A0758
Title: Issues in large covariance matrix estimation for portfolio risk prediction
Authors: Stjepan Begusic - Unversity of Zagreb (Croatia) [presenting]
Abstract: Most studies considering the problem of estimating large covariance matrices of asset returns from a short time window (high dimension low sample size (HDLSS) regime) focus on portfolio optimization applications. It is known that mean-variance optimization acts as error maximization (especially in the HDLSS regime), which has been mitigated by certain estimators, such as those based on the spiked covariance model. By imposing factor models or shrinking the covariance estimates towards them, the robustness of optimized portfolios can be improved, and their out-of-sample risk reduced. However, correcting the optimization error might lead to introducing errors in other applications, most notably those in portfolio risk prediction. The aim is to focus on these applications, specifically the estimation of portfolio variance in a spiked covariance model and the issues which arise with different covariance matrices in the HDLSS regime. The properties of the risk prediction errors and the effects of different eigenstructure shrinkage methods on these errors are considered. An experimental study is presented, for a range of dimensionality scenarios and various portfolios, together with some insights for practitioners and directions for future work.