Understanding Sentiment Through Context

We examine whether empirical results using text-based sentiment of U.S. annual reports depend on the underlying context, within documents, from which sentiment is measured. We construct a clause-level measure of context, showing that sentiment is driven by many different contexts and that positive and negative sentiment are driven by different contexts. We then construct context-level sentiment measures and examine whether sentiment works as expected at the context-level across four prediction problems. Our results demonstrate that document-level sentiment exhibits significant noise in prediction and suggest that document-level aggregation of sentiment leads to missed empirical nuances. The contexts driving sentiment results vary substantially by outcome, suggesting lower empirical internal validity for document-level sentiment. Using three additional sentiment measures, we document the same inferences, concluding that document-level aggregation likely leads to lower internal validity. Sentiment is thus best applied at the level of specific contexts rather than across whole documents