Reinforcement Learning for Single-Stock Portfolio Optimization in Equity Markets

Abstract In today’s volatile financial markets, flexible asset allocation and portfolio management are crucial for maximizing returns and lowering risk. Real-time portfolio allocation optimization is achieved by the Reinforcement Learning-based Adaptive Portfolio Management (RL-APM) system using the Deep Deterministic Policy Gradient (DDPG) approach. A state vector representing volatility, recent returns, and current portfolio weights across a range of equity assets is produced by RL-APM, which also analyzes daily market data, including Open, High, Low, Close, and Volume (OHLCV) values, and calculates technical indicators such as SMA, RSI, MACD, and Bollinger Bands. This state is used by the DDPG agent to create continuous portfolio weights that dynamically adapt to market fluctuations. In contrast to traditional DDPG-based techniques, the method introduces a novel multi-objective reward function that simultaneously maximizes volatility, Sharpe ratio, cumulative return, and transaction costs, improving stability and realism. RL-APM outperformed benchmark models like Mean-Variance Optimization (MVO) using daily S&P 500 index data from 2015 to 2024. Average returns increased by 12.4% and the Sharpe ratio by 15.7%. These results demonstrate that the system may scale efficiently for real-world financial scenarios, such as hedge funds, automated trading, and robo-advisory platforms, while also creating flexible allocation plans and maintaining robust risk management.