Abstract

This paper compares three approaches to estimating equity covariance matrices: a factor model, a market model and an unstructured asset-by-asset model. These approaches make different trade-offs between estimation variance and model specification error. We explore this trade-off with a simulation experiment and with an empirical analysis of UK equity portfolios. The factor model is found to perform best for large investment universes and typical sample lengths. The market model underperforms due to excessive specification error while an asset-by-asset model with a short half-life of 22 days underperforms due to high estimation variance. The importance of properly accounting for serial correlation is highlighted.