White paper · 7 min read

What Earnings Calls Reveal: Reading Management Language Systematically

Why the language executives use on earnings calls carries information about future results, why finance-specific methods beat generic sentiment tools, and where text signals go wrong.

SharpeMetrix Research · November 16, 2026

Four times a year, the management of most listed companies spends an hour describing results, answering analysts and setting expectations. The prepared remarks are reviewed by lawyers and investor relations. The question-and-answer session is less controlled. Both are recorded and transcribed.

SharpeMetrix’s Transcripts signal reads these calls systematically. This paper explains why management language carries information, how the method differs from generic sentiment analysis, and where text-based signals tend to fail.

Why language carries information

Financial statements report what happened. Earnings calls add what management thinks will happen and how confident it is. Three features make calls useful as data.

Calls contain soft information. Guidance, commentary on demand, explanations of margin changes and descriptions of the order book are not in the income statement. Some of this information reaches prices quickly. Some does not, especially when it is spread across a long transcript and expressed in hedged language.

Q&A is less scripted. Analysts ask follow-up questions that management did not choose. Price, Doran, Peterson and Bliss (2012) found that the tone of conference calls helps explain abnormal returns around the call and in the following weeks, and that the Q&A portion was particularly informative.

Language is hard to manage consistently. Executives can choose their words in a press release. Over many calls and many answers, word choice and hedging tend to reflect what management actually expects. Larcker and Zakolyukina (2012) found that language in calls preceding later restatements differed measurably from other calls, with more references to general knowledge and fewer references to shareholder value. Mayew and Venkatachalam (2012) went beyond words, finding that vocal cues of positive and negative affect in managers’ speech were associated with future firm performance.

The broader idea that text moves prices is not new. Tetlock (2007) found that pessimistic language in a widely read newspaper column predicted short-term downward pressure on market prices, followed by a reversal. Earnings calls are more useful than general media for stock selection because they are company-specific and come directly from the people running the business.

None of this means management is trying to mislead. Most of the information is simply detailed, qualitative and slow for the market to absorb across thousands of companies.

Finance-specific dictionaries versus generic sentiment

The simplest text method counts positive and negative words. The quality of the result depends on the word list.

Loughran and McDonald (2011) showed that a widely used general-purpose dictionary classified almost three-quarters of the “negative” words found in 10-K filings as negative when they were not negative in a financial context. Common examples include liability, tax, cost, capital, board and depreciation. These words describe the normal mechanics of a business. A generic tool scores a capital-intensive company as gloomy because it talks about its assets and liabilities.

They built finance-specific word lists (negative, positive, uncertainty, litigious and others) and showed that the finance-specific negative list was more closely related to market reactions than the generic one. Their lists are now a standard starting point in academic work.

ApproachStrengthWeakness
Generic sentiment dictionarySimple, transparentMisreads routine financial vocabulary
Finance-specific dictionaryTransparent, built for financial textTreats each word independently; misses context
Word embeddings and machine learningCaptures context and phrasingHarder to inspect; more prone to overfitting

SharpeMetrix’s engine combines the first-generation and later methods. It blends bag-of-words analysis, of the kind Loughran and McDonald formalised, with word embeddings that represent words by the contexts they appear in, and feeds the results into a machine learning model. The engine was trained on more than 200,000 earnings calls since 1999. It evaluates:

  • management tone and confidence,
  • word-choice patterns and linguistic indicators such as hedging,
  • forward-looking statements and the quality of guidance,
  • each call compared with the company’s own previous calls.

Tone change versus tone level

Some companies always sound upbeat. Some management teams are habitually cautious. Some industries use more negative vocabulary because of what they do: an insurer talks about losses and claims as a matter of course. A raw tone score therefore mixes information with style.

Comparing a company with itself removes much of this. A cautious management team that becomes noticeably more confident, or an upbeat one that starts hedging, is more informative than the absolute level of either. Cohen, Malloy and Nguyen (2020) found a related effect in annual reports: firms that changed the language of their 10-K filings substantially from the prior year tended to have lower subsequent returns, and the market was slow to notice these changes.

In practice this means three comparisons:

  1. Against the company’s own history. Is this call more or less confident than the last several?
  2. Against peers. Is the shift company-specific, or is the whole industry sounding more cautious?
  3. Between sections. Does the Q&A match the tone of the prepared remarks?

A worked illustration, without reference to any real company: two industrial firms each score in the 60th percentile of tone this quarter. Firm A has scored near the 60th percentile for three years. Firm B averaged the 85th percentile and has dropped for two calls in a row, with more uncertainty words in its answers on demand. The levels are the same. The information is not.

Pitfalls

Text signals have real weaknesses. Advisors evaluating any language-based strategy should ask how each of these is handled.

Look-ahead bias. A signal must be built only from transcripts available at the time. Dictionaries and models trained on later data can leak future information into historical tests.

Overfitting. Language offers a very large number of possible features. With enough features, a model can fit historical returns by chance. Harvey, Liu and Zhu (2016) argued that the standard for accepting a new return predictor should be higher than the conventional one, given how many have been tested. Long histories, out-of-sample testing and simple, stable model structure all help.

Boilerplate and coaching. As text analysis became common, firms learned which words attract attention. Prepared remarks have become more uniform. This is one reason the Q&A section and changes over time tend to be more useful than the scripted portion.

Context and negation. “Not a significant concern” contains a negative word and a positive meaning. Word-count methods miss this; context-aware models do better but are harder to inspect.

Decay. McLean and Pontiff (2016) found that return predictability weakens after academic publication. Simple tone measures are now widely known. Durable value is more likely to come from careful construction than from the basic idea.

Concentration. Text signals can cluster in particular sectors during particular periods, for example when an entire industry turns cautious. Portfolio construction limits, such as SharpeMetrix’s ±7% sector bounds relative to the benchmark, keep a language signal from becoming an unintended sector bet.

How the Transcripts score is used

The Transcripts family produces a 0–100 score for each company after each call, ranked against comparable companies. It is one of seven families in the SharpeMetrix composite. On its own, a text signal is noisy: one call is one data point, and management tone can shift for reasons that have nothing to do with the business. Combined with valuation, quality, momentum, volatility, earnings expectations and innovation, it contributes information the other six do not use.

That is the main argument for alternative data in a model portfolio. The source matters less than the overlap. A signal drawn from what management says is unlikely to repeat what price, accounting and analyst data already show.

References

  • Cohen, L., Malloy, C., & Nguyen, Q. (2020). Lazy prices. Journal of Finance, 75(3), 1371–1415.
  • Harvey, C. R., Liu, Y., & Zhu, H. (2016). …and the cross-section of expected returns. Review of Financial Studies, 29(1), 5–68.
  • Larcker, D. F., & Zakolyukina, A. A. (2012). Detecting deceptive discussions in conference calls. Journal of Accounting Research, 50(2), 495–540.
  • Loughran, T., & McDonald, B. (2011). When is a liability not a liability? Textual analysis, dictionaries, and 10-Ks. Journal of Finance, 66(1), 35–65.
  • Mayew, W. J., & Venkatachalam, M. (2012). The power of voice: Managerial affective states and future firm performance. Journal of Finance, 67(1), 1–43.
  • McLean, R. D., & Pontiff, J. (2016). Does academic research destroy stock return predictability? Journal of Finance, 71(1), 5–32.
  • Price, S. M., Doran, J. S., Peterson, D. R., & Bliss, B. A. (2012). Earnings conference calls and stock returns: The incremental informativeness of textual tone. Journal of Banking & Finance, 36(4), 992–1011.
  • Tetlock, P. C. (2007). Giving content to investor sentiment: The role of media in the stock market. Journal of Finance, 62(3), 1139–1168.

More research

See it with your own book

Every model, every holding, every trade. Open for 30 days.

Full platform access with no credit card. Or book a 20-minute walkthrough and we'll map the models to your current lineup.