This paper presents a comprehensive comparison between Deep Reinforcement Learning (DRL) and traditional Mean-Variance Optimization (MVO) for portfolio allocation. The authors design a DRL framework using Proximal Policy Optimization (PPO) with a Differential Sharpe Ratio reward function, which optimizes risk-adjusted returns at each timestep. The environment uses market replay of S&P 500 sector indices, with state features including historical log returns, current portfolio weights, and volatility indicators. For MVO, they use a 60-day lookback with Ledoit-Wolf shrinkage for covariance estimation and optimize for maximum Sharpe ratio. Both methods are tested on 10 annual backtests from 2012 to 2021, with DRL agents trained on rolling 5-year windows and validated on a burn year.

The results show that DRL significantly outperforms MVO across multiple metrics: average annual return (12.11% vs 6.53%), Sharpe ratio (1.17 vs 0.68), and maximum drawdown (-32.96% vs -33.03%). DRL also exhibits more consistent monthly returns and lower portfolio turnover, which implies lower transaction costs in practice. The study addresses common shortcomings in prior work by ensuring both methods optimize for the same objective (risk-adjusted returns) and providing detailed implementation details. The findings suggest that DRL can be a viable alternative to classical portfolio optimization, especially in dynamic market conditions.