33 Pages Posted: 12 Nov 2013
Date Written: November 10, 2013
Text sentiment extracts text’s attitude by counting negative words and has proved extremely useful in a variety of contexts. The literature interprets it in three ways: quantitative information, soft news, and psychological sentiment. We use a quasi-natural experiment to show that text sentiment reflects primarily omitted quantitative information and does not capture soft news or sentiment. We first extract text sentiment from earnings call transcripts with dictionary and supervised-learning methods, and then compare how it predicts returns during overnight and intraday calls; specifically, whether text sentiment explains a larger portion of stock returns for overnight calls. The overnight and intraday cases differ only in the timing of a quarterly report. Overnight calls are dominated by quantitative news from a quarterly report, while the intraday cases contain mostly soft news and sentiment. Text sentiment explains overnight returns well but fails to predict returns or volatility during intraday calls. Thus, text sentiment reflects news only during periods dominated by quantitative information.
Keywords: Text sentiment, Textual analysis, Earnings announcements, Soft information
JEL Classification: G12, G14, M41
Suggested Citation: Suggested Citation
Chebonenko, Tatiana and Muravyev, Dmitriy, What Does Text Sentiment Really Measure? Evidence from Earnings Calls (November 10, 2013). Available at SSRN: https://ssrn.com/abstract=2352524 or http://dx.doi.org/10.2139/ssrn.2352524