News sentiment and price prediction on stocks and Bitcoin: a comparative analysis of machine learning models

Abstract

This study investigates the differential impact of news sentiment on price prediction models across traditional and Web3 financial markets, comparing S&P 500 and Bitcoin performance. Using a comprehensive dataset spanning from October 2017 to January 2025, we develop a four-stage sentiment analysis approach utilizing a fine-tuned RoBERTa model, which achieves 80.25% accuracy in classifying financial headlines into positive, negative, neutral, or invalid categories. We implement and compare multiple prediction models including Linear Regression, Ridge, LASSO, ElasticNet, Support Vector Machines (SVM), and SVM with Principal Component Analysis (PCA), both with and without sentiment features. Results indicate that regularized models significantly outperform standard approaches, with LASSO regression achieving the highest predictive accuracy for both assets (R² of 0.45 for S&P 500 and 0.52 for Bitcoin). Notably, sentiment features demonstrate a more significant impact to Bitcoin prediction models, particularly when combined with PCA, while traditional S&P 500 models generally perform better without sentiment indicators. For both assets, lagged sentiment variables consistently outperform same-day sentiment, suggesting delayed market responses to news. These results show that sentiment affects different markets in different ways. Crypto markets tend to react more strongly to sentiment shifts, while traditional markets are generally more efficient and less influenced by this kind of data. This difference provides useful guidance for designing prediction strategies that are better suited to each type of asset.