Observational studies have established that significant correlations exist between Eurasian snow cover extent in October and the wintertime phase of the Arctic Oscillation (AO). A small number of modeling studies have also succeeded in reproducing these correlations when realistic or idealized snow cover is prescribed in the models. On the other hand, coupled GCMs with model-generated snow cover have not been able to reproduce the observed relationships. In line with these previous results, we demonstrate that hindcasts from the current version of the CFS (CFSv2) do not capture the observed atmospheric response to October snow cover variability, suggesting that some forecast skill may be gained from better representing the snow/AO relationship in the forecasts. Two general approaches are considered for improving the CFSv2 forecasts. First, the CFS model physics and dynamics could be improved to better capture the mechanisms linking snow with the AO. Towards this end, we identify weak links in how CFS model hindcasts represent the proposed snow/AO mechanism, including a failure in the model to generate sufficient snow cover variability. Second, the large ensemble of lagged seasonal CFS forecasts could be used more judiciously to emphasize ensemble members that best represent the expected atmospheric response to Eurasian snow cover. We use the CFSv2 seasonal hindcasts to identify ensemble members from each year that correctly capture various stages of the atmospheric response to snow, and examine whether these members show any additional skill in forecasting the wintertime AO.