Two decadal prediction ensembles, based on the same climate model (IPSL-CM5A-LR) and the same surface nudging initialization strategy are analyzed and compared with a focus on upper-ocean variables in different regions of the globe. One ensemble consists of 3-member hindcasts launched every year since 1961 while the other ensemble benefits from 9 members but with start dates only every 5 years. Analysis includes anomaly correlation coefficients and root mean square errors computed against several reanalysis and gridded observational fields, as well as against the nudged simulation used to produce the hindcasts initial conditions. The last skill measure gives an upper limit of the predictability horizon one can expect in the forecast system, while the comparison with different datasets highlights uncertainty when assessing the actual skill. Results provide a potential prediction skill (verification against the nudged simulation) beyond the linear trend of the order of 10 years ahead at the global scale, but essentially associated with non-linear radiative forcings, in particular from volcanoes. At regional scale, we obtain 1 year in the tropical band, 10 years at midlatitudes in the North Atlantic and North Pacific, and 5 years at tropical latitudes in the North Atlantic, for both sea surface temperature (SST) and upper-ocean heat content. Actual prediction skill (verified against observational or reanalysis data) is overall more limited and less robust. Even so, large actual skill is found in the extratropical North Atlantic for SST and in the tropical to subtropical North Pacific for upper-ocean heat content. Results are analyzed with respect to the specific dynamics of the model and the way it is influenced by the nudging. The interplay between initialization and internal modes of variability is also analyzed for sea surface salinity. The study illustrates the importance of two key ingredients both necessary for the success of future coordinated decadal prediction exercises, a high frequency of start dates is needed to achieve robust statistical significance, and a large ensemble size is required to increase the signal to noise ratio.< Leer menos