Information, Predictability, and Verification at the Thunderstorm Scale

Previous studies have shown that convective-allowing forecasts, despite reaching error saturation in the traditional predictability sense, still contain useful information to the forecaster. Indeed, new results presented here in the context of idealised simulations in buoyancy--shear space confirm that the theoretical predictability of thunderstorms may have two distinct predictability horizons depending on definitions: predictability loss measured by energy-based metrics is maximised by the turbulent churning inside supercells, while another “mode-based” predictability loss denotes a tipping point between the solutions of supercells and weak/dissipating cells in the experiment results presented herein.

These results highlight the need to measure a holistic forecast utility, and not just inherently gridpoint-based techniques, using a paradigm such as Information Theory. Furthermore, recent advances in scale-aware and feature-based verification must be merged with both a probabilistic framework and account for this extra value a forecast desires. Attributes mined from forecasts that may be useful to a forecaster, but are not explicitly considered by some verification metrics, include storm mode, storm speed, and object realism. In the case of storm objects, comparison between two ensembles of 3- and 1-km horizontal grid spacing yields only small improvements in traditional metrics, yet there are substantial differences in object (thunderstorm) attributes including size, movement, frequency, and updraught intensity when resolution increases. Herein, we propose a more appropriate sample climatology for assessing the predictability horizon at the thunderstorm scale, and implement more recent information-theory-based probabilistic verification to capture the information gain from an ensemble forecast system: gain (over a baseline climatology) that goes beyond a gridpoint-by-gridpoint evaluation.