Weighting climate model ensembles for mean and variance estimates

Abstract

Projections based on climate model ensembles commonly assume that each individual model simulation is of equal value. When combining simulations to estimate the mean and variance of quantities of interest, they are typically unweighted. Exceptions to this approach usually fall into two categories. First, ensembles may be pared down by removing either poorly performing model simulations or model simulations that are perceived to add little additional information, typically where multiple simulations have come from the same model. Second, weighting methodologies, usually based on model performance differences, may be applied, and lead to some improvement in the projected mean. Here we compare the effect of three different weighting techniques—simple averaging, performance based weighting, and weighting that accounts for model dependence—on three ensembles generated by different approaches to model perturbation. We examine the effect of each weighting technique on both the ensemble mean and variance. For comparison, we also consider the effect on the CMIP5 ensemble. While performance weighting is shown to improve the estimate of the mean, it does not appear to improve estimates of ensemble variance, and may in fact degrade them. In contrast, the model independence weighting approach appears to improve both the ensemble mean and the variance in all ensembles.