Why do ensemble prediction systems so often possess an apparent under-dispersiveness?

Ake Johansson
Swedish Meteorological and Hydrological Institute
2:30 pm in Room 2155

The relationship between spread and skill is an important indicator of the quality of an ensemble prediction system (EPS). When the commonly used spread-skill condition is used to quantify whether an EPS is under- or over-dispersive, there has historically been a clear bias towards under-dispersiveness. Arguments are given that suggests that an alternative to the commonly used spread-skill condition is needed to properly address the true relationship between spread and skill. Such a relationship is derived and given the name U2I spread-skill condition. Its properties are described both analytically and geometrically and the relationship to the commonly used condition is demonstrated and discussed. It is argued that it would provide not only a more appropriate and sharper tool to assess the spread-skill relationship, but also that it indicates how to improve upon the present design of an EPS. In particular, the presently prevalent practice of centering the perturbed ensemble members around the control analysis is shown to be substantially detrimental to the performance of an EPS and therefore should be replaced by a methodology that constructs equally likely ensemble members with the same quality as the control member. A recent experiment with a limited area ensemble prediction system is used to quantify the differences in perceived quality that is obtained by employing the two different spread-skill conditions. The differences are found to be substantial.