Decadal Variability in Ensemble Projections

A simple method to rank multi-model ensemble members of CMIP3 simulations by their representation of phases of decadal oscillations is introduced. A period of 22 years (1979-2000) from the 20th century simulations is used to generate ensemble projections of trends for an 11-year (2001-2011) lead time for the SRES A1B scenario. Although greenhouse-gas forcing is identical for all 20th century simulations, the phases of decadal oscillations are quite different. Thus, the suggested minimum requirements for a simple selection criterion for adequate ensemble members are that (a) trends in high-, mid-, and low-latitude zones need to be treated separately and (b) information about the state of teleconnections between the zones needs to be included, when projecting decadal variability and trends in climate. The new method indicates that half (19 out of 38) ensemble members retain their rank when each GCM is treated separately without any assumptions of which model might be superior. Thus, the overall ensemble size can be reduced without a large loss of information but with a greatly reduced range of uncertainty, when only the least well performing ensemble members of each GCM are omitted in an 11-year projection.