Can metric-based approaches really improve multi-model climate projections? The case of summer temperature change in France

The multi-model ensemble mean is generally used as a default approach to estimate climate change signals, based on the implicit hypothesis that all models provide equally credible projections. As this hypothesis is unlikely to be true, it is in theory possible to obtain more realistic projections by giving more weight to more realistic models according to a relevant metric, if such a metric exists. This alternative approach however raises many methodological issues. In this study, a methodological framework based on a perfect model approach is described. It is intended to provide some useful elements of answer to these methodological issues. The basic idea is to take a random climate model and treat it as if it were the truth (or "synthetic observations"). Then, all the other members from the multi-model ensemble are used to derive thanks to a metric-based approach a posterior estimate of the future change, based on the synthetic observation of the metric. This posterior estimate can be compared to the synthetic observation of future change to evaluate the skill of the approach. This general framework is applied to future summer temperature change in France. A process-based metric, related to cloud-temperature interactions is tested, with different simple statistical methods to combine multiple model results (e.g. weighted average, model selection, regression.) Except in presence of large observational errors in the metric, metric-based methods using the metric related to cloud temperature interactions generally lead to large reductions of errors compared to the ensemble mean, but the sensitivity to methodological choices is important .