Saturday, 19 March 2011: 15:30
The aim of this paper is to consider high dimensional multivariate realized volatility models for large dimensional datasets and also address the solution for noise problem coming out of volatility estimation in the presence of market microstructure effects. Standard models, where prices are contaminated with stochastically independent noise, are unable to explain the behavior of realized variance as the sampling frequency increases. We aim to extend the current analytic methods to the construction and assessment of volatility forecasts for continuous-time volatility models to the empirically important case of market microstructure noise via factors discussed by Bai, Ng (2002, 2004 and 2006) and principal component methodology of Stock and Watson (2002). The main contribution of this paper is to propose a novel way of conducting realized volatility, where integrated volatility takes a linear factor structure, facilitating the estimation of volatility factors while getting rid of the noise. These factors capture the market microstructure problem when applied to a large dimension of individual return series in a stock market. Two major models are suggested for the prediction of the realized volatility: Simple Factor Forecast (SFF) and Heterogeneous Autoregressive Factor Based Forecast (HAR-FF). Finally, forecasting based on the proposed models is studied. We show that the SFF model outperforms the other currently available approaches including HAR-RV, GARCH and AR models at various prediction horizons, not only in terms of minimizing the RMSE of the forecast, or high R² of the Mincer-Zarnowitz regressions, but also in terms of improving the volatility forecasts while dealing with the noise problem with the help of common factors.
Keywords: Realized Volatility; High-frequency Data; Factor models; Volatility Forecasting; Model Selection
JEL classification: C32; C50; C52; C53; C58