Portfolio Selection Using Artificial Intelligence:From Genetic Algorithms to Deep Reinforcement Learning, Graph Neural Networks, and Large Language ModelsA Revised and Extended Framework

More than a decade has passed since the genetic algorithm was first applied to the portfolio selection problem in Sefiane and Benbouziane (2012). That paper showed that arithmetic crossover could reliably identify efficient weight allocations in a constrained multi-objective setting. Much has changed since then. This paper revisits that earlier work not to replace it, but to situate it within the far richer methodological landscape that artificial intelligence now offers to quantitative finance. Three modern paradigms are integrated — deep reinforcement learning (DRL), graph neural networks (GNN), and large language models (LLMs) — alongside the original genetic algorithm (GA) framework, and a unified Hybrid AI (HAI) architecture is proposed. Two empirical illustrations are developed: the original five-asset dataset retained for direct comparability, and a ten-asset BRICS-inspired stylized portfolio spanning 2015 to 2024, calibrated to reflect elevated volatility, contagion dynamics, and regime breaks characteristic of emerging market investing. Seven figures accompany the analysis. Results confirm that GA-arithmetic crossover remains a strong interpretable baseline, while HAI dominates all benchmarks, improving the Sharpe ratio by 31 percent over GA on the original dataset and by 23 percent on the BRICS universe. The paper concludes with a targeted research agenda for the Algerian and MENA context, covering Arabic-language LLMs for regional financial sentiment, oil-price regime DRL, and Maghreb cross-border GNN construction.