Portfolio Construction Using First Principles Preference Theory and Machine Learning

Abstract

The author develops a novel approach to portfolio construction that builds on the first principles preference theory, incorporating characteristics of portfolios beyond the mean and variance of returns typical of the state-of-the art mean–variance optimization. Specifically, the author adds skewness and kurtosis of return distributions to the framework. This portfolio framework is proposed with the intent to make it easily accessible conceptually and easily implementable in practice so that it will be of pragmatic value to investment professionals. This imperative of practical feasibility leads to the utilization of standard machine learning techniques that proved effective in overcoming the methodological intricacies associated with mean–variance optimization. Furthermore, the author allows for a multiplicity of optimal portfolio outcomes, leaving to the investor some—albeit constrained—discretion in portfolio choice within a class of similarly optimal portfolios. In a Monte Carlo setting that is rich enough to incorporate co-skewness and tail dependencies aside from simple correlation structures between returns, the suggested framework generates more realistic portfolio solutions than mean–variance optimization.

TOPICS: Big data/machine learning, portfolio management, portfolio theory, portfolio construction, simulations

Key Findings

• • This article presents a portfolio construction framework that builds on first principles preference theory, incorporating higher moment return characteristics beyond mean and co-variance typical of mean–variance optimization. Specifically, we also characterize portfolio performance based on return skewness and kurtosis.

• • Our portfolio framework is easily accessible conceptually and easily implementable in practice, providing pragmatic value to practitioners. The utilization of standard machine learning techniques that proved effective in overcoming the intricacies of mean–variance optimization makes our framework feasible and stable.

• • We allow for a multiplicity of optimal portfolios. The investor has some discretion over a class of similarly optimal portfolios. In a Monte Carlo study that accounts for correlations, co-skewness, and tail-dependencies, our framework generates more realistic portfolios than mean–variance optimization.