Francisco J. Doblas-Reyes, Antje Weisheimer, Tim N. Palmer
ECMWF, Shinfield Park
RG2 9AX, Reading, UK
Michel Déqué
CNRM, Météo-France
42 Av. G. Coriolis, Toulouse, France
Noel Keenlyside
Leibniz-Institut für Meereswissenschaften
Düsternbrooker Weg 20, Kiel, Germany
Malcolm McVean, James M. Murphy, Doug Smith
Hadley Centre, Met Office
Fitzroy Road, Exeter, United Kingdom
Philippe Rogel
CERFACS
42 Av. G. Coriolis, Toulouse, France
The relative merits of three forecast systems addressing the impact of model uncertainty on seasonal/annual forecasts are described. One
system consists of a multi-model, whereas two other systems sample uncertainties by perturbing the parameterisation of reference models
through perturbed parameter and stochastic physics techniques. Ensemble re-forecasts over 1991 to 2001 were performed with coupled climate
models started from realistic initial conditions. Forecast quality varies between the systems due to the different strategies for sampling
uncertainties, but also to differences in initialisation methods and in the reference forecast system. Although the multi-model experiment has
an ensemble size larger than the other two experiments, most of the assessment was done using equally-sized ensembles. The three ensembles
show similar levels of skill: significant differences in performance typically range between 5 and 20%. However, a nine-member multi-model
shows better results for seasonal predictions with lead times shorter than five months, followed by the stochastic-physics and the
perturbed-parameter ensembles. Conversely, for seasonal predictions with lead times longer than four months, the perturbed-parameter ensemble
gives more often better results. Both the stochastic-physics and perturbed-parameter ensembles improve the reliability with respect to their
reference forecast systems, but not the discrimination ability. Annual-mean predictions showed lower forecast quality than seasonal
predictions, but a substantial number of cases had positive skill. Only small differences between the systems were found. The full multi-model
ensemble has improved forecast quality with respect to all other systems, mainly from the larger ensemble size for lead times longer than four
months and annual predictions.