Realized variance modeling: decoupling forecasting from estimation

This paper evaluates the in-sample fit and out-of-sample forecasts of various combinations of realized variance models and functions delivering estimates (estimation criteria). Our empirical findings highlight that: independently of the econometrician’s forecasting loss function, certain estimation criteria perform significantly better than others; the simple ARMA modeling of the log realized variance generates superior forecasts than the HAR family, for any of the forecasting loss functions considered; the (2,1) parameterizations with negative lag-2 coefficient emerge as the benchmark specifications generating the best forecasts and approximating long-range dependence as well as does the HAR family.