Eigen Portfolios: From Single Component Models to Ensemble Approaches

TL;DR

This paper explores eigen-portfolios derived from PCA of asset returns, demonstrating that ensemble approaches combining multiple eigen-portfolios outperform single-component strategies in out-of-sample tests.

Contribution

It formalizes eigen-portfolios using PCA and introduces an ensemble method that enhances out-of-sample performance over traditional single-component models.

Findings

Single eigen-portfolios often overfit and perform poorly out-of-sample.

Ensemble strategies combining multiple eigen-portfolios improve robustness.

Ensemble methods outperform benchmark portfolios in Sharpe ratio.

Abstract

The increasing integration of data science techniques into quantitative finance has enabled more systematic and data-driven approaches to portfolio construction. This paper investigates the use of Principal Component Analysis (PCA) in constructing eigen-portfolios - portfolios derived from the principal components of the asset return correlation matrix. We begin by formalizing the mathematical underpinnings of eigen-portfolios and demonstrate how PCA can reveal latent orthogonal factors driving market behavior. Using the 30 constituent stocks of the Dow Jones Industrial Average (DJIA) from 2020 onward, we conduct an empirical analysis to evaluate the in-sample and out-of-sample performance of eigen-portfolios. Our results highlight that selecting a single eigen-portfolio based on in-sample Sharpe ratio often leads to significant overfitting and poor generalization. In response, we propose an ensemble strategy that combines multiple top-performing eigen-portfolios. This ensemble method substantially improves out-of-sample performance and exceeds benchmark returns in terms of Sharpe ratio, offering a practical and interpretable alternative to conventional portfolio construction methods.