The undeniable impact of social media has revolutionized the dissemination of news and the simplified understanding of market sentiment, adversely shaping investors and market trading decisions. Consequently, examining news data scraped from social media platforms such as Twitter necessitates adequate processing of both textual and non-textual features. This study proposes a hybrid approach for stock market trend prediction that leverages historical price and tweet sentiment data, utilizing a rank-based weighting with time decay (RBWTD) method for non-textual features (viz. views, likes, retweets, and replies) based on the herd behavior, where individual investors mimic the actions of a crowd sharing and endorsing a common market perspective. The model architecture comprises two modules: a sentiment module employing the FinBERT model, specifically designed for handling financial sentiment data, classifying it into positive, neutral, or negative classes, and calculating the weighted sentiment score using the RBTWD for both textual and non-textual features improving on the typical overall sentiment score used. The price module utilizes temporal convolutional networks (TCNs) and gated recurrent units (GRUs) to handle and classify the stock trends into buy and sell labels. Experiments evaluated the model’s performance on multiple stocks from the STOCKNET dataset to predict one of either class: buy or sell. The proposed model was evaluated across the three time-series cross-validation folds and time-steps (7, 14, and 28 days). The experimental results demonstrate the model’s efficacy in achieving an average accuracy of over 86% for most stocks on shorter steps. Additionally, an ablation study confirmed the efficacy of the weighted sentiment score outperforming benchmarks across the various metrics. Statistical validation using the Wilcoxon signed-ranked test also substantiated the potential effectiveness of the proposed model in stock trend prediction applications.
Friday et al. (Sun,) studied this question.