Survey data on interest rate expectations are used to separate the forward interest rate into an expected future rate and a term premium. These components are used to test separately two competing alternative hypotheses in tests of the term structure: that the expectations hypothesis does not hold, and that expected future long rates over- or underreact. to changes in short rates. While the spread consistently fails to predict future interest rate changes, we find that the nature of this failure is different, for short versus long maturities. For short maturities, expected future rates are rational forecasts. The poor predictions of the spread can therefore be attributed to variation in term premia. For longer-term bonds, however, we are unable to reject the expectations theory, in that a steeper yield curve reflects a one-for-one increase in expected future long rates. Here the perverse predictions of the spread reflect investors' failure to raise sufficiently their expectations of future long rates when the short rate rises. We confirm earlier findings that bond rates underreact to short rate changes, but now this result cannot be attributed to the term premium