Temporal Modeling of Stock Directional Changes Using RNN and Attention Architectures with Sentiment Signals

Investors' opinions and attitudes toward an investment have become increasingly important in today's world. The increasing amount of platforms (such as Twitter) for sharing information has enabled people's sentiments (how they feel about something) to be leveraged as quantifiable data for academic analysis. The study conducted quantifiable methodology to ascertain whether or not sentiment as a standalone variable would be an appropriate metric to predict the direction of stock prices. The traditional method of using regression analysis to create predictive models was replaced with a binary-classification format where the sentiment around the company was combined with the previous day's closing price to produce a model predicting whether the stock would rise or fall. Tweets were collected about numerous companies and subjected to comprehensive NLP analysis (including tokenization, lemmatization, and removing stopwords) before being summed with the VADER model, creating a continually changing visual representation of investor sentiment. Once the models were developed, traditional machine-learning techniques and deep-learning architectures were utilized to compare performance. Overall the results showed significant improvements in predicting price movement direction by including investor sentiment as a metric in the pricing model compared with models based solely on historical data. The enhanced performance, particularly observed in temporal models like BiLSTM+Attention, provides robust evidence that human emotion, quantified through collective social discourse, significantly improves accuracy and flexibility in data-driven stock market prediction systems. This study validates the integration of behavioral insights as a critical, non-technical source of alpha in quantitative finance.