Abstract: This paper proposes new dynamic component models of realized covariance (RCOV) matrices
based on recent work in time-varying Wishart distributions.
The specifications are linked to returns for a joint multivariate model of
returns and covariance dynamics that is both easy to estimate and forecast.
Realized covariance matrices are constructed for 5 stocks
using high-frequency intraday prices based on positive semi-definite
realized kernel estimates. The models are compared based on a term-structure of density
forecasts of returns for multiple forecast horizons.
Relative to multivariate GARCH models that use only daily returns, the joint
RCOV and return models provide significant improvements in density
forecasts from forecast horizons of 1 day to 3 months ahead.
Global minimum variance portfolio selection is improved for forecast horizons
up to 3 weeks out.
Keywords: eigenvalues, dynamic conditional correlation, predictive likelihoods, MCMC
JEL Classification: C11; C32; C53; G17