How representative is the spread of climate projections from the 5 CMIP5 GCMs used in ISI-MIP?

In many studies that use data from Coupled Model Inter-comparisons Project Five (CMIP5) the large number of models included prohibits the use of data from all models. Studies based on small subsets of CMIP5 may therefore exclude a significant fraction of the plausible range of future climate changes. In the Inter-sectoral Impact Model Inter-comparison Project (ISI-MIP), a subset of five CMIP5 models is used in global climate impact studies. We find that the fraction of the full range of future projections captured across different regions and seasons by the ISI-MIP subset varies from 0.5 to 0.9 for temperature (median 0.75) and 0.3 to 0.8 for precipitation (median 0.55). The implication of this is that for many regions and seasons, this subset can be expected to underestimate both the total uncertainty in future climate impact, and the proportion of total uncertainty that is attributable to the use of different GCMs. The fraction of climate model uncertainty sampled could be improved only marginally by using a strategically selected global optimal subset of 5. In order to capture > 0.8 (0.9) of the range in >75% or more of regions and seasons, at least 13 (20) models would be required. However, the use of regionally optimised subsets could significantly improve the range of regional precipitation changes captured by a subset of restricted size. The results of impact studies based on subsets of CMIP5 could be presented alongside information about how well the subset captures uncertainties in future climate to aid interpretation of impact uncertainty ranges.