Explainable Deep Reinforcement Learning with Multi-Source Sentiment Fusion for High-Frequency Quantitative Asset Pricing

Automated algorithmic trading frameworks increasingly rely on deep reinforcement learning (DRL) agents to optimize real-time capital allocation; however, widespread adoption is frequently hindered by black-box decision models and inadequate handling of non-stationary market regimes. This paper proposes X-FinAgent, a modular deep reinforcement learning framework that integrates temporal cross-asset price dynamics with real-time sentiment streams mined from decentralized financial news and macroeconomic feeds. The model implements an adaptive attention mechanism to dynamic weights between order book microstructures and external textual signals, while employing Layer-wise Relevance Propagation (LRP) to provide interpretable attribution maps for every trade execution. Tested on high-frequency tick data across global equity and index derivatives spanning a five-year volatile window, X-FinAgent achieved a Sharpe Ratio of 3.42 and an annualized alpha of 18.7%, significantly outperforming standard benchmark strategies and standalone LSTM-based baseline models. The empirical results confirm that coupling explainable neural representations with sentiment-augmented policy networks delivers superior risk-adjusted returns while satisfying regulatory compliance and model risk management standards.