Hierarchical Portfolio Optimization

Abstract

The field of Portfolio Optimization has historically had a very hard time as the Mathematical Models at its availability are based on certain assumptions one can not afford to make in the financial markets, making naive approaches all-too enticing. In this project we have introduced the assumption that the different stocks in the financial markets have a hierarchical structure and have allowed ourselves to be inspired by it to build portfolios through a Machine Learning approach. We have employed the Hierarchical Risk Parity algorithm and tested minor variations relating to the dissimilarity measure it makes use of. The tests were conducted with historical daily closing price data from 2014 to 2020 for 440 stocks in the S&P 500 index. Results suggest most of the tested Hierarchical Risk Parity variants are robust and can compete with the Equal Weights Portfolio. We mainly encourage the use of two dissimilarity measures, the standard one, a correlation based metric and Dynamic Time Warping. The former is suggested to the pessimistic investor while the latter to the hopeful yet conservative investor. To optimistic investors with a high risk tolerance the recommendation would be to use the traditional Equal Weights portfolio among the asset allocation methods considered in this project.