When considering the use of output from short-range ensemble prediction (SREF) systems in operational settings, an often raised question is that of which individual ensemble forecast member is considered "best". We argue from a theoretical viewpoint that it is inappropriate to seek the answer to this question, specifically with regard to the concept of "return to skill." The theoretical arguments are then applied to a set of SREF cases collected in Spring 2003 by the Storm Prediction Center during an experiment designed to test the operational use of SREF products for predicting severe storms. Using a variety objectively defined approaches, most of the examples reveal tremendous spatial and temporal inhomogeneity in the individual member identified as "best". One case does reveal a consistently selected "preferred" solution, but we show that its initial perturbations may not be appropriate. Implications for the use of SREF products in operational settings are discussed.