We use a Bayesian vector autoregression with stochastic volatility to forecast government bond yields. We form the conjugate prior from a no-arbitrage affine term structure model. The model improves on the accuracy of point and density forecasts from a no-change random walk and an affine term structure model with stochastic volatility. Our proposed approach may succeed by relaxing the no-arbitrage affine term structure model's requirements that yields obey a factor structure and that the factors follow a Markov process. In the term structure model, its cross-equation no-arbitrage restrictions on the factor loadings appear to play a marginal role in forecasting gains.