This paper investigates the sensitivity of out-of-sample forecasting performance over a span of different parameters of l in the dynamic Nelson–Siegel three-factor AR(1) model. First, we find that the ad hoc selection of l is not optimal. Second, we find a substantial difference in factor dynamics between investment-grade and speculative-grade corporate bonds from 1994:12 to 2006: 4. Third, we suggest that the three-factor model is sufficient to explain the main variations of corporate yield changes. Finally, the parsimonious Nelson–Siegel three-factor AR(1) model remains competitive in the out-of-sample forecasting of corporate yields. Copyright © 2008 John Wiley & Sons, Ltd.