A new non-Gaussian evaluation method for ensemble forecasts based on analysis rank histograms

Abstract

The verification of ensemble forecasts is an active field of research. Today, more and more weather forecasting centers implement ensemble forecasts in their operational weather prediction cycles, giving rise to the demand for reliable verification methods especially designed for probabilistic forecasting. Beside well known scores (such as the Brier score or the continuous ranked probability score) analysis rank histograms or Talagrand diagrams are very popular amongst the ensemble forecasting community. However, a shortcoming of this verification method is its graphical character. In this paper, we present a score which can easily be derived from analysis rank histograms. The so called β-score holds the advantage of expressing the graphical character of the histogram, and therefore the character of the ensemble spread, in just one single number. We show results of a β-score analysis of operational ensemble forecasts from different centers for different regions and forecast lead times. In addition to the β-score, we present the so called β-bias which quantifies the shift of analysis rank histograms to lower or higher values and therefore a relative measure of ensemble forecast bias. Results of the β-bias are also shown in this paper.