HR: 1340h
AN: H13E-1620    [Abstracts]
TI: Can multi-model combination really enhance the prediction skill of probabilistic ensemble forecasts?
AU: * Weigel, A P
EM: andreas.weigel@meteoswiss.ch
AF: Federal Office of Meteorology and Climatology MeteoSwiss, Kraehbuehlstrasse 58, P.O. Box 514, Zurich, 8044, Switzerland
AU: Liniger, M A
EM: mark.liniger@meteoswiss.ch
AF: Federal Office of Meteorology and Climatology MeteoSwiss, Kraehbuehlstrasse 58, P.O. Box 514, Zurich, 8044, Switzerland
AU: Appenzeller, C
EM: christof.appenzeller@meteoswiss.ch
AF: Federal Office of Meteorology and Climatology MeteoSwiss, Kraehbuehlstrasse 58, P.O. Box 514, Zurich, 8044, Switzerland
AB: The success of multi-model ensemble combination has been demonstrated in many studies. However, given that a multi-model contains information of all participating models, including the less skillful ones, the question remains as to why, and under which conditions, a multi-model can outperform the best participating single model. It is the aim of this presentation to resolve this supposed paradox. The study is based on a synthetic forecast generator, allowing the generation of perfectly calibrated single model ensembles of any size and skill. Additionally, the degree of ensemble underdispersion (i.e. overconfidence) can be prescribed. Multi-model ensembles are then constructed from both weighted and unweighted averages of these single model ensembles. Skill is measured with the discrete ranked probability skill score (Weigel et al. 2007a), which is favorable in the context of multi-model studies since it is insensitive to changing ensemble size. Applying this toy model, systematic model-combination experiments are carried out. We evaluate how multi- model performance depends on skill and overconfidence of the participating single models. It turns out that multi- model ensembles can indeed outperform a "best model approach", but only if the single model ensembles are overconfident. The reason is that multi-model combination reduces overconfidence, i.e. that ensemble spread is widened while the average ensemble mean error is reduced. This implies a net gain in prediction skill, because probabilistic skill scores penalize overconfidence. It is under these conditions that even the addition of a consistently poorer model can enhance multi-model skill, as long as the model's poor performance is due to high overconfidence rather than due to low potential predictability. Using real seasonal forecasts from the DEMETER data set, it is shown that the conclusions drawn from the toy model experiments equally hold in a real multi-model ensemble prediction system. References: A.P. Weigel et al., 2007a, Mon. Wea. Rev., 135, 2778-2785 A.P. Weigel at al., 2007b: Can multi-model combination really enhance the prediction skill of probabilistic ensemble forecasts? Quart. J. Roy. Met. Soc. Under review
DE: 3309 Climatology (1616, 1620, 3305, 4215, 8408)
DE: 3333 Model calibration (1846)
DE: 3337 Global climate models (1626, 4928)
SC: Hydrology [H]
MN: 2007 Fall Meeting