Current challenges include the representation of forecast uncertainty due to the use of imperfect numerical models. Model uncertainty can be addressed through the use of multimodel ensembles (in which each single model run is deterministic), or through stochastic representations of parameterized physical processes, as implemented in the ECMWF medium-range ensemble, thereby introducing randomness into the model runs (13). Both options link flow-dependent forecast uncertainty and model-related errors, and it remains to be seen whether they are superior in any way to approaches based purely on statistical postprocessing (7, 14). Nor has the debate on selective versus Monte Carlo sampling of initial condition uncertainty been resolved, although it may evolve in novel directions as operational experience with various methods of sequential data assimilation accrues.
From daily to seasonal time scales, probabilistic forecasts based on ensembles have become a prominent part of numerical weather prediction. The ability of ensemble systems, in concert with statistical postprocessing, to improve deterministic forecasts—in that the ensemble mean forecast outperforms the individual ensemble members—and to produce probabilistic and uncertainty information to the benefit of weather-sensitive public, commercial, and humanitarian sectors has been convincingly established. More work needs to be done to routinely provide fully reliable, flow-dependent probabilistic forecast distributions, particularly of weather fields, as opposed to forecasts at individual sites. In keeping with the remarkable pace of progress since the early 1990s, we anticipate notable improvements in deterministic and probabilistic forecast skill through the continued development of multimodel, multi-initial condition ensemble systems and advanced, grid-based statistical postprocessing techniques.
Additional effort is required in the communication, visualization, and evaluation of probabilistic forecasts, and differing interpretations of probability need to be reconciled, to avoid the risk of perfecting ensemble methodologies without a clear aim (15, 16). To this end, the paradigm of maximizing the sharpness of the probabilistic forecasts under the constraint of calibration may offer guidance. Calibration refers to the statistical consistency between the probabilistic forecasts and the observations; sharpness refers to the spread of the predictive distributions and is a property of the forecasts only. The goal is to increase sharpness in the forecasts, without compromising the validity of the probability statements.