A statistical approach to correct ensemble seasonal forecasts is formulated based on the regression of the forecast model's leading forced singular value decomposition (SVD) patterns and the observed 500 hPa geopotential height. This technique is applied to the winter forecasts from two general circulation models (GCMs). The performance of the corrected forecasts is assessed by comparing their cross-validated skill with that of the original GCM ensemble mean forecasts. We are particularly interested in the forecast skill of the Pacific/North American (PNA) and the North Atlantic Oscillation (NAO). In the case of the PNA, the technique significantly improves the skill of the less skillful of the two models, and does not modify significantly that of the other model, which produces very good PNA forecasts even before the correction. For the NAO, the correction significantly improves the forecast skill of both models.