Understanding Sentiment Through Context

Abstract

We examine the extent to which results based on financial sentiment of U.S. annual reports are conditional on the underlying context from which financial sentiment is derived, as well as the extent to which financial sentiment is related to the underlying context of the annual report. To achieve this, we construct a measure of context that is based on the grammar, syntax, and content of sentences in each report. We then apply sentiment measures to the phrases within each context to examine how sentiment is related to each context, and under which contexts financial sentiment works as expected or not for a variety of prediction problems. We show that sentiment encompasses a wide variety of contexts, and that positive and negative sentiment respond to different contexts. In addition, we show that there is significant noise in predicting various outcomes (stock return, volume, volatility, and material weaknesses). Specifically, only select contexts drive the primary results of each analysis, and these select contexts vary by the outcome being predicted. Furthermore, under some contexts we find results opposite to expected predictions, indicating a nontrivial amount of systematic noise or error in sentiment classification.