Ensemble reduction is the task of selecting a subset of the members of a global ensemble prediction system (EPS) to specify the initial and boundary conditions for the integration of a limited‐area EPS. Cluster analysis is often used for this purpose, even if random member selection would be a legitimate approach as well. Clustering algorithms organize forecasts from different ensemble members into groups, based on the degree of similarity between selected forecast fields. Reduction is performed by choosing one representative member from each cluster. Ensemble reduction degrades forecast accuracy, measured by the continuous rank probability score. The degree of degradation depends primarily on the size of the reduced ensemble and becomes larger as the ensemble gets smaller. We estimate the loss of forecast accuracy caused by different ensemble reduction methods by comparing the probabilistic forecasts obtained from the 51‐member EPS run by ECMWF with those from several reduced ensembles. We show that different ensemble reduction methods cause marginally different loss of accuracy and that, generally, clustering methods are not significantly better at ensemble reduction than random sampling. Clustering typically results in reduced ensembles with significantly lower spread than both the parent ensemble and randomly defined subsets. The effectiveness of clustering depends on the forecast range and on the variables used to cluster the global ensemble members; not all meteorological parameters are equally good clustering variables. Clustering is most effective at ensemble reduction when it detects meaningful differences between the ensemble members. This is only possible at forecast ranges beyond about 3 days and when variables with a low degree of small‐scale spatial variability are used as object descriptors. The figure uses colour to display the result of a cluster analysis. Individual objects are the members of an ensemble forecast system and three clusters are detected. Ensemble reduction consists of choosing one representative member from each cluster. Ensemble reduction by clustering gives useful results only at forecast ranges longer than 3 days.