Integrating sentiment analysis with CNN–LSTM models for stock price forecasting

Stock price prediction remains challenging due to markets’ non-linear, volatile nature. While existing studies analyse numerical data or textual news separately, their synergistic integration remains underexplored. We propose a novel dual-stream deep learning framework combining wavelet-processed price data with Financial Bidirectional Encoder Representations from Transformers (FinBERT)-based news sentiment analysis through intermediate feature fusion. Our architecture employs convolutional neural networks–long short-term memory (CNN–LSTM) networks to model cross-modal temporal dependencies. Comprehensive evaluation demonstrates superior performance (test mean squared error (MSE) = 0.0220), highlighting the value of multimodal integration for financial forecasting. Results suggest that this approach enhances predictive accuracy in algorithmic trading, though performance depends on data quality and temporal alignment between information sources.