This paper introduces a transparent framework for extracting interpretable alpha signals from financial tweets by leveraging large language models (LLMs) to perform multi-label event classification. Unlike traditional sentiment analysis that only captures polarity, the proposed method assigns fine-grained event categories (e.g., rumor/speculation, retail investor buzz, brand boycott) to each tweet, enabling the construction of structured, event-driven factors. Using a dataset of over 85,000 company-related tweets, the authors build cross-sectional factors by aggregating net tone scores per event label and evaluate their predictive power through portfolio sorts and long-short backtests over 1- to 7-day horizons.

The empirical results reveal that certain event labels, particularly Speculation/Rumor and Geopolitical Tension, consistently generate statistically significant negative alpha, with Sharpe ratios as low as -0.38 and information coefficients exceeding 0.05. These findings suggest that high volumes of discussion around specific themes act as contrarian indicators. The study also demonstrates that the predictive power of these event-based factors is orthogonal to market beta, confirming they represent genuine alpha rather than exposure to known risk factors. Additionally, a baseline lexicon-based sentiment strategy achieved an impressive Sharpe ratio of 5.0, underscoring the overall value of social media sentiment.

A key contribution of this work is its commitment to transparency and reproducibility: all code, methodologies, and processed data are publicly available. The framework provides a principled path toward next-generation behavioral finance research, bridging advances in NLP with interpretable quantitative modeling. Future work includes extending the framework to other text sources, fine-tuning domain-specific LLMs, and modeling temporal dynamics of narrative decay.