Out-of-Sample Predictability of Bond Returns

We test the out-of-sample predictive power for one-year bond excess returns for a variety of models that have been proposed in the literature. We find that these models perform well in sample, but have worse out-of-sample performance than the historical sample mean. We write the one-year excess return on a n-maturity bond at time t + 1 as the difference between n times the n-maturity bond yield at time t, and the sum of n 1 times the (n 1)-maturity bond yield at time t + 1 and the one-year bond yield at time t. Instead of forecasting returns directly, we forecast bond yields and replace them in the bond excess return definition. We use two bond yield forecasting methods: a random walk and a dynamic Nelson-Siegel approach proposed by Diebold and Li (2006). An investor who used a simple random walk on yields would have predicted bond excess returns with outof-sample R-squares of up to 15%, while a dynamic Nelson-Siegel approach would have produced out-of-sample R-squares of up to 30%.