We Asked an AI 500 Questions About 174,996 Earnings Calls. Three Answers Predict a Crash.
Here is a game. Take any earnings call. Do not look at the numbers, the guidance, the sector or the stock. Answer three yes-or-no questions about the words, and we will tell you, roughly, how often that stock fell twenty points behind the market over the following year. Zero yeses: about one in three. Three yeses: nearly two in three.
That is not a model with a thousand parameters. It is three questions we pulled out of five hundred, and the five hundred are the part of this story worth explaining first.
- We ask every earnings call the same 500 yes-or-no questions, 174,996 calls so far, 87 million answers. We ranked all 500 by how well a “yes” predicted the stock over the next twelve months, since 2021, requiring the sign to hold in every single year.
- The most dangerous words on a call are “we believe”. When management described the future with “we believe” more than “we expect”, 51% of those stocks fell 20 points or more behind the S&P 500 within a year, against 42% when “we expect” won.
- Tell two: “excited” more than “confident”. 49% crashed against 42%. Tell three: an analyst’s question left unanswered. 55% against 38%.
- Stack them. Zero tells: 35% crashed. One: 43%. Two: 54%. Three: 62%. The typical three-tell stock trailed the index by 37 points. It holds in every year since 2021, in every size of company, and in raw returns with no index at all.
- The boring words win. The CFO saying “free cash flow”, “record”, “strong”, an analyst asking about capex: all good. “Headwinds”, “synergies” and “thank you” predict nothing at all.
- Before 2021 the tells were half as strong, and in 2019 and 2020 they did not work. Five straight years is a regime, not a law, and we say so below.
Five hundred questions
Every stock page on this site carries the same 500 questions, answered yes or no by a language model that has read the transcript: does the operator announce more than ten questioners, does the CFO say “free cash flow”, does anyone describe the future with “we believe” more than “we expect”, is any question left unanswered. They were written to be checkable, not clever; most of them are things you could count with a highlighter if you had the afternoon.1
Which of the 500 matter? We had never checked. So we took the 116,738 calls with a full year of prices after them, split them by the answer to each question, and compared the two groups on three things: how often the stock beat the S&P 500 over the next twelve months, how often it fell twenty points or more behind, and the median. Then we did the cruel part: we kept only the questions where the gap had the same sign in every year from 2021 to 2025. A question that worked in three years and flipped in two does not make the list.2
Most questions do nothing. The ones that survive are not the ones we would have guessed.
The three tells
The single most consistent bad sign on an earnings call is a two-word phrase:
| Question the model answers on every call | Yes on | Crashed, yes | Crashed, no | Beat index, yes | Beat index, no |
|---|---|---|---|---|---|
| Does anyone describe the future with "We believe" more than "We expect"? | 33% | 51% | 42% | 30% | 35% |
| Does an executive use the phrase "we believe"? | 88% | 46% | 38% | 33% | 37% |
| Is the word "excited" used by the CEO more than the word "confident"? | 39% | 49% | 42% | 31% | 35% |
| Executive uses the word "excited" more than twice. | 31% | 50% | 43% | 30% | 35% |
| Does management use the phrase "we believe" at least three times? | 59% | 48% | 41% | 32% | 36% |
| Does any speaker say "operator" aloud during the call? | 57% | 48% | 41% | 32% | 35% |
| Does any executive say "we believe" at least twice? | 73% | 47% | 40% | 33% | 36% |
| Does anyone say "we're excited" or "we are excited"? | 51% | 48% | 42% | 32% | 36% |
| Does anyone say "we are excited about" without naming a number? | 48% | 48% | 42% | 32% | 35% |
| Does management mention a specific number of stores, locations, or markets? | 45% | 47% | 43% | 32% | 35% |
Read down the left column and a pattern jumps out. Belief instead of expectation. Excitement instead of confidence. “Excited about” with no number attached. “Leverage” as a verb. Praising the question instead of answering it. These are not the words of a management that knows what next quarter looks like. They are the words of one that hopes.
We took the two strongest and added a third from our call-level flags, whether any analyst question was left unanswered, and called them the three tells:
- “We believe” beats “we expect”. 51% of those stocks crashed within a year, against 42%.
- “Excited” beats “confident”. 49% against 42%.
- A question is left unanswered. 55% against 38%.
Here is what each one sounds like, from calls you will recognise. The returns are raw, twelve months from the call.
“We believe.” Carvana, November 2021: “Given our opportunity, we believe this is clearly the right choice.” One year later: −98%. Coinbase, the same week: “We believe that that’s a huge validation of the whole space.” −85%. Snap, July 2021: “We believe we are well positioned to drive business results for advertisers over the long-term.” −87%.
“Excited.” Nikola, February 2024: “I couldn’t be more excited to be a part of Nikola as I am now.” −99%. Upstart, November 2021: “We are really excited about what we’ve done, but even more excited about what’s in front of us.” −92%. Opendoor, the next day: −91%.
The unanswered question. Bed Bath & Beyond, April 2022: −99%. AMC, May 2023: −95%. Carvana again, February 2022: −94%.
Each is modest on its own. Together they are not:
A call with all three tells: 62% of those stocks fell twenty points or more behind the market within a year, and the median trailed it by 37. A call with none: 35%, and a median of −8.
About 8% of calls since 2021 carry all three tells. That is roughly one company in twelve, every quarter, telling you something in the way it talks.
| Company | Call | One year later |
|---|---|---|
| Carvana | November 2021 | −98% |
| Snap | July 2021 | −87% |
| Coinbase | November 2021 | −85% |
| Beyond Meat | November 2021 | −81% |
| Lucid | November 2021 | −80% |
| Twilio | July 2021 | −77% |
| Affirm | May 2021 | −74% |
| Robinhood | August 2024 | +547% |
That last row is there on purpose. Among the 317 three-tell calls from well-known companies in our data, 129 went up over the following year, and a few went up a lot: Robinhood, MicroStrategy, Oklo. Three tells is a base rate, not a verdict. It moves the odds of a crash from one in three to nearly two in three, and it does nothing to the odds of a rocket. If you own one of the rockets, you know it, and you also know what it sounds like on the call.
Does it survive?
Three checks, because a result this clean deserves suspicion.
Size. Excited, believing, evasive managements cluster in small companies, and small companies had a terrible run against a capitalisation-weighted index. So we split by the number of analysts on the call, a decent proxy for size. The gap holds in every bucket. Among the largest companies, ten or more analysts, “we believe” calls crashed 41% of the time against 36%; among the smallest, 57% against 49%.
The index. Drop the S&P altogether and use raw returns. Since 2021 the median “we believe” stock lost 11% over the following year and 40% of them went up; the median “we expect” stock lost 2% and 47% went up. Same story, no benchmark required.
Time. The crash rate for zero, one, two and three tells, by year:
| Year | 0 | 1 | 2 | 3 | Gap |
|---|---|---|---|---|---|
| 2017 | 26% | 34% | 35% | 45% | +19 |
| 2018 | 31% | 37% | 43% | 49% | +18 |
| 2019 | 51% | 51% | 51% | 51% | +1 |
| 2020 | 26% | 28% | 32% | 34% | +8 |
| 2021 | 25% | 39% | 55% | 70% | +45 |
| 2022 | 31% | 39% | 52% | 59% | +28 |
| 2023 | 42% | 49% | 58% | 64% | +22 |
| 2024 | 39% | 45% | 52% | 58% | +19 |
| 2025 | 36% | 41% | 48% | 53% | +18 |
Every year since 2021 the ladder runs the right way, with a gap of 17 to 45 points. In 2017 and 2018 it ran the right way at half the strength. In 2019 and 2020 it did not run at all: the tells were as likely to precede a winner as a loser. That is the honest shape of this finding. The words have carried information for five straight years, they carried some before, and there were two years when they carried none.
The market, for once, notices something on day one. The morning after a three-tell call, 43% of stocks went up; after a zero-tell call, 53%. Three months out the gap is already 10 points on the hit rate. A year out it is 14.
The words that work
The other end of the table is the mirror image, and it is almost funny how dull it is:
| Question the model answers on every call | Yes on | Crashed, yes | Crashed, no | Beat index, yes | Beat index, no |
|---|---|---|---|---|---|
| Does any analyst's question include the word "capex" or "cash flow"? | 30% | 40% | 47% | 38% | 32% |
| Does the CFO use the phrase "free cash flow" at least once? | 31% | 40% | 47% | 38% | 32% |
| Does management use the word "strong" at least three times? | 82% | 43% | 56% | 35% | 29% |
| Does the word "strong" appear more than five times? | 62% | 40% | 53% | 36% | 30% |
| Does the phrase "record" appear at least twice in the transcript? | 36% | 39% | 48% | 37% | 32% |
| Does the transcript contain the word "record" preceding a noun (e.g., "record margin")? | 42% | 39% | 49% | 37% | 31% |
| Does at least one analyst ask a question about margins or costs? | 82% | 43% | 53% | 35% | 29% |
| Does any question explicitly reference the company's prior guidance or outlook? | 79% | 43% | 52% | 35% | 30% |
| Does any analyst preface a question by summarizing the company's stated results? | 83% | 43% | 54% | 34% | 30% |
| Does the word "resilient" or "resilience" appear? | 31% | 39% | 48% | 37% | 32% |
Free cash flow. Record. Strong. Capex. Margins. An analyst who quotes the numbers back to management before asking. A management that says “base case” or “scenario”. The good calls sound like accountants and the bad calls sound like a launch event. The CFO saying “free cash flow” is the best single word on the list: since 2021, 27% of those stocks crashed in raw terms against 37% when nobody said it, and the median one went up while the median silent one lost 7%. Palantir’s CFO, August 2024, zero tells: “$149 million in adjusted free cash flow, representing a margin of 22%.” One year later: +575%. Super Micro, May 2023, zero tells, positive free cash flow of $190 million: +469%. Dell, June 2023: +179%. Nobody on those calls was excited about anything.
The words that mean nothing
| Question the model answers on every call | Yes on | Crashed, yes | Crashed, no | Beat index, yes | Beat index, no |
|---|---|---|---|---|---|
| Does the CEO speak first among named executives in the Q&A? | 73% | 45% | 44% | 34% | 34% |
| Does anyone use the phrase "the vast majority"? | 15% | 44% | 45% | 34% | 34% |
| Does anyone say "headwind" or "headwinds" more than once? | 28% | 42% | 46% | 34% | 34% |
| Does an analyst thank management at the start or end of their turn? | 82% | 45% | 45% | 34% | 34% |
| Does any question begin with the words "can you" or "could you"? | 58% | 45% | 46% | 34% | 34% |
| Does the word "synergies" appear? | 16% | 43% | 45% | 34% | 34% |
| Does management decline to answer by citing a future event ("we'll address that in..." or "at our upcoming...")? | 48% | 45% | 45% | 34% | 34% |
| Does the phrase "thank you" appear more than ten times? | 63% | 45% | 46% | 34% | 34% |
“Headwinds” is the word every financial newsletter tells you to fear. In our data it predicts nothing: 34% of headwind calls beat the index, and 34% of the others did. “Synergies”, nothing. Whether the CEO speaks first, nothing. Whether anyone says “thank you” more than ten times, nothing. The cliches you have been taught to listen for are noise, and the tells are phrases nobody told you to notice.
Two things that fell out on the way
We will write both of these up properly, but they belong here in one line each. First, the model’s own grade, which is built from all 500 answers: since 2021, 66% of the calls it graded F crashed within a year and 22% beat the index; for the A calls, 28% and 43%. Second, founder-led companies. From 2017 to 2020 they beat the field, 44% ahead of the index against 39%. Since 2021 they have been the worst large group in the data: 30% ahead against 35%, a median of −24 against −13. A founder-led company with all three tells crashed 66% of the time.
It is not all clean
The regime is the big caveat and we have already shown it: two of the last nine years, the tells did nothing. A language model answered the 500 questions, and a model can miscount; we spot-check, but 87 million answers are not hand-verified. The price histories carry some survivorship bias, which flatters every group about equally. Correlation is doing the work here, not causation: a management that hopes instead of expects is usually a management in a business that is going badly, and the business, not the grammar, sinks the stock. The words are a symptom you can read a year early, which is exactly why they are useful.
What we are watching
The 2026 cohort is the largest yet and its outcomes arrive through 2027. We will publish the ladder for it whichever way it runs. In the meantime every stock page shows all 500 answers for every call, so you can count the tells on any company in about ten seconds, and the rankings update after every call.
- The 500 questions were written in rounds and kept for stability (the model gives the same answer on re-runs) and information (the answer is not the same for nearly every call). They are answered from the transcript alone; the model does not see prices. Every answer is public on the company’s stock page. ↩
- Outcome window: first close after the call to the same date twelve months later, minus the S&P 500 (SPY) over the same window. Calls under 2,500 words are excluded. Universe: 116,738 calls from 2016 to 2025 with a full year of prices, of which 61,663 are from 2021 onward. Crash: excess return of −20% or worse. A question qualified for the ranking if both the yes and no groups had at least 1,500 calls since 2021 and the yes-minus-no gap in hit rate had the same sign in every year from 2021 to 2025 with at least 200 calls a side. ↩