Some recent studies have examined the utility of a dynamic MOS approach (e.g. MAO et al., 1997) in which statistical relationships are built using only recent weather observations and model forecasts. The dynamic approach has the advantage that it can adapt to changes in the model's configuration automatically, as well as providing relationships which are more relevant to the ongoing weather regime. On the other hand, the use of limited datasets upon which the regressions are built, tends to occasionally produce erronous relationships that can yield large errors when applied.
To alleive the latter effect, a ensemble approach to dynamic-MOS has been tested using the ETA model output from the fall of 2000. In this experiment, a large set of semi-independent regression equations (as opposed to just the leading equation) were determined for each forecast variable. The resulting regression set was then applied to each forecast to produce an ensemble of forecast estimates. From the set of ensemble predictions, outlier forecasts resulting from unstable regressions could easily be identified and rejected from the set. From the remaining members, a final median forecast was extracted. The skill of the median forecast was compared to the skill of the single best regression equation. The ensemble approach was found to be considerably superior to the single leading equation in both reducing the number of large-error forecasts associated with statistically unstable regressions, but also in the mean error of the forecasts. The results were consistent across forecasts of a number of weather parameters including temperature, probability of precipitation, winds and clouds. Approximately 100 forecast locations were used in the experiments.