The response of numerical weather prediction systems to fgge level iib data. Part II: Forecast verifications and implications for predictability

The purpose of our two‐part study is to assess the importance of the differences between three independent analyses of the same FGGE level IIb observational dataset. Part I is concerned with the nature and origin of the differences among the analyses. Part II of the study is concerned with the implications of the analysis differences for forecast skill, and for estimates of predictability. Our experimental material is a set of forecasts by two models from the three ensembles of analyses. the energetics of the analyses are different, and this is reflected in differences in the energetics of the forecasts based on the analyses. A study of the objective forecast verifications shows that there are clear differences in the forecast skill of the two models we used. the verifications also show a marked sensitivity to the choice of initial data. We studied in some detail how forecasts made with the same model from different analyses diverge from one another, with particular attention to three aspects: (i) the growth rate at small amplitude; (ii) the time taken for the differences to reach an asymptotic level; (iii) the amplitude of this level relative to that of persistence. the results for small amplitude growth rates are consistent with earlier results in giving a doubling time of 2·0 days in the height field in the day 1‐2 forecasts but with slower growth rates (doubling times of 2·6 days) in the wind field. the baroclinic waves show faster growth, with about 0·25 days shorter doubling times. the forecast differences reach an asymptotic level more rapidly in this than in earlier investigations. This suggests that earlier estimates of the purely dynamical potential for predictability of instantaneous weather patterns may be overoptimistic. We combined the measures of forecast skill and forecast divergence in an error budget which separated the contributions to forecast error arising from model error and analysis error, subject to some simplifying assumptions. the results show that between day 2 and day 5 the model error grows linearly in time and the analysis error grows exponentially. the model is the main source of forecast error in this time range. the analysis error is the main source of forecast error in the short‐range forecasts, and is a substantial contributor after about day 5. These results are very sensitive to the magnitude of the initial analysis errors. Taken together with the synoptic results of part I, these statistical results emphasize the impo