EVALUATING ML MODELS IN MODERN PORTFOLIO DIVERSIFICATION – A THEORETICAL EMPIRICAL APPROACH

The integration of machine learning (ML) into portfolio diversification has emerged as a transformative approach to address the limitations of classical financial frameworks. This study evaluates the efficacy of ML models, such as Gradient Boosted Decision Trees (GBDTs) and Long Short-Term Memory networks (LSTMs), against traditional methods like mean-variance optimization and risk parity. By adopting a theoretical-empirical approach, the research highlights ML’s capability to capture non-linear dependencies in financial markets—such as regime-switching dynamics and tail risks—through advanced techniques like copula models and graphical dependency networks. The analysis reveals that ML models excel in high-dimensional data environments, offering superior risk-adjusted returns and diversification benefits, particularly in volatile regimes. However, challenges such as overfitting, interpretability trade-offs, and sensitivity to hyperparameter tuning persist. Classical models, while transparent and theoretically grounded, underperform in scenarios requiring adaptability to time-varying correlations and structural market breaks. The study further proposes enhancements to classical frameworks through ML-driven regularization and adaptive feature engineering, bridging the gap between computational innovation and financial theory. Empirical findings underscore ML’s dominance in long-horizon forecasting and complex non-linear modeling, though hybrid approaches combining ML’s predictive power with classical interpretability are advocated for balanced portfolio strategies. This work contributes to the evolving discourse on modern portfolio management by delineating scenarios where ML augments traditional practices and identifying critical areas for future research.