Francisco J. Doblas-Reyes, Renate Hagedorn, Magdalena Alonso
Balmaseda and Tim N. Palmer
ECMWF, Shinfield Park
RG2 9AX, Reading, UK
The forecast quality of the DEMETER seasonal multi-model ensemble predictions
has been assessed over the North Atlantic/Europe region. The analysis shows
that singlemodel ensembles have positive though low skill in almost every
season. The skill decreases with lead time, although the agreement in skill
between the models increases. This may be due to a more relevant role of the
decadal variability versus the initial conditions at longer lead times. The simple
multi-model constructed from equally weighted single-model ensembles proves to
have an average forecast quality superior to any of the single models and to
persistence predictions, in particular in a probabilistic framework. In
spite of the relatively low skill, a perfect model approach indicates that
potential skill is much higher than actual skill. However, potential skill
estimates depend strongly on the model used. It is suggested that a multi-model
system may provide more realistic predictability estimates. The relevance of
these results for operational seasonal forecasts and its users is discussed.