Post-processing of seasonal predictions – Case studies using the EUROSIP hindcast data base

Abstract. Seasonal predictions from climate models are increasingly invoked in various sectors like water management, energy and transport to cite a few. This study investigates the post-processing of the seasonal predictions of the EUROSIP multi-model system. The hindcasts comprise samples of 23 to 36 years and ensembles of 10 to 28 members depending on the 5 models included. Skill scores both deterministic and probabilistic are calculated in order to compare the impact of the post-processing and help selecting – if any – the multi- or single-model and the post-processing method best suited for a specific location, target season and lead-time. The presence of trends and the cross-validation setting add some complexity to the already heterogeneous database. This study focuses on six cases in Western Europe and the Mediterranean Region. The forecasts of three monthly averages of surface temperature and of sea mean sea level pressure are compared with the corresponding ERA Interim reanalysis data whereas the forecasts of precipitation are evaluated with the rain-gauge data from the Global Precipitation Climatology Centre.

The skills of seasonal predictions in the extra-tropics are limited and our results are no exception. There is a significant skill for the spring temperature forecast with model initiation in March for all but one case studies and the skill is extending to the initiation begin of February for Belgium. There is is also a significant skill for the summer temperature for the case studies in the Mediterranean region. For these area, the skill comes in large part from the global warming so that after having de-trended the data, a null improvement cannot be excluded. Autumn temperature in UK and in the Turkey is forecast with some skill as well as winter temperature in UK and Greece. Precipitation is even more difficult to forecast: the two spots where skill scores are significantly positive are Sweden and Greece during winter with initialisation on the first December. It has been shown that multi-model ensemble improve the skills in many cases and that taking into account the longest common period of hindcasts results in better and less uncertain skill scores. For all these cases, the post-processing method and the model or model combination resulting in the best skill score have been selected.