Accurate stock price forecasting is increasingly dependent on models that can integrate heterogeneous market signals beyond traditional quantitative indicators. Classical sentiment analysis tools and lexicon-based methods often fail to capture the contextual and domain-specific nuances of financial language, limiting their predictive contribution. To address these limitations, this study introduces a multimodal forecasting framework that leverages FinGPT, an advanced financial large language model specifically trained for market-related text understanding. FinGPT is employed to extract high-resolution sentiment polarity scores from Twitter data, replacing conventional sentiment models and providing a more robust representation of investor psychology. These sentiment features are fused with historical stock prices and technical indicators to form a comprehensive input space. The integrated dataset is used to evaluate four deep learning architectures: CNN, LSTM, Attention-based LSTM (ALSTM), and a newly proposed Hybrid ALSTM-CNN model with an attention mechanism. Experiments performed on eight major NASDAQ and NYSE stocks demonstrate that incorporating FinGPT-derived sentiment significantly enhances forecasting accuracy across all models. The Hybrid model consistently achieves the best performance, effectively capturing both long-term dependencies and short-term market dynamics. Results confirm that FinGPT-driven sentiment modeling provides substantial predictive value compared with traditional sentiment approaches, enabling more realistic, behavior-aware stock price forecasting. For instance, in the case of AMD stock, it recorded the lowest MSE (46.06) and the highest R² (0.624), highlighting its superior ability to capture nonlinear dependencies and market dynamics. Overall, the proposed sentiment-enhanced hybrid framework provides a more robust and realistic procedure for stock price prediction by integrating quantitative and qualitative market signals.
Golabzaei et al. (Sun,) studied this question.