Subselection of seasonal ensemble precipitation predictions for East Africa

The Greater Horn of Africa (GHA) is highly vulnerable to climate and weather hazards such as drought, heat waves, and floods. There is a need for accurate seasonal forecasts to prepare for risks (such as crop failure and reduced grazing opportunities) and take advantage of favorable conditions (rains arrive on time and where they are needed) when they arise. As such, information at finer spatial scales than current state-of-the-art global prediction models can provide is needed. Dynamical downscaling is one method employed to obtain information at finer scales. However, providers of seasonal forecasts over the GHA are hampered by limited computational resources and time constraints that restrict the number of global model ensemble members that can be downscaled. Some ensemble subselection criteria must be employed. Currently, providers take an uninformed (or random) approach. Specifically, forecasters simply take the first ensemble member of the global model seasonal forecast ensemble. However, recent work, focused on decadal prediction, has shown that subselecting global model ensemble members in an informed way, that is, according to their ability to reproduce key features of the climate system, results in improved predictions. This emerges from the fact that the climate system is likely more predictable than our models would have us believe. Seeing an opportunity for improvement, we apply the same thinking to the seasonal context and assess several procedures for subselecting ensemble members from seasonal predictions with exchangeable members. Such informed subselections have the potential to take advantage of information in an ensemble of global simulations that might be missed by random selection. Three subselection methods are investigated, with a focus on seasonal predictions for rainfall over GHA. We demonstrate that informed subselection leads to systematically higher skill than random subselection. We find that (1) for small subsample sizes, such as would be chosen for dynamical downscaling and/or downstream impact modeling, informed subselection nearly always outperforms random subselection, (2) subselecting based on well-known teleconnections benefits those seasons in which such pathways are active, such as OND and JJAS, and (3) k$$ k $$-means subselection outperforms random selection for small ensemble sizes throughout all seasons, including the notoriously difficult to predict MAM season. These techniques require only input that is available at the time of the forecast release and are easy to apply operationally.