Assessing uncertainty in mesoscale numerical weather prediction models

Current methods of meteorological forecasting are largely deterministic and produce predictions with unknown levels of uncertainty. Specifically, forecast errors and uncertainties arise from an incomplete knowledge of initial conditions and from shortcomings in model physics. We report on the early stages of a project aimed at developing calibrated probabilistic forecasts based on operational ensembles of mesoscale numerical weather predictions. Our approach builds on the general ideas of Bayesian model averaging, conditionally heteroscedastic regression, and simulation-enhanced ensembles. Joint work with Fadoua Balabdoui, Yulia Gel and Anton Westveld.