Comparative Analysis of Traditional and Machine Learning Portfolio Optimization Models Using Public Financial Data

Comparative Analysis of Traditional and Machine Learning Portfolio Optimization Models Using Public Financial Data

Warwick Business School

Yingqi Wang

2025

For portfolio managers seeking implementable portfolio optimization approaches, it is essential to evaluate the models and select the one that delivers the highest after-cost and risk-adjusted performance. This study compares traditional portfolio optimization models (Mean-Variance Optimization and the Black-Litterman model) with machine learning approaches (Random Forest and XGBoost) using the same dataset. The 30 S&P 500 stocks are randomly selected, and their price data from 2015 to 2024, obtained at a monthly frequency from Yahoo Finance, are split into a training window (2015-2021), a validating window (2022), and a testing window (2023-2024). All investment strategies are long-only, fully invested, and rebalanced monthly. An equally weighted portfolio consisting of all 30 selected stocks, applying a buy-and-hold strategy, serves as the benchmark model. The model comparison involves evaluating risk-adjusted portfolio performance and predictive accuracy. The empirical test results from the testing period show that machine learning models, particularly Random Forest, can add significant value to portfolios and achieve good predictive accuracy, but they are cost-sensitive and unstable. In contrast, traditional models perform relatively poorly but are stable under different cost settings.

Portfolio Optimization, Investment Management, Machine Learning, Random Forest, XGBoost, Black-Litterman, Mean-Variance Optimization