Subsampling members in a seasonal forecast ensemble

Climate Services based on seasonal forecasts are a powerful tool for adaptation in a changing climate and they attract growing interest from different sectors. However, operational seasonal forecasts have traditionally been issued following a subjective procedure, combining information from different sources, such as observation, empirical and dynamical models. Although it adds value by incorporating expert knowledge, the subjective procedure usually results in graphic products, with limited traceability, and not suitable for objective skill assessment or coupling sectoral applications. Identifying this issue, WMO encourages Regional Climate Centers and RCOFs to develop an objective procedure. The purpose is to increase the reliability of our results and to provide the basis for future climate services. With that aim, we explored ways of developing an objective approach that adds value to raw model forecasts in the Mediterranean region.
As is usually accepted, the starting point is a multimodel ensemble, which in our case combines seven Copernicus seasonal forecast models, hoping to minimize the weaknesses of individual models. The work focuses on looking for ways of subsampling the ensemble data based on comparing observational patterns with the evolution of ensemble members at the beginning of the period. Therefore, we did not use the latest model run, choosing instead earlier initializations and applying techniques such as cluster analysis or subsampling a fixed number of members to select those that were closer to reality.
First, we performed a cluster analysis to the ensemble forecast for winter (DJF) 2023-2024 and chose the cluster which best predicted the values of precipitation in October, which would be the last month with complete data when producing the winter seasonal forecast. However, we did not find a significant increase in skill in the Mediterranean region, possibly due to the great differences in cluster population between each year of the hindcast.
Then, we tested an alternative method by selecting a fixed number of members for the forecast and each year of the hindcast. We subsampled the group of members which best predicted precipitation in October and found a significant increase in skill in certain areas. However, there were not consistent improvements along the whole region, with some areas showing lower skill.
A comparison of the methodology using different model runs was conducted, finding better performance for the September run.