Realized Volatility Forecasting in the Presence of Time-Varying Noise

Abstract

Observed high-frequency financial prices can be considered as having two components, a true price and a market microstructure noise perturbation. It is an empirical regularity, coherent with classical market microstructure theories of price determination, that the second moment of market microstructure noise is time-varying. We study the optimal, from a finite-sample forecast mean squared error (MSE) standpoint, frequency selection for realized variance in linear variance forecasting models with time-varying market microstructure noise. We show that the resulting sampling frequencies are generally considerably lower than those that would be optimally chosen when time-variation in the second moment of the noise is unaccounted for. These optimal, lower frequencies have the potential to translate into considerable out-of-sample MSE gains. When forecasting using high-frequency variance estimates, we recommend treating the relevant frequency as a parameter and evaluating it jointly with the parameters of the forecasting model. The proposed joint solution is robust to the features of the true price formation mechanism and generally applicable to a variety of forecasting models and high-frequency variance estimators, including those for which the typical choice variable is a smoothing parameter, rather than a frequency.