Risk mitigation through noise reduction in hierarchical portfolio selection

Abstract

Risk parity portfolio methods rely solely on covariance estimates to minimize risk, ignoring expected returns due to their high estimation error. This approach can be unstable when dealing with a reduced number of observations. We address this limitation by improving the signal-to-noise ratio in covariance and correlation matrix estimation within hierarchical portfolio selection models. Our approach combines shrinkage covariance estimation, a backbone network extraction, and density-based clustering method. We test two workflows: one for covariance and one for correlation matrices across four real-world market datasets (S&P, Dow Jones, Euro Stoxx 50, Ibex 35) and a synthetic dataset. Results show improved out-of-sample performance in terms of value-at-risk and conditional value-at-risk, offering a more robust alternative to standard hierarchical risk parity.