Forecasting the yield curve: art or science?

The objective of our work is to analyze the forecast performance of the dynamic Nelson-Siegel yield curve model and, for comparison, the first order autoregressive (AR(1)) model applied to a set of US bond yield data that covers a large timespan from November 1971 to December 2008. As a reference we take the random walk model applied to the yield data. For our analysis, we make use of a simple parameter representing the relative forecast performance to compare forecasting results of different methods. Our findings indicate that none of the yield curve models convincingly beats the random walk model. Furthermore, our results show that deriving conclusions on basis of model testing for a limited time period is inadequate.