This paper investigates the performance of soundings generated from the National Centers for Environmental Prediction’s Short Range Ensemble Forecast (NCEP SREF). The NCEP SREF is an operational ensemble forecast model with 15 members. Rank histograms are used as the primary tool to investigate consistent bias problems as well as ensemble dispersal. For the period spanning 1 May 2003 and 19 July 2003, nine different locations scattered about the continental U.S. are validated with rawinsonde data. Ensembles modified by a lagged bias correction and ensembles modified by both a lagged bias correction and the addition of observational errors are considered.

Rank histograms constructed from the unmodified ensemble imply either severe bias problems in the ensemble or a significantly underdispersed ensemble, depending on the variable examined, forecast time, pressure level, and location. Because forecasts between the different locations are poorly correlated, the assumption of independence is acceptable and rank histograms for each location are merged into combined rank histograms for all cities for a given variable, forecast time, pressure level, and location to produce adequate sample sizes. Combined rank histograms constructed from the bias corrected ensemble are U-shaped, which may be caused either by an under-dispersed ensemble, a non-homogeneous bias structures, or observational errors. However, including observational errors with the bias correction often results in uniform or, occasionally over-dispersed, rank histograms. Analysis of other factors, including non-homogeneous biases of the ensemble, is shown to help understand the combined rank histograms. Without the bias correction, this ensemble if of limited utility, but the lagged bias correction greatly enhances the ensemble performance.