Machine Learning and Factor-Based Portfolio Optimization

Abstract

We examine machine learning and factor-based portfolio optimization. We find that factors based on autoencoder neural networks exhibit a weaker relationship with commonly used characteristic-sorted portfolios than popular dimensionality reduction techniques. Machine learning methods also lead to covariance and portfolio weight structures that diverge from simpler estimators. Minimum-variance portfolios using latent factors derived from autoencoders and sparse methods outperform simpler benchmarks in terms of risk minimization. These effects are amplified for investors with an increased sensitivity to risk-adjusted returns, during high volatility periods or when accounting for tail risk.

Keywords: Autoencoder, covariance matrix, dimensionality reduction, factor models, machine learning, minimum-variance, principal component analysis, Partial least squares, portfolio optimization, sparse principal component analysis, sparse partial least squares

JEL Classification: C38, C4, C45, C5, C58, G1, G11