A1363
Title: Coherent forecasting of realized volatility
Authors: Marius Puke - University of Hohenheim (Germany) [presenting]
Karsten Schweikert - University of Hohenheim (Germany)
Abstract: The QLIKE loss function is the favored choice in the literature for out-of-sample evaluations of volatility forecasts. However, the state-of-the-art models for forecasting realized volatility (RV), such as the heterogeneous autoregressive (HAR) model, traditionally minimize the squared error loss for in-sample parameter estimation. To the best of knowledge, no prior research has investigated the potential benefits of using the QLIKE loss function also for in-sample parameter estimation within an extensive volatility forecasting exercise. In doing so, the HAR model is embedded within the class of M-estimators, which theoretically justifies the use of the rich class of Bregman loss functions, including, among many others, the squared error and QLIKE loss functions. While asymptotically, the squared error and QLIKE loss functions should yield the same solution, our empirical findings reveal forecast performance gains from a HAR model directly estimated using the QLIKE loss in finite samples. Results are based on the current cross-section of stocks in the Dow Jones index, where the newly proposed HAR model extension demonstrates significant improvements in forecast performance, particularly for short-term horizons.
