We propose a novel approach to decompose realized jump measures by type of activity (finite/infinite) and sign, and also provide noise-robust versions of the ABD jump test (Andersen et al., 2007b) and realized semivariance measures. We find that infinite (finite) jumps improve the forecasts at shorter (longer) horizons; but the contribution of signed jumps is limited. As expected, noise-robust measures deliver substantial forecast improvements at higher sampling frequencies, although standard volatility measures at the 300-s frequency generate the smallest MSPEs. Since no single model dominates across sampling frequency and forecasting horizon, we show that model averaged volatility forecasts – using time-varying weights and models from the model confidence set – generally outperform forecasts from both the benchmark and single best extended HAR model. Finally, forecasts using volatility and jump measures based on transaction sampling are inferior to the forecasts from clock-based sampling.