Bond Yield Predictability and Estimation of Affine Term Structure Models

Recent studies by Dai and Singleton (2002), Duffee (2002), and Duarte (2004) show that affine term structure models that match the time variability of the expected returns of bond yields do not generate time variation in the volatility of interest rates. This failure indicates that affine models fail to match one stylized fact of the term structure of U.S. interest rates. However, in this paper I show that the empirical limitation can be solved by allowing a more flexible specification of the risk-premia. I find that the affine models perform poorly at short forecasting horizons, but perform very well at longer horizons. Some affine models produce more accurate out-of-sample forecast than the random walk and other benchmark models. My empirical work also shows that the Kalman filter has the ability to filter discrepancy in zero-coupon bond yields occurring from different choices of splines, whereas the factor inversion method does not. When applied to different spline techniques, the factor inversion method gives larger differences in in-sample and out-of-sample bond yield forecasts than the Kalman filter method. Thus, I recommend estimating affine models by the Kalman filter rather than the factor inversion