Multi-model seasonal hindcasts over the Euro-Atlantic: skill scores and dynamic features

A group of multi-model seasonal hindcast experiments for Europe are verifie d and analysed using as reference the European Centre for Medium-range Weat her Forecasts re-analysis and Xie and Arkin precipitation data. Each model' s systematic error is described. Hindcast skill scores are evaluated comput ing anomaly correlation coefficients. The values of the scores are highly d ependent on the variable, on the region and on the season considered. Score s are particularly low over Europe for all seasons, reaching their maximum during winter. The presence of occasional poor hindcasts affects the multi- model ensemble results substantially. In order to see whether or not the sk ill inconsistencies are linked to the model's inability to forecast the evo lution of some particular patterns, hindcast skill scores are computed for the four large-scale patterns which explain most of the observed low-freque ncy variance over the Euro-Atlantic region, during winter. These scores are strongly dependent on the pattern. Multi-model hindcasts are better than t he best single model hindcast only for those patterns for which the model b iases cancel each other. In all cases, substantially better multi-model hin dcast scores for all patterns can be obtained by combining the four model r esults using optimal weights, computed for each model and for each pattern with the technique suggested by Thompson. All results show no dependence on the ensemble size considered. Skill scores are finally computed for severa l indices, which measure the variability of selected weather regimes over E urope. Regimes scores are consistent with the scores obtained for the corre spondent Euro-Atlantic EOF patterns? and it is shown that the removal of ea ch model's systematic error from its hindcasts does not improve the final r egime hindcast skill.