The impact of ensemble forecasts on predictability

Abstract

AbstractAn estimate of the mean effect of ensemble‐averaging on forecast skill, under idealized ‘perfect model’ conditions, is obtained from a set of eight independent 50‐day winter ensemble forecast experiments made with a hemispheric version of the Meteorological Office (UKMO) 5‐level general circulation model. Each ensemble forecast consisted of seven individual integrations. Initial conditions for these were obtained by adding spatially correlated perturbations to a given wintertime analysis, and a further integration created in the same manner was used to represent nature, giving the perfect model approach.The ensemble‐mean forecast shows a clear improvement in amplitude and phase skill compared with individual forecasts, the period of significant predictability for daily fields being increased by 50%. The improvement in skill is consistent with simple theoretical estimates based on the perfect model assumption. These calculations are used to deduce how ensemble‐mean forecast skill should vary with the size of ensemble. The superiority of the ensemble‐mean is maintained when forecasts are spatially smoothed or time‐averaged.The spread of an ensemble distribution can in principle give an a priori indication of forecast skill. A moderate level of correlation between ensemble spread and the forecast skill of the ensemble‐mean is found on the hemispheric scale.The extent to which the potential benefits of ensemble forecasting may be achieved in reality depends on the model's practical forecast skill. Since the practical skill of the 5‐level model is rather low, an ensemble‐mean forecast is on average no better than an individual forecast up to the normal limit of deterministic predictability. However, in four experiments where the individual forecasts show skill beyond this point, the ensemble‐mean forecast does give increased skill.Spatial variations in both the practical and perfect model skills of an ensemble‐mean anomaly field are found to be related to corresponding variations in the statistical significance of the anomaly field. For example, the average perfect model skill, in regions where the ensemble‐mean anomaly is significantly different from zero, exceeds the full field skill in all experiments for forecast days 1–15, and in all but two cases for days 16–30.