Currents measured during the winter of 1985--1986 by four moorings on the inner Scotian Shelf are used to assess the hindcast skill of a three-dimensional circulation model forced by local wind stress and coastal sea level. The model is linear and barotropic. The integration scheme is based on the modification of the Galerkin spectral method proposed recently by Sheng and Thompson <1993>. The skill of the model is measured by the variance of the hindcast errors divided by the variance of the observations (henceforth &ggr;2). The model is most effective within 30 km of shore (0.51≤&ggr;2≤0.66). At the mooring in the Nova Scotia Current, a surface intensified southwestward jet with its center approximately 50 km from shore, the model fails to capture the bulk of the variance at current meters within 70 m of the surface (0.81≤&ggr;2≤0.87). The skill of the model is lowest at the mooring 65 km from shore (0.92≤&ggr;2≤0.94). To put these measures of skill into perspective, the currents are hindcast using a linear statistical model with the same inputs as the circulation model. The statistical model is optimal in the sense no other linear model with these inputs can achieve a lower hindcast error variance. For half of the current meters within 30 km of shore and two in the Nova Scotia Current the skill of the circulation model is not significantly lower than that of the statistical model. The largest discrepancies in the skill of the two models are found at the offshore mooring. We show that the suboptimal performance of the circulation model is due in part to the assumption of a spatially uniform wind field. We speculate that another contributing factor is the assumed form of the sea level profile along the open boundary that is upstream in the sense of coastal trapped wave propagation. Finally, the statistical model is used to estimate the increase in skill of the circulation model that may result from the assimilation of hydrographic data and additional coastal sea levels. ¿ 1997 American Geophysical Union