Achieving reliable and accurate storm forecasts remains a scientific challenge due to uncertainty in grasping the initial conditions (e.g., limited observations and ability to assimilate them), shortcomings in model physics (e.g., parameterization of subgrid-scale processes) and computational capabilities (e.g., resolution in space and time, processing speed), and limitations of our understanding of how nature works (e.g., missing processes that might be relevant). The forecast skill of observation-driven expert systems decreases rapidly with increasing lead-time, while numerical weather prediction models typically exhibit a limited forecast ability within the first few hours after initialization primarily due to spin-up problems.
Probabilistic prediction attempts to capture the forecast uncertainties by means of techniques that include ensemble forecasts (e.g., created by assembling time-lagged forecast runs, combining multiple parallel forecast runs, selecting multiple analogue forecasts from past runs, and/or creating an ensembles of diagnostics) and statistical post-processing (e.g., based on data mining and/or artificial intelligence approaches).