Portfolio Selection via Deep Reinforcement Learning: Comparative Analysis with Classical Strategies

Abstract

Portfolio management involves the strategic allocation of assets to achieve investment objectives while minimizing risk. This study explores portfolio management by comparing traditional Mean-Variance Optimization (MVO), the Black-Litterman Model, and Deep Reinforcement Learning (DRL) for optimizing portfolio allocations. MVO utilizes the Ledoit-Wolf shrinkage method to estimate covariance matrices and applies Efficient Frontier techniques for optimization. The Black-Litterman Model extends MVO by incorporating investor views into the equilibrium market returns derived from the Capital Asset Pricing Model (CAPM), offering a more balanced approach. In contrast, Proximal Policy Optimization (PPO) is used as the DRL method to dynamically adjust portfolio weights. Empirical analysis, based on backtesting with historical market data, shows that the DRL approach significantly outperforms both the MVO and the Black-Litterman Model across various performance metrics, including cumulative return, annual return, volatility, and Sharpe ratio. These results highlight the potential of DRL, particularly PPO strategies, as robust tools for dynamic and adaptive portfolio management, capable of achieving superior returns and effectively managing risks effectively in modern financial markets.