Decoding the news: a sentiment-driven comparison of crisis-period market responses

Gund, Reshmi Santosh; Dey, Maitreyee; Rana, Soumya Prakash

London Met Repository

Lists

Tools

Gund, Reshmi Santosh, Dey, Maitreyee and Rana, Soumya Prakash (2026) Decoding the news: a sentiment-driven comparison of crisis-period market responses. In: 14th International Conference on Frontiers of Intelligent Computing: Theory and Applications (FICTA-2026), 8-9 June 2026, London (UK) / Online. (In Press)

Abstract
Documents
Details
Record

[+][-]

Abstract

Financial markets do not react only to numbers; they react to news. Over the last six years, it has been observed that a single headline about a war, a tariff announcement, or a wave of artificial intelligence (AI) layoffs can move billions in market value within hours. This paper investigates whether automatically measured financial-news sentiment is associated with S&P 500 returns across different crisis types. A two-phase system is developed: first, six sentiment classification approaches are compared on the FinancialPhraseBank (FPB) benchmark, ranging from a financial lexicon and classical machine-learning baselines to fine-tuned transformers and a weighted ensemble; second, FinBERT, the best individual transformer model, is applied to 32,405 financial news headlines collected from the Global Database of Events, Language and Tone (GDELT) between March 2020 and March 2026, covering COVID-19, the Russia-Ukraine war, Israel-Gaza, US tariff escalation, AI-driven layoffs, and the Iran crisis. The weighted ensemble achieved 96.0% macro-F1, with FinBERT at 93.3% and RoBERTa at 91.7%. McNemar tests showed that the ensemble's improvement over FinBERT was not statistically significant (p = 0.22), while FinBERT significantly outperformed RoBERTa (p = 0.034). Daily sentiment scores were compared with S&P 500 log returns using Pearson correlation, Spearman correlation, and Granger-causality tests. Israel-Gaza showed an FDR-significant sentiment-return association (r = 0.27, q = 0.030), while Russia-Ukraine showed a weaker exploratory association (r = 0.16, q = 0.078) that did not survive the 5% FDR threshold. No event-period Granger results survived FDR correction. These findings suggest that sentiment-market coupling is contingent on crisis type rather than universal.

Documents