Portfolio optimization solutions for developed and emerging financial markets during a period of increased geopolitical tension and economic fragmentation (2023-2024) are examined in this study. Specifically, the study compares the effectiveness of a model-free Deep Reinforcement Learning (DRL) strategy using the Twin-Delayed Deep Deterministic Policy Gradient (TD3) algorithm with that of Modern Portfolio Theory (MPT) - through Maximum Sharpe Ratio (MS) and Minimum Variance (MV) strategies. The top 30 large-cap equities in four sample markets - the United States and Sweden (developed) – China, and Vietnam (emerging) - were used to construct the portfolios. Models were trained on data prior to 2023 and evaluated on out-of-sample performance for 2023–2024 using historical data from 2013–2024. Evaluation indicators included the Sharpe and Sortino ratios, volatility, maximum drawdown, and annualized return.

While DRL provides theoretical adaptability and creativity, further refinement is necessary to enhance its practical implementation in financial markets. This study adds to existing literature by critically evaluating the readiness of AI models for active portfolio management and emphasizing the importance of context-driven strategy selection.