Earnings calls, deception and text analysis: questions answered
Plain answers about what EchoSignal measures, what the research says about deception cues and earnings call tone, and where the tool's limits are.
Can you tell if someone is lying from what they write?
Not reliably. Some word patterns are associated with deception in research, but no single word or pattern proves a lie.
Forensic statement analysis looks at patterns such as equivocation, negation, omissions and an unusually long opening section.
Practitioners stress that the cues interact with each other and with context, so a cue on its own is weak evidence.
Several reviews warn that results on statements written for experiments do not always carry over to real criminal cases.
A cue that appears in one statement may be ordinary in another, so compare against the speaker's usual wording where possible. Treat any pattern as a reason to look at the statement more closely, not as proof.
Forensic statement analysis is a set of methods for reading a written or spoken account for the way it is worded: what is included, what is left out, and how the speaker describes events. Its practitioners use it as an investigative aid, not as proof.
Methods such as SCAN (Scientific Content Analysis) look for omissions, changes in wording, and the sequence in which events are described.
Research on written witness statements finds that the weight of a cue depends on the other cues present and on the context, so no cue is read alone.
This is why EchoSignal scores are read alongside the wording, not as a checklist of cues. Each cue should prompt a question about the statement, not a conclusion about the person.
Does the tone of an earnings call predict the share price?
Tone is linked to share price reactions in some research, but that research does not establish that tone can predict share price moves you could trade on afterwards.
Druz, Wagner and Zeckhauser (NBER working paper) study negative tone in management communication and its link to share prices.
The Market page sets out the tests we run, including analyst downgrades and amended filings, so you can see what each one measures.
Each call is measured against its sector, so a market-wide move does not count as a reaction to the call. The day-of reaction and the moves after it are tested separately. Read the result alongside the sector comparison, not on its own.
The CEI is EchoDepth's score for how a piece of writing is delivered. It reads each sentence for its emotional content and combines the readings into one score.
For earnings calls, the score is given for the whole call and for its parts: the prepared remarks, the executives' answers in the Q&A, and the analysts' questions.
A CEI score describes how a passage is delivered. It is not a measure of truthfulness.
Compare like with like: the prepared remarks with other prepared remarks, and the Q&A with other Q&A. A score for one call is more useful against the same company's previous calls.
Use the score as a prompt for reading the transcript, not as a rating of the company. The score is most useful as one of several readings of the same call.
What do trust and sensitivity mean in EchoSignal?
They are scores from the EchoSignal text rules. Trust is 100 minus the density of sensitive, emotionally charged or evasive phrasing in short windows of text; a higher score means fewer such markers. Sensitivity is the strongest local concentration of those markers.
The score is built from the indicator rules listed on the EchoSignal Rules page. Each indicator contributes in proportion to its weight.
Both are descriptive scores. They describe the language of a call and are not forecasts of share price moves.
Compare passages from the same speaker or company over time, since a change is more informative than one score. Scores from different tools or different versions of EchoSignal are not directly comparable. A score is most useful when read with its neighbouring passages.
Does EchoSignal predict analyst downgrades?
Analyst downgrades are one of the outcomes we test. The scores are not presented as a predictor of them, and the tests are set out on the Market page.
The tests use 11,405 Russell 1000 earnings calls.
Downgrades are a stand-in outcome. They show how the market and analysts reacted, not whether anything in the wording was hidden.
A call counts as followed by a downgrade if an analyst downgrades the stock within 60 days of the call. Calls are included only where that full 60-day window is in the data.
Each call is also tested against its earnings surprise, which is a known influence on downgrades. This shows what the language adds beyond the numbers.
Read the downgrade test together with the earnings surprise test on the Market page, not on its own.
Can EchoSignal tell if management is lying?
No. EchoSignal describes how writing is delivered. It has not been validated against confirmed cases of deception, and it should not be used as evidence that anyone is lying.
Our tests use share price moves and analyst downgrades as outcomes. Neither is a confirmed deception outcome, and they are the most readily available stand-ins for it.
A proper test would need outcomes such as restatements or enforcement actions that are confirmed independently. We do not yet have that data.
The question that matters is whether a claim in the writing is later shown to be false. Tone alone cannot answer that.
EchoSignal can help you decide which passages to question, such as a sudden change in tone or a passage that avoids a specific question. Any finding about a person should come from evidence, not from a score.
Should I use EchoSignal to make investment decisions?
No. The scores are not forecasts, and the Market page publishes the tests behind them in full.
The tests are published on the Market page. They measure each call from the first trading day after it, against the stock's sector.
Use the scores to read how a call was delivered and what changed since the last call, alongside the numbers and your own judgement.
The scores are research output, not investment advice. Read them with the filings, the numbers and your own judgement, and take regulated advice for decisions about money.
The Market page shows the tests behind the scores, so you can check the method before you rely on any figure. Past patterns in the tests are no promise of future returns. Use the scores as one input among several.
How do I read an earnings call with EchoSignal?
Start with the earnings numbers against expectations, then use the scores to see where the call was unusually cautious, warm or evasive, and read those passages in the transcript.
Compare the scores for the prepared remarks with the executives' answers in the Q&A. A large gap can mean the answers were less confident than the script.
Look at what changed since the company's previous call. A change is more informative than a single score.
Treat each score as a prompt for questions, not as a verdict.
Note the tone of the opening remarks, since they are often scripted and set the baseline for the rest of the call. Check whether the same topics come up in both the prepared remarks and the Q&A.
Record what you find, so the next call can be compared with it.
Can the language of an earnings call reveal financial misreporting?
Some research links language and vocal cues to later restatements. The links are weak, and they do not support flagging a single company or executive.
Larcker and Zakolyukina labelled about 30,000 calls from 2003 to 2007 as truthful or deceptive, using later restatements, and built models from word categories linked to deception in earlier research.
Hobson, Mayew and Venkatachalam scored CEO speech for vocal dissonance on 1,572 calls and compared the scores with later restatements.
The "deceptive" label is a proxy. The authors say they cannot be certain the executives knew about the problems when they answered questions.
Why do general sentiment word lists misread financial text?
Because many words that sound negative are routine in financial reporting. Loughran and McDonald showed that general-purpose dictionaries flag such words as negative in financial text. "Liability" is the standard example. Finance needs its own word lists.
Their 2011 Journal of Finance study used 10-K annual filings from 1994 to 2008. They built finance-specific lists for negative and positive tone, uncertainty, litigiousness and modal words.
The current master dictionary is hosted by the Notre Dame SRAF project. It is free for academic research, and commercial use needs a licence from the authors.
A tone score built on a general dictionary will label routine financial language as negative. Check any score against the words that produced it.
Split the transcript into prepared remarks and the question-and-answer section, compare each with the company's previous call, and check both against the reported numbers. A tone score is a place to start reading, not a conclusion.
Start with the numbers: results against analyst expectations and against guidance. Then read the passages where the tone or the specifics change.
Compare the executives' answers with the questions they were asked. Note questions answered on a different topic, or not answered at all.
In EchoSignal, paste a section into the Analyse page, or two versions into Compare, and read the flagged sentences in context.
Mark any question that was not answered directly, and check whether the same question returns on the next call. Keep the reading to a short list of passages, so the follow-up stays manageable.
How can you tell if executives are being evasive on an earnings call?
Look for answers that do not address the question asked, and for passages that stay general where figures, dates or named items would be expected. Research on deceptive calls has linked some word patterns to restatements, but no single sign proves evasion.
These are reading prompts. They give you a reason to ask a follow-up question, not a finding about an individual executive.
Look for the specifics a question asked for, such as a figure, a date or a named customer, and note when they are missing.
Compare the answer with the executive's prepared remarks on the same topic. A clear gap between them is worth a follow-up question. Keep a note of the question and the answer, so the follow-up can be specific.
EchoSignal is a text analysis tool from Cave Fish Labs. It scores how a piece of writing is delivered, on trust, credibility, sensitivity and specificity. It is aimed at people who read statements and earnings calls closely. It is research software, and it does not detect lies.
The Analyse page scores any pasted text. The Market research applies the scores to earnings calls for the Russell 1000, FTSE 100 and FTSE 250.
Scores describe delivery. They are not measures of truth, and the questions on deception explain why.
It does not show who is right, and it does not replace reading the filings or the transcript. The Market page and the EchoSignal pages explain what each score measures.
Use it as one input alongside the filings, the transcript and the numbers.
Is EchoSignal accurate?
The scores are not forecasts. The Market page publishes the tests on 989 Russell 1000 companies, with data to 16 October 2026, so you can read the results yourself.
Each call is measured from the first trading day after it, against the stock's sector. The live model and an out-of-sample test are both published on the Market page.
The scores describe how a call was delivered. They are not a forecast.
Accuracy depends on what is measured. Each test on the Market page names its outcome, its sample and its method.
Read the sample size before the result, because small samples can swing a lot from one period to the next.
The Market page is updated as calls are added. A single test does not settle whether the scores are useful; a pattern across tests and periods is more informative.
Last reviewed: 2026-10-18. EchoSignal is research software. Nothing here is investment advice, and the scores are not evidence that anyone is lying. Privacy