Assessing the Ensemble Spread–Error Relationship

Abstract The potential ability of an ensemble prediction system (EPS) to represent its own varying forecast error provides strong motivation to produce an EPS over a less expensive deterministic forecast. Traditionally, this ability has been assessed by correlating the realized forecast error with the ensemble's dispersion. This paper revisits the limitations of the skill–spread correlation, but uses aspects of the correlation to introduce two metrics to assess an EPS's capacity to provide a reliable likelihood of its own error. Using a perfect EPS, skill–spread correlation is shown to be limited by its dependence on how “skill” and “spread” are defined and, perhaps most fatally, by its inability to distill the skill–spread reliability from the stability properties of the physical system being modeled. Building from this, it is argued there are two aspects of an ensemble's dispersion that should be assessed. First, is there enough variability in the dispersion to justify the expense of the EPS? The factor that controls the theoretical upper limit of the spread–error correlation can be useful in diagnosing this. Second, does the variable dispersion of an ensemble relate to a variable expectation of the forecast error? Representing the spread–error correlation in relation to its theoretical limit can provide a simple diagnostic of this attribute. A context for these concepts is provided by assessing two operational ensembles: western U.S. temperature forecasts and Brahmaputra River flow before and after postprocessing. It is shown that “skill–spread” reliability can be improved by postprocessing to that of a perfect EPS, but at the cost of the potential information content of the EPS's variable dispersion.