Real-Time Forecasting of U.S. Bond Yields and Their Excess Returns

Some bond investors form expectations for future interest rates by either assuming that yields will not change over time (i.e. random walk), or that future spot rates will evolve to match current forward rates (i.e., the expectations hypothesis). Recent academic research argues that we can do significantly better by modeling term structure dynamics as some time-varying function of current spot and forward rates. Using a new Federal Reserve Board dataset, we examine the out-of-sample forecasting ability of these and other dynamic models for U.S. Treasury bonds and their excess returns. Broadly speaking, we find that none of these models has consistently and statistically outperformed a random walk across bond maturities since 1985. We reconcile our findings with those that claim stronger predictability by tracing the deterioration in out-of-sample forecasts to changes in the U.S. inflation process. From this, we specify simple yield dynamics that out-forecast a random walk in real time. We discuss avenues for future research