By Belen Telahun

Faculty Mentor: Melody Denhere

Abstract

This study examines whether the tone of financial news can predict the next-day return direction of large-cap technology stocks, specifically focusing on Meta Platforms, NVIDIA, and Intel. The primary objective is to determine if sentiment extracted from news offers additional predictive value beyond standard market indicators. The analysis integrates more than 61,000 news articles from 2021 to 2025, assigning sentiment scores using a finance-specific language model (FINBERT). The dataset is meticulously aligned to prevent look-ahead bias, and logistic regression models are used to compare a baseline specification with an enhanced version that incorporates sentiment features.

The results show that news sentiment does not significantly improve predictive performance. The baseline model achieves 51.73% accuracy and a ROC-AUC of 0.532, while sentiment-augmented models offer only marginal enhancements (accuracy: 0.523–0.547; ROC-AUC: 0.543–0.548). Sentiment variables are statistically insignificant (p > 0.10) and display near-zero correlations with returns (0.02–0.03). In contrast, short-term volatility is significant at the 5% level (p < 0.05) and consistently improves model fit. Model diagnostics indicate good calibration (Hosmer–Lemeshow p-values: 0.34–0.91) but reveal minimal explanatory power, with little change in deviance or AIC. Overall, these findings suggest that daily news sentiment is noisy and rapidly absorbed into market prices, limiting its effectiveness as a short-term signal. The results underscore the challenges of using aggregated sentiment data for practical forecasting and highlight the need for more refined and targeted methodologies in future research.