Optimal versus Naive Diversification: False Discoveries, Transaction Costs and Machine Learning

Abstract

This paper shows that sophisticated diversification strategies never underperform the 1/N rule when adjusting for multiple testing; however, their edge is severely undermined by transaction costs. As a way forward, this paper provides a machine learning approach for ex-ante strategy selection. By linking the characteristics of investment scenarios to the out-of-sample performance of strategies, the algorithm never underperforms the 1/N rule, even in the presence of relatively high transaction costs.

Keywords: Portfolio selection, Multiple-testing adjustment, Machine learning

JEL Classification: G11, G14, C58, B26, N20