Empirical Analysis of Realized Variance and Covariance Estimators

Presented: December 4th, 2007

Speaker: Scott Payseur, Senior Financial Engineer, Insightful Corporation

View the on-demand web cast. Download instructions on how to view the web cast.

Download the web cast presentation.

Theoretically, realized estimates of integrated second moments are unbiased.  However, this result depends on an absence of market-microstructure noise at a very high sampling frequency.  Bid-ask-bounce creates an upwards bias in realized variance estimates and non-synchronous trading creates a bias towards zero in realized covariance estimates. Sampling at lower frequencies helps eliminate the bias of realized estimators but comes at the expense of high variability.  In the first half of this webcast I will introduce the S-PLUS and R realized variance toolkit.  This toolkit contains many of the new bias correcting estimators (kernel and sub-sampling based), as well as, tools to help visualize the performance of each.  In the second half, I will demonstrate why realized estimates can be so powerful.  This will be done by comparing ex-ante forecasts of multivariate GARCH conditional second moments with ex-post realizations of realized variance.  Without realized estimates our best guess for the latent ex-post variance process is squared returns and cross returns, which are extremely biased.