It has long been known that the information content of weather forecasts extends beyond the model-variablesthat share the same name as the forecast target-variables of interest. Traditional model output statistics(MOS) algorithms extract information from any model-variable deemed relevant to estimating a given target-variable, especially when the “corresponding” model-variable, taken at face value, forecasts “poorly”.Mathematically, this form of MOS can be seen as adopting a “projection” operator between model-state spaceand observations that is more complex than the identity operator. Ensemble forecasts allow the introduction of a new twist. Typically, one treats each individual ensemble member as a viable scenario, projecting it intoobservation space as a (dressed) forecast, and then combining all ensemble members; an alternativeapproach is to condition the probability forecast of the target value upon properties of the joint distribution of allthe ensemble members (in a potentially multi-model ensemble). Thus eMOS goes beyond MOS in that it notonly aims to locate information in each individual model run, but also considers the ensemble as a whole, notmerely as a collection of scenarios. The approach is illustrated in precipitation forecasts, and more generalinterpretations relevant to THORPEX's core aims are noted.
[1]
Leonard A. Smith,et al.
Evaluating Probabilistic Forecasts Using Information Theory
,
2002
.
[2]
Robert Tibshirani,et al.
Bootstrap Methods for Standard Errors, Confidence Intervals, and Other Measures of Statistical Accuracy
,
1986
.
[3]
H. Glahn,et al.
The Use of Model Output Statistics (MOS) in Objective Weather Forecasting
,
1972
.
[4]
Leonard A. Smith,et al.
A forecast reliability index from ensembles: a comparison of methods
,
2001
.