Attention-enhanced reinforcement learning for dynamic portfolio optimization

We propose a deep reinforcement learning framework for dynamic portfolio optimization that combines a Dirichlet policy with cross-sectional attention mechanisms. The Dirichlet distribution enforces feasibility by construction, accommodates tradability masks, and provides a coherent geometry for exploration. Our architecture integrates per-asset temporal encoders with a global attention layer, allowing the policy to adaptively weight sectoral co-movements, factor spillovers, and other cross-asset dependencies. We evaluate the framework on a comprehensive S&P 500 panel from 2000 to 2025 using purged walk-forward backtesting to prevent look-ahead bias. Empirical results show that attention-enhanced Dirichlet policies deliver higher terminal wealth, Sharpe and Sortino ratios than equal-weight and reinforcement learning baselines, while maintaining realistic turnover and drawdown profiles. Our findings highlight that principled action parameterization and attention-based representation learning materially improve both the stability and interpretability of reinforcement learning methods for portfolio allocation.