Document Type : Original/Review Paper

Authors

• javad mohammadzadeh ¹
• Hossein Mohammadifar ²
• Mahdi Mazinani ³
• abbas Koochari ²

¹ Department of Computer Engineering, Ka.C., Islamic Azad University, Karaj, Iran.

² Department of Computer Engineering, SR.C., Islamic Azad University, Tehran, Iran

³ Department of Electrical and Electronic Engineering, ShQ.C., Islamic Azad University, Shahr-e Qods, Iran

Abstract

A reinforcement learning-based framework is proposed for optimizing investment portfolio allocation in dynamic financial markets. Traditional portfolio optimization approaches often struggle to capture the nonlinear dynamics, uncertainty, and regime shifts inherent in financial environments, limiting their ability to adapt to changing market conditions.To address these challenges, an ensemble reinforcement learning framework is developed by integrating A2C, SAC, and an adaptive DDPG (ADDPG) with an adaptive exploration-noise mechanism. The framework combines handcrafted financial indicators with 16 latent features extracted by a fully connected neural network (FCN) and LSTM trend forecasts. Together with portfolio-state variables (asset holdings, cash balance, and unrealized profit and loss), these inputs construct the Markov Decision Process (MDP) state space for sequential trading. Finally, aggregating agent decisions enhances robustness across market regimes under diversification and minimum allocation constraints. Reinforcement learning agents are trained using step-wise rewards derived from changes in portfolio value after transaction costs, while portfolio performance is evaluated using Cumulative Return Ratio (CRR), Sharpe Ratio (SR), Sortino Ratio (SoR), and Downside Risk (DR). The experimental results demonstrate complementary performance among A2C, SAC, and ADDPG across different asset classes, while the portfolio-level ensemble methods provide competitive and robust allocation results under different diversification constraints. The findings highlight the potential of integrating heterogeneous feature representations, trend forecasting, adaptive reinforcement learning, and ensemble portfolio optimization for robust and adaptive financial decision-making.

Keywords

• Deep Reinforcement Learning
• Ensemble Learning
• Long-Term Investment
• Neural Feature Extraction
• Portfolio Optimization

Main Subjects

• H.3. Artificial Intelligence