The present study addresses the issue of optimal distribution of observations for short-to-medium range forecasting. An error estimate is obtained from the difference of two equally credible analyses of the same state. This error measure assumes that the differences of modern analyses would be negligibly small for a perfect observing system. We focus on a five day period leading to the east coast "superstorm" of March 1993 and use ECMWF and NCEP Reanalysis differences in order to estimate observation uncertainty. Two separate global predictions are made from the two Reanalyses. The predictions diverge with time, and the forecast differences, which become substantial after 4 days, provide an estimate of predictive sensitivity to initial state uncertainty. Several experiments are done to determine the benefits associated with targeting local regions in reducing forecast errors. The results show that local targeting reduces the errors, but not more substantially than targeting the external, global domain. The spatial scales responsible for most of the uncertainty growth correspond to wavenumbers 0-15 in the initial uncertainty field. The present results suggest that an observing system which resolves these relatively large scales would provide most of the forecast benefits of a fully global, high resolution observing system