Aligning social signals with market outcomes: A reinforcement learning approach to noisy sentiment data in stock forecasting

Abstract

Stock price prediction relies heavily on market sentiment, but it is still difficult to identify truly meaningful signals from social media because of the volume of noise and flimsy engagement measurements. In this paper, sentiment signal extraction is formulated as a delayed reward recommendation issue using a unique reinforcement learning framework. The method uses actual market input instead of static labels by simulating the task as an agent that learns to suggest tweets based on their potential to increase predicted accuracy. Due to the scarce and delayed nature of market signals, the agent may differentiate between those that are useful and those that are misleading or irrelevant by using a reward function that aligns with 48-hour post-publication stock movements. When combined with a Long Short-Term Memory (LSTM) network for price forecasting, the suggested approach shows that, in contrast to employing sentiment variables directly, integrating the RL-based suggestion enhances prediction performance. Tests conducted on Apple stock data from 2014 to 2016 demonstrate that using agent-selected tweets improves the system's R-squared scores and reduces prediction errors. The findings indicate that dynamically adjusting comment selection in unstable financial settings can be achieved using reinforcement learning. This study combines market-based incentive design, financial text mining, and reinforcement learning to provide a viable method for improving the reliability of sentiment-driven stock predictions. To further validate this framework's robustness, future research may expand it to include more equities and explore integrating it with other market signals.