Integrating price and textual data for next-day stock movement prediction: a study using StockNet dataset

This research enhances next-day stock movement prediction by integrating price data with social media text from Twitter using the StockNet dataset. We employ diverse preprocessing techniques and modeling approaches, including advanced embedding methods, sentiment analysis, and generative adversarial networks (GANs). Results reveal notable performance trends across embedding techniques and sentiment-based methodologies. Our adversarially fine-tuned BERT model, trained on the Financial Phrasebank dataset, achieves top performance with 74.2% accuracy and 74.1% macro F1, surpassing other techniques. Introducing thresholds significantly improves model performance, with accuracy increasing from 70 to 76.4% and macro F1 score from 70 to 76.3%. Ensemble voting methods further enhance predictive accuracy, reaching 77% with thresholds applied. The integration of GANs provides modest additional improvements. Our approach achieves an F1 score of 78.5%, substantially outperforming previous benchmarks on the same dataset (57.5–60.5%). These findings underscore the importance of data integration, sophisticated modeling methodologies, and careful threshold optimization in achieving accurate and timely stock market predictions.