Abstract
A series of one-month forecasts were carried out for eight January cases, using a particular prediction model and prescribing climatological sea-surface temperature as the boundary condition. Each forecast is a stochastic prediction that consists of three individual integrations. These forecasts start with observed initial conditions derived from datasets of three meteorological centers. The forecast skill was assessed with respect to time means of variables based on the ensemble average of three forecasts. The time or space filter is essential to suppress unpredictable components of atmospheric variabilities and thereby to make an attempt at extending the limit of predictability. The circulation patterns of the three individual integrations tend to be similar to each other on the one-month time scale, implying that forecasts for the 10 day (or 20 day) means are not fully stochastic. The overall results indicate that the 10-day mean height prognoses resemble observations very well in the first ten days, and then start to lose similarity to real states, and yet there is some recognizable skill in the last ten days of the month. The main interests in this study are the feasibility of one-month forecasts, the adequacy of initial conditions produced by a particular data assimilation, and the growth of stochastic uncertainty. An outstanding problem turns out to be a considerable degree of systematic error included in the prediction model, which is now known to be “climate drift.” Forecast errors are largely due to the model's systematic bias. Thus, forecast skill scores are substantially raised if the final prognoses are adjusted for the model's known climatic drift.
