Probabilistic Forecasting with Nmme

Abstract:

The North American Multi-Model Ensemble (NMME) forecasting system has been continuously producing seasonal forecasts since August, 2011. The NMME, with its suite of diverse models, provides a valuable opportunity for characterizing forecast confidence using probabilistic forecasts. The current experimental probabilistic forecast product presents the most likely tercile-based class for the monthly mean value, out of “above normal”, “near normal”, or “below normal”, using a non-parametric counting method to determine the probability of each class. A first-order bias correction is applied when the local model climatology is removed and replaced with the climatology from observations; the counting method also corrects for gross biases in distribution, such as the standard deviation. In preparation for the development of more precise calibration methods, the skill of the monthly-mean and three-month-mean probabilistic forecasts of 2 m surface temperature, precipitation rate, and sea-surface temperature is assessed using the 29-year (1982-2010) NMME hindcast database. Three configurations are considered: a full six-model NMME, CFSv2 alone (24 members), and a 24-member “mini-MME”, comprised of four members from each of the six models. Skill is assessed using the Brier skill score and ranked probability score; forecast reliability is also assessed. For 2 m surface temperature and SST, assessed in the hindcast study, it is found that probabilistic forecasts using the full NMME have higher skill and reliability than the “mini-MME”, which in turn have higher skill and reliability than the CFSv2 alone. Verification of the real-time probabilistic forecasts is also reviewed.