A Market-Aligned Reinforcement Learning Framework for Financial Sentiment Analysis

Financial sentiment analysis(FSA) plays a critical role in algorithmic trading and market risk assessment. However, most existing approaches define sentiment purely through linguistic polarity, ignoring whether such sentiment aligns with actual market reactions reflected in price movements. At the same time, commonly used financial sentiment datasets and training pipelines are not designed to support the end-to-end learning of sentiment signals grounded in market behavior. To address these limitations, we propose a market-aligned reinforcement learning(RL) framework for FSA. First, we align sentiment with post-news price movements and construct a dataset of over 16,000 annotated data across short- and medium-term horizons. Second, we develop a class-weighted RL algorithm that optimizes model predictions for consistency with market-aligned labels. Extensive experiments on both BERT-like models and generative models show that our approach consistently outperforms supervised fine-tuning, achieving improvements of up to 23% in Macro-F1 and 22% in accuracy. Furthermore, back-testing results demonstrate that models trained with our RL framework produce more stable and reliable trading signals than semantic sentiment baselines. Our code is publicly available.

2026 International Joint Conference on Neural Networks (IJCNN), 1-8. IEEE