I propose a strategy for forecasting the term structure of interest rates that may produce significant gains in predictive accuracy. The key idea is to use the restrictions implied by Gaussian, no-arbitrage, affine term structure models on a vector autoregression as prior information instead of imposing the restrictions dogmatically. This allows us to account for possible model misspecification. We use the proposed method to forecast a system of five U.S. yields up to 12 months ahead, and we find it provides significant gains in forecast accuracy.