The ENSEMBLES multi-model and perturbed-parameter decadal re-forecasts are used to assess multi-year forecast quality for global-mean surface air temperature (SAT) and North Atlantic multi-decadal sea surface temperature variability (AMV). Two issues for near-term climate prediction, not discussed so far, are addressed with these two examples: the impact of the choice of the observational reference period, and of the number of years included in the forecast average. Taking into account only years when both observational and model data are available, instead of using the full record, to estimate observed climatologies produces systematically (although not statistically significantly different) higher ensemble-mean correlations and lower root mean square errors in all forecast systems. These differences are more apparent in the second half of the decadal prediction, which suggests an influence of non-stationary long-term trends. Also, as the forecast period averaged increases, the correlation for both global-mean SAT and AMV is generally higher. This also suggests an increasing role for the variable external forcing as when forecast period averaged increases, unpredictable internal variability is smoothed out. The results show that predicting El Niño-Southern Oscillation beyond one year is a hurdle for current global forecast systems, which explains the positive impact of the forecast period averaging. By comparing initialized and uninitialized re-forecasts, the skill assessment confirms that variations of the global-mean SAT are largely controlled by the prescribed variable external forcing. By contrast, the initialization improves the skill of the AMV during the first half of the forecast period. In an operational context, this would lead to improved predictions of the AMV from initializing internal climate fluctuations. The coherence between the multi-model and perturbed-parameter ensemble supports that conclusion for boreal summer and annual means, while the results show less consistency for boreal winter.