We document two stylised facts of US short- and long-term interest rate data incompatible with the pure expectations hypothesis: Relatively slow adjustment to long-run relations and low contemporaneous correlation. We construct a small structural model which features three types of randomness: While a persistent monetary policy shock implies immediate identical reactions through the term structure, both a transitory policy shock and an autocorrelated risk premium allow for the sustained decoupling observed in the data. Indeed, we find important impacts and persistence of risk premia and a decomposition of policy shocks judging a larger part as transitory the longer the investment horizon.