Predictive Power of Social Media: A Comparative Study of Classical ML, Lexicon-Based, and LLM Approaches for Football Sentiment Analysis

This paper presents a systematic comparison of four sentiment analysis approaches applied to social media discourse surrounding the England national football team during UEFA Euro 2024. We train a TF-IDF Logistic Regression baseline and its three-class extension on the TweetEval benchmark dataset (Barbieri et al., 2020; N = 57,763), evaluate VADER as a rule-based comparison, and apply Claude Haiku as a zero-shot LLM classifier with a domain-adapted prompt. All four models are applied to a corpus of 871 tweets from the tournament's final stages (July 7–21, 2024), covering England's semi-final victory over the Netherlands (2–1) and final defeat to Spain (1–2). The binary Logistic Regression achieves the strongest benchmark performance (accuracy = 0.862, AUC = 0.927), outperforming VADER (accuracy = 0.778) and the three-class extension (accuracy = 0.646). However, only Claude Haiku detects a statistically significant difference in public sentiment between match-win and match-loss days (t = 2.52, p = .012), while classical approaches fail to reach significance. Analysis reveals that 79.2% of tournament tweets are classified as neutral by the LLM, suggesting football Twitter is predominantly informational rather than emotionally charged. Cross-model agreement of 81.3% provides convergent validity for the observed sentiment pattern. These findings demonstrate that benchmark accuracy and domain sensitivity are dissociable properties of sentiment classifiers.