Predicting the winter climate over Eurasia is challenging as both the initialized predictions and uninitialized climate projections show limited skill in reproducing observed variability on multi-annual to decadal timescales. This study presents a novel approach to constrain variability in projection simulations over Eurasia by exploiting the spatial patterns of teleconnection between the North Atlantic Oscillation (NAO) and the near-surface air temperature (SAT). The constraint evaluates simulated and observed NAO-temperature teleconnection patterns during 20 year windows prior to making predictions. The resulting top ranking members are used to make predictions for the next 10 years. We find here that the constrained ensemble is skillful in predicting winter SAT on multi-annual to decadal timescales over Eurasia. We also find that the constrained multi-annual predictions yield higher correlation skill compared to both the unconstrained ensemble and the initialized predictions, especially up to 5 year mean winter SAT predictions. Therefore, we argue that for the Eurasia region, the constraint could provide skillful winter SAT predictions on multi-annual timescales with minimal cost compared to the state of the art initialized decadal climate predictions for which added value from initialization wanes quickly after the first forecast year.