A Transformer-Based Stock Trend Prediction with Sentiment Analysis of Relevant and Urgent Tweets

Stock market price movement primarily focuses on accurately classifying buy and sell signals to enable traders to make informed trading decisions. The integration of sentiment analysis into stock trading modeling has become increasingly vital for capturing behavioral signals that influence market trends. In this study, we propose a dual-scoring sentiment framework for tweet sentiment analysis for stock price movement classification, including a relevance score that captures the semantic proximity of tweet content to financial actionability and an urgency score using domain-specific keywords that signal imminent market reactions. The sentiment classification module is based on a frozen FinBERT encoder architecture, ensuring the stability of pretrained financial language representations while adapting the output layer to emotion and sentiment labels. A modified transformer-based architecture is designed and employed, which fuses the historical price features and the tweet attributes for the final stock movement classification. Experiments conducted on the STOCKEMOTION dataset containing over 35,000 labeled financial tweets and synchronized with stock price data demonstrate the efficacy of the proposed model achieving substantial improvements across all five considered stocks taken (AAPL, AMZN, META, MFST, TSLA) metrics with an average accuracy (92.72%), precision (91.56%), recall (94.00%), F1-score (92.76%), AUC (0.98), and matthews correlation coefficient (MCC = 0.89). These results highlight the potential of incorporating urgency-aware and relevance-driven signals in real-world financial sentiment models and stock movement prediction pipelines.