Discrepancy Between Construct and Predictive Validity in Financial Sentiment

Testing five sentiment instruments including FinBERT and LLM annotators against a 23-year corpus of securities class actions, this study finds that predictive validity for one-day stock return leads depends heavily on sampling conventions. Results show human agreement aligns better with same-day associations than with predictive one-day signals.