Machine learning vs traditional portfolio optimisation under different market conditions in exchange traded funds portfolios.

Abstract

The rapid advancement of machine learning (ML), a branch of artificial intelligence that enables systems to learn from data, has significantly reshaped financial decisionmaking. This study examines the comparative performance of traditional and MLbased portfolio optimization methods in Exchange-Traded Funds (ETFs), which provide diversified exposure to equities, bonds, commodities, and alternative assets. ETFs’ liquidity, transparency, and cost-effectiveness make them an ideal context for testing optimization strategies. The research employs a three-stage methodology. First, a Systematic Literature Review (SLR) assesses the global application of ML techniques in portfolio optimization, highlighting models such as Random Forest (RF), Long Short-Term Memory (LSTM), and Support Vector Machines (SVM). Second, an empirical analysis uses daily return data from 33 JSE-listed ETFs spanning 2013–2023 to compare three traditional models Markowitz Model (MM), Conditional Value-at-Risk (CVaR), and Mean Absolute Deviation (MAD) with three ML models: Enhanced Radial Basis Function Artificial Neural Network (K4-RANN), Genetic Algorithms (GA), and Extreme Gradient Boosting (XGBoost). Third, event study methodology evaluates portfolio resilience during South African market shocks, namely the COVID-19 pandemic (2020–2022), the July 2021 social unrest, and the April 2022 KwaZulu-Natal floods. Performance is measured through mean return, standard deviation, and Sharpe ratio. Findings show that ML-based models generally outperform traditional approaches by generating higher risk-adjusted returns and improving diversification. Under normal market conditions, GA and K4-RANN produced portfolios with lower risk and stronger returns. During crises, XGBoost proved most effective, striking a balance between returns and volatility and demonstrating adaptability in turbulent markets. These results confirm the potential of ML models to enhance portfolio stability and resilience in both normal and distressed environments. The study concludes that portfolio managers and institutional investors should integrate ML-based optimization, particularly ensemble models like XGBoost, into dynamic allocation strategies. For policymakers and financial institutions in emerging markets, investment in ML infrastructure and data capabilities is recommended to drive innovation in asset management. Future research should investigate hybrid frameworks that merge traditional financial theory with ML to advance more comprehensive portfolio optimization approaches.