Dynamical Ensembles: A Critical Assessment

Over the past decades, ensemble forecasting has become a ubiquitous part of the practice of Numerical Weather Prediction (NWP). Today, an estimated 75% of computational and other NWP research and operational resources are devoted to ensembles. Their dynamical generation and use are predicated on reality being a random draw from the ensemble members. Yet this foundational assumption is found to be violated not only in operational, but also in idealized or perfect ensembles. Representative sampling that is easily accomplished in a single dimension is shown to be theoretically unattainable in the multidimensional space of atmospheric dynamics. After an analysis of the forecast skill degrading effect of this widely used technique, alternative approaches for the filtering of unpredictable forecast variance, the generation of plausible forecast scenarios, and the estimation of forecast error variance will be discussed.