A Signal-driven Hierarchical Cross-Attention Portfolio Agent with Deep Reinforcement Learning

Deep Reinforcement Learning (DRL) has shown promise for automated portfolio management, yet agents often trade excessively on market noise rather than on high-conviction signals. To address this, we introduce the Signal-driven Hierarchical Cross-Attention Portfolio Agent (SHCAPA), a novel framework featuring a hierarchical cross-attention mechanism. The core innovation is a structured state representation that separates global market features from asset-specific data, including a heuristic Transaction Turning Point Signal derived from the MACD indicator. This structured state is processed by a novel cross-attention module where the global market context acts as a query to dynamically attend to individual asset features. This architecture enables the agent to make context-aware decisions, prioritizing high-conviction trading opportunities flagged by our signal while filtering out market noise. Validated on DJIA 30 stocks for the 2024 calendar year, SHCAPA substantially outperforms DRL and traditional baselines, achieving a cumulative return of 28.96% and a Sharpe ratio of 1.33. Our work demonstrates that combining heuristic financial signals with a financially-aware attention architecture provides a robust solution for practical DRL-based portfolio management.