The Myth of Diversification Reconsidered

Abstract

That investors should diversify their portfolios is a core principle of modern finance. Yet there are some periods in which diversification is undesirable. When the portfolio’s main growth engine performs well, investors prefer the opposite of diversification. An ideal complement to the growth engine would provide diversification when it performs poorly and unification when it performs well. Numerous studies have presented evidence of asymmetric correlations between assets. Unfortunately, this asymmetry is often of the undesirable variety: It is characterized by downside unification and upside diversification. In other words, diversification often disappears when it is most needed. In this article, the authors highlight a fundamental flaw in the way some prior studies have measured correlation asymmetry. Because they estimate downside correlations from subsamples in which both assets perform poorly, they ignore instances of successful diversification (i.e., periods in which one asset’s gains offset the other’s losses). The authors propose instead that investors measure what matters: the degree to which a given asset diversifies the main growth engine when it underperforms. This approach yields starkly different conclusions, particularly for asset pairs with low full-sample correlation. The authors review correlation mathematics, highlight the flaw in prior studies, motivate the correct approach, and present an empirical analysis of correlation asymmetry across major asset classes.

TOPICS: Portfolio theory, portfolio construction, quantitative methods, statistical methods, performance measurement

Key Findings

• ▪ There is strong empirical evidence that asset class correlations are asymmetric, which poses complications in portfolio construction.

• ▪ Investors prefer diversification when a portfolio’s main growth engine performs poorly and unification when it performs well.

• ▪ To measure correlation asymmetry caused by nonnormality, investors must adjust for changes in correlation that arise mathematically when part of a sample is excluded.

• ▪ Unlike prior research, investors should condition correlations on the performance of a single asset, not two assets.