Diversified Spectral Portfolios: An Unsupervised Learning Approach to Diversification

Abstract

The question of how to diversify an investment portfolio is one with many possible answers. Over the past couple of years, the industry and academic literature have been shifting focus from an asset-driven answer to a factor-driven one, sparking special interest in the use of implicit factors identified through unsupervised learning. However, issues around the stability and implementation of these, in the context of diversification, have left a gap between what is an academic exercise and what is an implementable methodology. This article aims to fill this gap by presenting a diversification-focused portfolio construction methodology that takes advantage of singular value decomposition to identify implicit factors and uses hierarchical agglomerative clustering to address some of the challenges surrounding its implementation. In out-of-sample Monte Carlo simulations, this methodology can provide better risk-adjusted performance than other commonly used portfolio diversification approaches.