Seasonal skill and predictability of ECMWF PROVOST ensembles

Abstract

AbstractVariations in seasonal‐forecasting skill and predictability during the 15 years (1979–93) of the European Centre for Medium‐Range Weather Forecasts (ECMWF) re‐analysis (ERA), have been studied using 120‐day ensemble integrations of the ECMWF model. These integrations form part of the European Union PROVOST (PRediction Of climate Variations On Seasonal to interannual Time‐scales) project. Observed sea surface temperatures (SSTs) were updated daily at the model's lower boundary. Two major and three moderate El Niño Southern Oscillation (ENSO) events occurred during the ERA period. Results are interpreted as giving an upper bound on the predictive skill of a coupled ocean‐atmosphere system as a function of season, location, and state of ENSO.The model systematic error was found to be comparable with a typical amplitude of interannual variation. When standardized by the corresponding ERA anomaly variance, systematic error appears to be largest in boreal spring in the northern extratropics, and in boreal summer in the tropics.Ensemble‐mean skill scores were found to be positive overall. Apart from the northern winter season, the ensemble‐mean skill for months 2–4 drops significantly when compared with months 1–3. The interannual variation of skill scores is much larger for the European region than for the hemispheric domain. Over the northern hemisphere, skill is much higher when only ENSO years are considered; for Europe, the enhancement in skill for ENSO years is much weaker.Estimates of intrinsic predictability were made for each year of the dataset. These estimates, defined both by a t‐test and variance ratio, indicate generally high predictability in years when ENSO was strong. Apart from northern winter, the predictability estimates also showed a systematic drop between months 1–3 and months 2–4. It is therefore concluded that the fall in skill scores between months 1–3 and 2–4 indicates more a weakening of the impact of initial conditions (ICs) than, say, an increase in the effects of model error. In order to study this further, the relative impacts of SSTs and ICs, including land surface ICs, on interannual variation of precipitation have been examined in an additional set of experiments. Overall, SSTs have a dominant role, though the impact of ICs is not negligible.The predictability of tropical and extratropical precipitation is also discussed. The level of skill for precipitation in the extratropics is generally lower than in the tropics. However, within the tropics there are regions where the precipitation exhibits chaotic behaviour and is correspondingly less predictable.