A1318
Title: Regularizing stock return covariance matrices via multiple testing
Authors: Richard Luger - Laval University (Canada) [presenting]
Abstract: A new method is developed for the regularization of stock return covariance matrices in well-diversified portfolio settings. The framework is quite general and allows for the presence of heavy tails and multivariate GARCH-type effects of unknown form. The approach proceeds by simultaneously testing all pairwise correlations and then sets to zero the elements that are not statistically significant. Distribution-free Monte Carlo test procedures are proposed for control of the familywise error rate. A subsequent shrinkage step ensures that the covariance matrix estimate is positive definite and well-conditioned, while preserving the achieved zeros. When compared to alternative estimators, the new regularization method is found to perform well both in simulation experiments and in an actual portfolio optimization application with low-volatility stocks.
