Reproduce “Innovative Sentiment Analysis and Prediction of Stock Price Using FinBERT, GPT-4 and Logistic Regression: A Data-Driven Approach” on OpenResearch

About this paper

Olamilekan Shobayo, Sidikat Adeyemi-Longe, Olusogo Popoola, Bayode Ogunleye

Researchers from Sheffield Hallam University and the University of Brighton evaluated FinBERT, GPT-4, and Logistic Regression for financial news sentiment analysis and stock index prediction using Nigerian financial news. The study found that a simpler Logistic Regression model achieved superior performance, with an 81.83% accuracy and 89.76% ROC AUC, significantly outperforming FinBERT (63.33% accuracy) and GPT-4 (54.19% accuracy) on this specific task.

Key insights

• A simpler, well-tuned Logistic Regression model unexpectedly outperformed more complex state-of-the-art models like FinBERT and GPT-4 for financial news sentiment analysis and stock prediction on the given dataset.
• The performance of advanced NLP models like GPT-4 can be limited by a lack of flexibility, adaptability to specific data patterns, or insufficient understanding of subtle, context-specific financial terminology when using predefined approaches.
• Effective feature engineering and pragmatic model selection, rather than sheer model complexity, are crucial for optimal performance in dynamic financial market applications.

Summary by alphaXiv.