Extensions of the Eta Model forecasts at NCEP to 60 and then 84 h have enabled a new look at the limited area predictability: what skill additional to that of the "driver" global model can be achieved by a limited-area model (LAM), and for how long? The traditional view is that "the contamination at the lateral boundaries ... limits the operational usefulness of the LAM beyond some forecast time range". What exactly is that time range? Note that this would also represent the limit of the usefulness of the LAM ensemble approach, given that usefulness of the control is a prerequisite for the usefulness of the ensemble as a whole.

Looking for some answers, we have examined the trend of the Eta vs Avn precipitation scores, the rms fits to raobs of the two models as a function of time, and the errors of these models in forecasting the position of major lows. Note that in these comparisons the Eta, getting its lateral boundary condition from the Avn runs initialized 6 h before, is facing a handicap additional to those of the standard lateral boundary errors.

In the two months May-June 2001 with rms scores available to 84 h, we find little evidence of the Eta mid- and upper-tropospheric scores falling systematically behind those of the Avn as the forecast time range is approaching 84 h. Regarding the position errors in forecasting major lows, in winter, when the impact of the lateral boundary influence is the fastest, the Eta is found at 60 h to be significantly more accurate than the Avn in placing major lows over our verification region -- approximately the United States east of the Rockies.

We interpret these results as indicating that a LAM, provided its domain is large enough and it has specific advantages over its driver global model such as higher resolution, better treatment of major topography, etc., is able not only of "downscaling" but also of improving upon large-scale features at time ranges when its verification domain is well affected by the lateral boundary errors. In other words, not only downscaling but also "upscaling" can and in our case would have to be taking place. If compensation for the lateral boundary errors is of a sufficiently significant impact, we envisage that a LAM could indefinitely stay competitive or, in principle, even generally superior to its driver global model over the region of interest.