When it comes to financial disclosure, headline numbers are not all that matters. Equally important are what company executives talk about – and even how they talk about it. A large and growing body of literature in finance and accounting establishes the usefulness of narrative disclosure for predicting a company’s future performance and returns. Many of these papers attempt to extract predictions from text using lexicon or machine learning approaches and, in general, interpret their results as capturing “sentiment.”
While it is clear that this sentiment, however measured, is useful, it is not clear why. One classic paper interprets the usefulness of text as capturing “otherwise hard-to-quantify aspects of firms’ fundamentals”. But what is the nature of these hard-to-quantify aspects of firms’ fundamentals? After all, firms provide lengthy communications to investors through annual reports, earnings press releases, earnings calls, and special announcements, along with their numerical disclosure.
In a recent paper, we argue that information in narrative disclosure comes in two distinct forms – facts and opinions – identify facts and opinions at the sentence level, and show that this distinction matters. We focus on earnings calls, a market-relevant and timely source of information about public firms. In earnings calls, managers present facts that can be hard to convey numerically and opinions that interpret facts. In principle, both can be informative to investors and would affect firm fundamentals and returns.
We suggest a new methodology that allows us to automate the classification of sentences into “objective” and “subjective.” The first category contains sentences stating facts that are, in principle, falsifiable; for example, “We opened a new store in Washington DC.” The second category contains sentences that cannot be falsified; for example, “We had a great quarter.” The last category contains sentences that are irrelevant to the facts or opinions distinction; for example, “We now turn to slide 13.” We develop this methodology by manually tagging a large number of sentences from earnings calls and then training a machine-learning model to capture the relevant text attributes associated with each.
It is important to note that this approach merges the benefits of human intuition with the statistical or machine-learning approaches that are often treated in the literature as being at odds with each other. That is, we identify the object of interest through human tagging of text and only then scale it by finding the associated word attributes. We do that precisely because there is no empirical outcome that can tell us whether the disclosure was objective or subjective in a way that returns can be used to train models that classify disclosure as good or bad news, for example. This allows us to go further than the standard sentiment lexicons and distinguish between statements that are positive in facts (“Revenues improved by 10 percent”) and positive in coverage (“Revenues improved substantially”).
Having identified facts and opinions in earnings calls, we show that this distinction helps up understand the heterogeneity of disclosure across firms and executives and that it has important implications for the incorporation of information by investors. First, despite the potential view that only facts should matter, subjective sentences make up roughly half of earnings calls. This composition appears to be stable over the sample period but varies systematically with the complexity of the company — growth and large firms use more subjective language than value and small firms, with a similar variation being observed across industries.
Second, executives’ language varies systematically, with CEOs using more subjective language than CFOs and both using more subjective language during the Q&A section than in the opening remarks section. Examining changes in executives over time, we find that the prevalence of opinions in CEO communication depends more on who the CEO is rather than the identity the company, while the opposite is true for CFOs. This is consistent with the idea that CEOs have more leeway than CFOs in how they communicate with investors.
Third, we examine how the form of communication affects the incorporation of information. A large body of literature examined how the release of information affects investors’ disagreement. We contribute to that literature by showing that more subjective information is associated with higher disagreement, as measured by standard proxies for disagreement (for example, the amount of trading activity) or by more direct measures of retail traders’ disagreement based on StockTwits data. Next, we link the form of communication to anomalous returns (predictable returns non-zero market-adjusted returns attributable to various factors described in the literature). Previous work has found that anomalous returns are more pronounced around the release of earnings and argued that the release of public information pushes investors’ beliefs closer to the truth, thereby sending the prices of under (over) valued firms up (down). We repeat the analysis of that work while separating instances when the information is more and less subjective. We find strong evidence that the high anomalous returns are associated with the release of subjective information.
Finally, we apply machine learning to study how the language in objective and subjective sentences, separately, is linked to the accounting performance and returns of the announcing firms. We apply our model and find that accounting performance in the quarter discussed in the earnings call corresponds more closely to facts presented in the call, rather than opinions expressed in it. However, facts and opinions are both necessary to explain market returns following the announcement and future firm performance, suggesting that even opinions about past events are forward-looking in nature.
Earnings calls contain about an equal mix of facts and opinions. We find evidence suggesting that investors find it harder to agree on the significance of opinions but also find more opportunities to take advantage of anomalous returns when the disclosure is more subjective. When it comes to explaining current and future firm performance, facts and opinions play different roles. Facts convey information about both current and future firm performance. In contrast, opinions expressed by managers don’t have a strong correlation with how the firm is doing now, but they are as informative as facts about the future.
 TETLOCK, P.C., SAAR-TSECHANSKY, M. and MACSKASSY, S. (2008), More Than Words: Quantifying Language to Measure Firms’ Fundamentals. The Journal of Finance, 63: 1437-1467. https://doi-org.philfedlib.idm.oclc.org/10.1111/j.1540-6261.2008.01362.x
 See COOKSON, J.A. and NIESSNER, M. (2020), Why Don’t We Agree? Evidence from a Social Network of Investors. The Journal of Finance, 75: 173-228. https://doi-org.philfedlib.idm.oclc.org/10.1111/jofi.12852
 ENGELBERG, J., MCLEAN, R.D. and PONTIFF, J. (2018), Anomalies and News. THE JOURNAL OF FINANCE, 73: 1971-2001. https://doi-org.philfedlib.idm.oclc.org/10.1111/jofi.12718
This post comes to us from Shimon Kogan at DC Herzliya – Arison School of Business and the University of Pennsylvania – The Wharton School and from Vitaly Meursault at the Federal Reserve Bank of Philadelphia. It is based on their working paper, “Corporate Disclosure: Facts or Opinions?” available here. The views in the post and the paper are solely those of the authors and do not necessarily reflect the views of the Federal Reserve Bank of Philadelphia or the Federal Reserve System. Any errors or omissions are the responsibility of the authors.