Quant Investing in Cluster Portfolios

This paper discusses portfolio construction for investing in N given assets, e.g. constituents of the Dow Jones Industrial Average (DJIA) or large cap stocks, which is based on partitioning the investment universe into clusters. The clusters are determined from the trailing correlation matrix via an information theoretic algorithm that uses thresholding of high-correlation pairs. We calculate the Principal Eigenvector of each cluster from its correlation matrix and the corresponding eigenportfolio. The cluster portfolios are combined into a single N-asset portfolio based on a weighting scheme for the clusters. Various tests conducted on components of DIA and a thirty-stock basket of large-cap stocks indicate that the new portfolios are superior to the DIA and other Mean-Variance portfolios in terms of risk-adjusted returns from 2009 to 2019. We also tested the cluster portfolios for the larger basket of 373 S&P500 components from 2001 to 2019. The test results give convincing evidence that cluster-based portfolio can outperform passive investing.KEY MESSAGES1) A correlation-based set partitioning algorithm that divides the investment universe dynamically into clusters of assets proposed.2) The principal eigenvector of each cluster from its correlation matrix is calculated and the corresponding eigenportfolio. The cluster portfolios of varying sizes combined into a single N-asset portfolio based on a weighting scheme for the clusters.3) The proposed portfolio outperforms hierarchical risk parity (HRP) portfolio, eigenportfolio (EP), and a few other portfolio constructions and relevant ETFs based on several tests performed with market data.4) The performance comparisons give convincing evidence that cluster-based long-only investment portfolio can outperform passive investing