We study the relation between realized and implied volatility in the bond market. Realizedvolatility is constructed from high-frequency (5-minute) returns on 30 year Treasury bond futures.Implied volatility is backed out from prices of associated bond options. Recent nonparametric statisticaltechniques are used to separate realized volatility into its continuous sample path and jumpcomponents, thus enhancing forecasting performance. We generalize the heterogeneous autoregressive(HAR) model to include implied volatility as an additional regressor, and to the separateforecasting of the realized components. We also introduce a new vector HAR (VecHAR) modelfor the resulting simultaneous system, controlling for possible endogeneity of implied volatility inthe forecasting equations. We show that implied volatility is a biased and inefficient forecast in thebond market. However, implied volatility does contain incremental information about future volatilityrelative to both components of realized volatility, and even subsumes the information contentof daily and weekly return based measures. Perhaps surprisingly, the jump component of realizedbond return volatility is, to some extent, predictable, and bond options appear to be calibrated toincorporate information about future jumps in Treasury bond prices, and hence interest rates.