently plausible scenarios. For this purpose we use mixed scenarios, which are risk factor distributions rather than realisations. A Maximum Loss theorem explicitly gives the worst case distribution. The method is illustrated in a number of example applications: linear and quadratic portfolios, stressed default probabilities, stressed correlations, macroeconomic stress tests.