A comparison of high-frequency realized variance measures: Duration- vs. return-based approaches

We study the accuracy of a variety of parametric price duration-based realized variance estimators constructed via various financial duration models and compare their forecasting performance with the performance of various non-parametric return-based realized variance estimators. Our financial duration models consist of an ACD(1,1), its logarithmic version, Log-ACD(1,1), and its long-memory version, FIACD(1,1), as well as the Markov-switching multifractal duration (MSMD) model and the factorial hidden Markov duration (FHMD) process. In an empirical study using high-frequency data on ten stocks traded on the New York Stock Exchange (NYSE), our in- and out-of-sample results show that the parametric price duration-based realized variance (RV) estimators, especially the ACD-based RV estimator, perform better than the non-parametric return-based RV estimators. Furthermore, we also find that the price duration-based and return-based RV models produce more accurate and valid Value-at-Risk forecasts than the GARCH(1,1) model.