Global Model selection for evaluation of Climate Change projections in the Eastern Tropical Pacific Seascape

Two methods for selecting a subset of simulations and/or general circulation models (GCMs) from a set of 30 available simulations are compared: 1) Selecting the models based on their performance on reproducing 20th century climate, and 2) random sampling. In the first case, it was found that the performance methodology is very sensitive to the type and number of metrics used to rank the models and therefore the results are not robust to these conditions. In general, including more models in a multi-model ensemble according to their rank (of skill in reproducing 20th century climate) results in an increase in the multi-model skill up to a certain point and then the inclusion of more models degrades the skill of the multi-model ensemble. In a similar fashion when the models are introduced in the ensemble at random, there is a point where the inclusion of more models does not change significantly the skill of the multi-model ensemble. For precipitation the subset of models that produces the maximum skill in reproducing 20th century climate also showed some skill in reproducing the climate change projections of the multi-model ensemble of all simulations. For temperature, more models/simulations are needed to be included in the ensemble (at the expense of a decrease in the skill of reproducing the climate of the 20th century for the selection based on their ranks). For precipitation and temperature the use of 7 simulations out of 30 resulted in the maximum skill for both approaches to introduce the models