This paper exploits a recent and granular data set for 1,500 German LSIs to conduct a residential mortgage stress testing exercise. To account for model uncertainty when modeling PD dynamics we use a benchmark-constrained Bayesian model averaging approach that combines standard BMA with a benchmark derived from a quantile mapping between the historical PD distribution and the historical distribution of macro variables. To link LGD to current LTV we derive a reduced-form meta-dependency. In the baseline model, we quantify expected as well as unexpected losses. We show that German LSIs, though being mostly sufficiently capitalized, are susceptible to a corrective movement in house prices with a median CET1 ratio reduction of 1.5pp in the severely adverse scenario. We quantify the impact of RWA modeling on stress test results and show that the Standardized Approach leads to an up to 33% lower stress impact relative to the more risk-sensitive "pseudo-IRB" approach.
stress test Bayesian model averaging quantile mapping survey data German residential mortgage market model uncertainty