A Bayesian high-frequency estimator of the multivariate covariance of noisy and asynchronous returns

A multivariate positive definite estimator of the integrated covariance matrix of noisy and asynchronously observed asset returns is proposed. We adopt a Bayesian Dynamic Linear Model where microstructure noise is interpreted as measurement error, and asynchronous trading as missing observations in an otherwise synchronous series. Missing observations are treated as any other parameter, as typical in a Bayesian framework. An augmented Gibbs algorithm is used since all full conditionals are available and its convergence and robustness are discussed. A realistic simulation study compares our estimator with existing alternatives, under different liquidity and microstructure noise conditions. The results suggest that our estimator is superior in terms of RMSE particularly under severe conditions, such as portfolios of assets with heterogeneous liquidity and high level of microstructure noise. The application to the empirical dataset of ten tick-by-tick stock price series confirms the simulation results.