Quantile-based short-range QPF evaluation over Switzerland

Quantitative precipitation forecasts (QPFs) are often verified using categorical statistics. The traditionally used 2×2 contingency table is modified here by applying sample quantiles instead of fixed amplitude thresholds. This calibration is based on the underlying precipitation distribution and has beneficial implications for categorical statistics. The quantile difference and the debiased Peirce skill score split the total error into the complementary components of bias and debiased pixel overlap. It is shown that they provide a complete verification set with the ability to assess the full range of rainfall intensities. The technique enables the potential skill in a calibrated forecast to be estimated without spurious influences from the marginal totals and the problem of hedging is therefore avoided. To exemplify the feasibility of quantile-based contingencies, the method is applied to 6.5 years of operational rainfall forecasts from the Swiss Federal Office of Meteorology and Climatology (MeteoSwiss). Daily accumulations of the COSMO model at 7 km grid size are compared to a high-quality gridded observational record of spatially interpolated rain gauge data. The quantile-based scores are applied to single grid points and to predefined regions. A high-resolution error climatology is then built up and reviewed in terms of typical error characteristics in the model. The seasonal QPF performance exhibits the most severe overestimation over the Northern Alps during winter, indicative of the impact of the model ice phase. The QPF performance related to model updates, such as the introduction of the prognostic precipitation scheme, is also evaluated. It is demonstrated that the potential skill continuously increases for subsequent versions of the COSMO model. Over the entire time period, a strong gradient of the debiased Peirce skill score is evident over the Alps, meaning that the potential skill is much higher on the Alpine south side than on the north side.