Positive temperature skill has been found in decadal predictions over the North Atlantic. This skill can be linked to skilful predictions of the Atlantic multi-decadal variability (AMV). The ENSEMBLES multi-model, the DePreSys perturbed-parameter and the EC-Earth decadal ensemble re-forecasts are used to illustrate the skill of indices that characterize the North Atlantic climate variability. All systems yield skilful predictions, although in the first few years the forecast quality of the initialized predictions is significantly better than the skill of simple statistical and uninitialized dynamical predictions. Besides, the perturbed-parameter and multi-model ensembles have better forecast quality than the single-model systems. The AMV skill is interpreted in terms of its regional distribution, where the subpolar and subtropical regions show a different behaviour, and the role of the correct representation of the changing radiative forcing. An assessment of the impact of the start date frequency is carried out using DePreSys and illustrates the poor sampling of the CMIP5 decadal prediction experimental setup. The EC-Earth system is used to illustrate the impact of different strategies to generate enough ensemble spread that addresses prediction uncertainty.