Reducing Estimation Risk in Mean-Variance Portfolios with Machine Learning

TL;DR

This paper introduces a machine learning framework for portfolio optimization that reduces estimation risk by effectively predicting optimal weights, outperforming traditional and shrinkage methods across simulations and datasets.

Contribution

It develops a novel ML-based approach for estimating portfolio weights that encompasses traditional methods as special cases and offers improved risk reduction.

Findings

ML significantly reduces estimation risk compared to traditional methods.

The framework outperforms equal weight strategies in simulations and real datasets.

New estimation methods derived from ML literature enhance portfolio optimization.

Abstract

In portfolio analysis, the traditional approach of replacing population moments with sample counterparts may lead to suboptimal portfolio choices. I show that optimal portfolio weights can be estimated using a machine learning (ML) framework, where the outcome to be predicted is a constant and the vector of explanatory variables is the asset returns. It follows that ML specifically targets estimation risk when estimating portfolio weights, and that "off-the-shelf" ML algorithms can be used to estimate the optimal portfolio in the presence of parameter uncertainty. The framework nests the traditional approach and recently proposed shrinkage approaches as special cases. By relying on results from the ML literature, I derive new insights for existing approaches and propose new estimation methods. Based on simulation studies and several datasets, I find that ML significantly reduces estimation risk compared to both the traditional approach and the equal weight strategy.