Stock Price Prediction Using News Article Sentiment Analysis: A Comparative Study of NLP Classification Techniques

Abstract

This paper presents an enhanced approach for predicting stock price movements by leveraging sentiment analysis of financial news articles. We evaluate a range of classifiers, including traditional lexicon-based methods, Naive Bayes, and Support Vector Machines (SVM), alongside modern deep learning architectures such as FinBERT and Long Short-Term Memory (LSTM) networks. Using the Reuters news corpus combined with Quandl stock price data, we conduct comprehensive experiments to assess classifier performance across multiple temporal prediction horizons, including one-day and two-day forecasts. Experimental results demonstrate that FinBERT consistently outperforms classical models, achieving an accuracy of 65.7% for two-day price movement predictions. Additionally, we perform rigorous statistical significance testing to validate the robustness of our results, accompanied by detailed error analysis to identify common sources of misclassification. We also investigate the impact of varying labeling thresholds on classifier performance, highlighting the trade-offs between precision and recall in sentiment-driven prediction tasks. Our findings suggest that while sentiment analysis provides valuable signals for stock movement prediction, integrating domain-specific Named Entity Recognition (NER) and contextual modeling is essential to fully capture the complexities inherent in financial text. Future work will explore incorporating event-driven representations and multi-modal data sources, such as social media and market indicators, to enhance predictive accuracy and reliability further.