The impact of financial news sentiment on stock market volatility
Ahmed, Saima (2026)
Pro gradu -tutkielma
2026
School of Business and Management, Kauppatieteet
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe20260617100072
https://urn.fi/URN:NBN:fi-fe20260617100072
Tiivistelmä
This thesis examines whether firm-level financial news significantly predict short-term stock market volatility (proxied by the 5-day rolling standard deviation of daily stock returns). Existing literature focuses predominantly on aggregate indices rather than firm-level dynamics, with limited out-of-sample validation using transformer-based NLP. In this study, FinBERT is applied to Reuters headlines for 15 large-capitalization U.S. companies (March 2023–March 2026), constructing four daily sentiment aggregations mean, sum, maximum, and count. Ordinary Least Squares regression model with a chronological 80/20 train-test split is used to investigate the relationship between financial news and stock market volatility. Performance is evaluated using in-sample R², RMSE, and out-of-sample R_"OOS" ^2.
Findings reveal limited and predominantly firm-specific explanatory power. The count aggregation method achieves the strongest in-sample fit (14 firms significant as p-values below the 5% significance level) among all other methods but captures news volume rather than sentiment direction and deteriorates out-of-sample. The sum aggregation method is the most informative tone-based aggregation, yielding positive out-of-sample performance for Exxon Mobil (R_"OOS" ^2 = +0.023) and Amazon (R_"OOS" ^2 = +0.026). Energy and healthcare firms exhibit the most stable relationships, while technology and financial firms are largely unresponsive at the daily horizon. These patterns are broadly consistent with the semi-strong Efficient Market Hypothesis.
Three contributions emerge from this study: it establishes that news volume and sentiment tone are distinct predictive dimensions; it demonstrates that choice of aggregation method influence the magnitude and direction of estimated effects; and it shows that in-sample significance substantially overstates practical forecasting value. Sector-stratified cumulative-intensity models represent the most promising direction for future research.
Findings reveal limited and predominantly firm-specific explanatory power. The count aggregation method achieves the strongest in-sample fit (14 firms significant as p-values below the 5% significance level) among all other methods but captures news volume rather than sentiment direction and deteriorates out-of-sample. The sum aggregation method is the most informative tone-based aggregation, yielding positive out-of-sample performance for Exxon Mobil (R_"OOS" ^2 = +0.023) and Amazon (R_"OOS" ^2 = +0.026). Energy and healthcare firms exhibit the most stable relationships, while technology and financial firms are largely unresponsive at the daily horizon. These patterns are broadly consistent with the semi-strong Efficient Market Hypothesis.
Three contributions emerge from this study: it establishes that news volume and sentiment tone are distinct predictive dimensions; it demonstrates that choice of aggregation method influence the magnitude and direction of estimated effects; and it shows that in-sample significance substantially overstates practical forecasting value. Sector-stratified cumulative-intensity models represent the most promising direction for future research.