Quantifying the Value of Fine-Tuning: A Reproducible Framework for Financial Sentiment Analysis in the Technology Sector

Abstract

Financial sentiment analysis in the technology sector remains difficult because market-moving news combines firm-specific events, macro signals, and highly technical vocabulary. This study evaluates whether task-specific fine-tuning improves on zero-shot use of FinBERT when the corpus is restricted to English-language Technology news collected from MarketAux between 1 January and 30 June 2025. The final gold set contains 894 manually reviewed headline-description pairs labeled as negative, neutral, or positive and split into 625 training observations, 134 validation observations, and 135 test observations. On the held-out test set, zero-shot FinBERT reaches an accuracy of 0.637 and a macro F1 score of 0.626, whereas the fine-tuned version reaches 0.585 and 0.532. An exact McNemar test does not indicate a statistically significant difference between the two classifiers (p = 0.450). The findings suggest that, under a small sector-focused design and consumer-grade computational constraints, fine-tuning is not automatically superior to a strong domain-pretrained baseline.