We compare the differences between the Chinese and U.S. bond risk premia. We find that the expectations hypothesis fails in the two bond markets: We identify the Chinese and U.S. bond time-varying risk premia by forecasting the corresponding excess return of n-year bond using the n-year forward rate and n-year forward spread, respectively. To focus on the systematical forecasts, we then combine the forward rates at different maturities as the return-forecast factors. Unlike the one-factor model introduced by Cochrane and Piazzesi (2005), a two-factor model including level- and slope-based factors explains significantly Chinese bond premia with R2 up to 68%. More importantly, the slope-based factor sharply improves the performance of test. The results are robust with respect to measurement errors, multicollinearity and small-sample biases. Out-of-sample tests show that, in recent years, the U.S. bond market changes drastically, and tends to be like the Chinese market. We use the empirical results to calibrate the parameters of affine model, and find that the differences of bond premia between the two markets are caused by the differences of dynamics of state variables and risk attitude of investor.