Diversifying Estimation Errors with Unsupervised Machine Learning

We investigate how unsupervised machine learning can reduce estimation errors in mean-variance portfolio optimization. Our approach clusters assets into equally weighted sub-portfolios that then feed into classical minimum-variance optimization. The resulting clustered-variance strategies significantly outperform equally weighted portfolios, while avoiding the extreme weights of plug-in minimum variance. The optimal number of clusters approximates N/4 when varying from one cluster (equally weighted) to N clusters (plug-in minimum variance), suggesting a structural bias-variance sweet spot in portfolio construction.

Working Paper (University of Liechtenstein)