Robust Portfolio Selection by Bayesian Estimation Methods

This research proposes several performance-based measures to evaluate the estimation of a high-dimensional covariance matrix, where the number of variables can be close to the number of observations. In case of high-dimensional data, the sample covariance contains a lot of estimation error, and is not even invertible in some cases. This research makes a comparative analysis of covariance estimation based on dierent priors for the Bayesian estimation methodology. The performances of the resulting portfolios are judged in the realistic capital asset environment of the New York Stock Exchange. The sample covariance points out to be an unt estimator in case of high-dimensional data. When a proper prior is chosen for the Bayesian estimation method, this method does not suer from this dimensionality issue and clearly outperforms the natural estimator.