Graph attention-based heterogeneous multi-agent deep reinforcement learning for adaptive portfolio optimization

Abstract

Abstract Traditional portfolio optimization methods face significant limitations in capturing complex asset relationships and adapting to dynamic market conditions. This paper proposes a novel graph attention-based heterogeneous multi-agent deep reinforcement learning framework that addresses these challenges through innovative integration of graph neural networks and specialized agent architectures. The framework employs graph attention networks to model time-varying asset correlations and dependencies, while utilizing three heterogeneous agents specialized in risk assessment, return prediction, and market environment perception. An adaptive optimization strategy dynamically adjusts parameters based on real-time market conditions and regime changes. Comprehensive experiments on S&P 500, NASDAQ 100, and Russell 2000 datasets demonstrate superior performance, achieving 16.8% annualized returns, 1.34 Sharpe ratio, and 8.2% maximum drawdown, significantly outperforming traditional mean–variance optimization, equal-weight portfolios, and existing deep learning approaches. Ablation studies confirm the critical contributions of each framework component, while sensitivity analysis validates robustness across varying market conditions. The proposed framework represents a significant advancement in computational finance, offering enhanced adaptability and risk management capabilities for modern portfolio optimization challenges.