Economics Seminar -John Cotter

We examine whether a factor-based framework to construct the covariance matrix can enhance the out- of-sample performance of minimum-variance portfolios. We study two refinements to the method of factor-based covariance estimation. First, we consider dimensionality reduction approaches that induce sparsity or introduce non-linearities to the latent factors. Second, we compare the performance of static and dynamic specifications of the covariance matrix. Our key findings show that sparse and non-linear latent factors lead to portfolios that significantly outperform simpler benchmarks in terms of risk minimization. Portfolios based on latent factors tend to produce weights which are smaller, less volatile throughout time and more diversified than models based on observed factors. Dynamic covariance specifications often lead to additional gains over the static covariance, but at the cost of higher turnover.