Dual stream deep learning for fake news-aware stock prediction: Integrating technical indicators and sentiment analysis

Stock prediction remains challenging for investors to achieve substantial returns. Although artificial intelligence models have emerged, they often focus on historical price data, while fewer studies jointly integrate technical indicators and sentiment analysis. Furthermore, many of the modern sentiment analysis techniques overlook the detection of fake news, making predictions unreliable. To address these limitations, this study proposes a fake news-aware dual-stream predictive framework that integrates daily return features, technical indicators, and sentiment analysis via feature-level concatenation. The methodology follows a dual-parallel architecture; the model filters stock-related tweets using a BERT-based fake news detection module, followed by a RoBERTa model, a one-directional convolutional neural network (1D-CNN), and a Bi-directional Long-Short-Term Memory (Bi-LSTM) to extract daily sentiment features. Simultaneously, the model processes historical stock prices and technical indicators. Finally, a fully connected neural network processes unified features to predict stock prices. Evaluated on 88 stocks, the model achieved a mean accuracy of approximately 81%. For Apple Inc. (AAPL), which represents the best-performing single-stock case, the mean accuracy reached 93.44%. Overall, the model demonstrated improved performance on benchmark datasets, particularly for stocks with high social media activity, achieving higher average Accuracy, F1-score, AUC, and MCC compared to recent models under the same evaluation setting.