We employ a nonlinear, nonparametric method to model the stochastic behavior of changes in several short and long term U.S. interest rates. We apply a nonlinear autoregression to the series using the locally weighted regression (LWR) estimation method, a nearest-neighbor method, and evaluate the forecasting performance with a measure of root mean square error (RMSE). We compare the forecast performance of the nonparametric fit to the performance of two benchmark linear models: an autoregressive model and a random-walk-with-drift model. The nonparametric model exhibits greater out-of-sample forecast accuracy than that of the linear predictors for most U.S. interest rate series. The improvements in forecast accuracy are statistically significant and robust. This evidence establishes the presence of significant nonlinear mean predictability in U.S. interest rates, as well as the usefulness of the LWR method as a modeling strategy for these benchmark series.