Understanding the sensitivity and dynamical origins of the tails of some standard ensemble diagnostics

Operational ensemble algorithms only formally optimize error variance over the short range and, in the case of non-Gaussian errors, don’t fully constrain the upper tails of error. Further, ensemble scorecards tend to measure system performance from an average, rather than extreme, standpoint. Thus, the effect of ensemble design choices on the upper tail of any given ensemble diagnostic distribution is somewhat of an unknown quantity. Here, the sensitivity of the distributional tails of several standard ensemble diagnostics including the continuous ranked probability score (CRPS) and the spread-skill ratio is examined as a function of common ensemble design elements such as horizontal resolution, ensemble size, and stochastic forcing. Additionally, case studies are used to understand the dynamical/physical circumstances which give rise to the diagnostics’ extreme tail values.