Abstract A detailed study of hemispheric forecast-error (f.e.) statistics for the operational Canadian large-scale forecast model was made in preparation for reparameterization of the correlation function used in the data assimilation step of the forecast cycle. The choice of an appropriate function type for geopotential height and temperature f.e. lag-correlation representations was a major concern. A number of possible functional representations, both isotropic and anisotropic, were considered, and the goodness-of-fit of the various candidate functions to correlations computed from observed f.e. data were compared. The following special form of the second-order autoregressive function was adopted as the function of choice: (1 + c|s|)e−c|s|, where |s| is geographic location separation. An examination of f.e. correlation structure on a regional basis revealed large differences from region to region. Latitudinal, pressure-level, and seasonal dependencies of the statistical structure of constant pressure-su...