Forecasting sovereign bond realized volatility using time-varying coefficients model

This paper studies predictability of realized volatility of U.S. Treasury futures using high-frequency data for 2-year, 5-year, 10-year and 30-year tenors from 2006 to 2017. We extend heterogeneous autoregressive model by Corsi (2009) by higher-order realized moments and allow all model coefficients to be time-varying in order to explore dynamics in forecasting power of individual predictors across the term structure. We find realized kurtosis to be valuable predictor across the term structure with robust contribution also in out-of-sample analysis for the shorter tenors. Time-varying coefficient models are found to bring significant out-of-sample forecasting accuracy gain at the short end of the term structure. Further, we detect significant asymmetry in forecasting errors present for all the tenors as the constant-coeffi cient models were found to generate systemic under-predictions of future realized volatility.