Forecasting the Term Structure of Interest Rates with Potentially Misspecified Models

We compare the out-of-sample predictive accuracy of a mixture of bond yield models with that of the individual models. The individual models considered here are the dynamic Nelson--Siegel model, arbitrage-free Nelson--Siegel model, and random-walk model. Out-of-sample forecasts for U.S. bond yields show that none of the individual models dominates the others across all maturities and forecast horizons, although the random-walk model performs well in most cases. We then assess the predictive accuracy for two subsamples: before and during a period of zero interest rates. In the first subperiod, overall the mixture of the three models outperforms the individual models, whereas the random-walk model seems to forecast better than all combinations for the period of zero interest rates. We show that these mixed results on the forecasting ability of the mixture models across the subsamples can be attributed to the zero interest-rate policy