The basic premise of ensemble forecasting is that the distribution of ensemble members at any lead time and location can be considered as a probability forecast for the atmospheric variable considered. Some earlier studies suggested that it is only the first (mean), and perhaps the second (spread) moment of the distribution that carries all useful forecast information. The question whether any information beyond these basic distribution characteristics can be trusted, however, has not been addressed. In the present study the NCEP global ensemble forecast system, comprising of 10 forecasts run out to 16 days lead time twice a day (0000 UTC and 1200 UTC), will be evaluated in this respect. In particular, cases where the forecast distribution is bimodal, will be verified in a statistical sense. If the distribution of verifying analysis values corresponding to the bimodal forecast cases also exhibits bimodality, that is an indication that the bimodal forecast distributions are reliable and can be trusted in forecast applications. This would also be a clear indication that ensemble forecasts can in general provide more information than that given in the first (two) moment(s) of their distribution. The more general question of how much detail one should retain from an ensemble given the size (membership) of the ensemble will also be addressed.