Evolution and Taxonomy of Portfolio Optimization: Traditional Theories vs. Neural Network Innovations

Abstract This paper examines the evolution of portfolio theory, beginning with the foundational mean-variance framework introduced by Markowitz, and explores advancements in computational methods, particularly artificial neural networks (ANNs), for addressing modern portfolio optimization challenges. While traditional models like the efficient frontier remain pivotal, they often struggle with real-world constraints such as dynamic data interactions, non-linear relationships, and high-dimensional datasets. Through a comparative analysis, we demonstrate how machine learning techniques—especially Long Short-Term Memory (LSTM) networks and hybrid deep learning architectures—overcome these limitations by enabling robust predictions of asset returns and optimal weight allocation. Empirical evidence from reviewed studies highlights that neural network models significantly outperform classical approaches in terms of Sharpe ratios, risk-adjusted returns, and adaptability to complex financial environments. The study underscores the necessity of integrating traditional financial theories with advanced machine learning to enhance portfolio diversification, risk management, and return optimization. This synthesis not only clarifies the strengths and weaknesses of existing models but also provides a roadmap for future research in adaptive, data-driven portfolio strategies. Keywords: Portfolio optimization, Markowitz model, efficient frontier, neural networks, LSTM, risk-return tradeoff, machine learning in finance. JEL classification: G11, G17.