By Evan Pondel, IRC
If you log in to Kalshi on any given day and search for “earnings” on the prediction markets platform, more than four dozen posts will appear instantly, allowing users to bet on what companies will say on their next earnings call.
From General Motors to Netflix, prediction market enthusiasts can place bets for pennies on the dollar that an executive will say the word “domestic” or “Kpop Demon Hunters” on an upcoming call. While insider trading laws apply on these platforms, there are no rules governing how these investors engage investor relations (IR) professionals.
That raises a question every IR professional may eventually face: should anyone in the IR department be watching these prices? A new working paper from Duke University offers some of the first hard evidence, and its answer is a qualified “no.” The bets are not pure noise, the researchers find, but they tell IR teams very little that their own disclosures don’t already reveal.
A Coin Flip It Is Not
Duke University researchers William J. Mayew and Boyan Zhu studied more than 1,400 Kalshi bets on earnings call language, covering more than 100 calls at 72 companies during the prediction market’s first nine months of existence. The researchers are in the process of submitting a revised paper that more than triples the size of the study with effectively similar results.
The companies are household names, closely followed by Wall Street analysts. The most frequently listed contracts centered on hot-button words such as “tariff,” “shutdown,” and “China.”
The researchers staged a simple contest. On one side were the Kalshi odds just before each call, typically minutes before it began and, in nearly every case, after the earnings release was out. On the other was material IR teams already produce, including what the company said on recent calls, in press releases, and in SEC filings, plus what competitors had been saying.
It was essentially a tie. Both did a strong job of predicting which words executives would say, and neither clearly beat the other.
Combining the two helped, but only a little. Prices do carry some information beyond disclosures, the authors conclude, but the gain is “economically modest.” Their bottom line: “There is little novel information in prediction market prices beyond what investor relations officers (IROs) already provide to meet information demand through disclosures.”
The strongest predictors on the disclosure side were how often a word appeared in past call presentations and in press releases. Words buried in 10-K and 10-Q forms added nothing.
Signal or Slot Machine?
Mayew and Zhu frame the debate as two competing views. In the “information in price” view, retail traders dig up issues management has missed, and rising prices flag topics investors want addressed. In the “gambling as a leisure activity” view, trades are entertainment, and prices carry little worth acting on.
The market favors the second view. Many contracts go days without a single trade, and the average contract trades only about $9,400 over its life. The typical contract opens about four weeks before the call, deep inside the quiet periods large companies observe, when fresh private information is hardest to obtain.
Yet the evidence doesn’t fit pure gambling either. Prices predicted outcomes well in the scripted presentation, where most of the words were said. In the Q&A, they did little better than a coin flip.
That pattern, the authors argue, is consistent with managers “responding to prediction market prices as if they contain an additional, albeit economically small, signal of stakeholder information demand.” The presentation is where executives proactively decide what to say, so that is where a price signal would show up.
The researchers are careful about what they can’t prove. They cannot see management actions directly, so they cannot fully rule out tipping, in which insiders leak planned wording to traders. Companies with disclosed quiet periods showed no weaker price effect, a test the authors concede cannot rule tipping in or out.
Practitioners are split too. One author asked some NIRI members how they would adjust disclosures based on Kalshi prices. One member said the prices should be ignored since Kalshi traders aren’t interested in fundamental value. Another warned against chasing “gamblers and speculators rather than the investors you want to attract,” while allowing that high-volume contracts on relevant topics might serve as surveillance. A third argued that IR has entered the prediction markets era, and the money flowing in makes prices matter.
The Machines Are Listening, Too
“Mention markets” are the newest audience parsing executives’ words, but not the first. For years, quantitative funds and analysts have run earnings call transcripts through software that scores tone and flags risk.
Those tools have grown sharper. A 2025 study by S&P Global Market Intelligence found that large language models can read context that old word lists miss, recognizing that a phrase such as “growth slowed less than expected” is good news. A long-short strategy built on large language model signals returned 8.4% a year, double the 4.2% of a traditional dictionary approach, according to Fortune.
The tools reach beyond the transcript. A student team in the CFA Institute’s 2026 Artificial Intelligence (AI) Investment Challenge built EarningsIQ, which pairs a finance-tuned language model with audio analysis of executives’ vocal delivery, according to the CFA Institute.
Executives have noticed. Researchers at Georgia State University and Columbia University found that companies where filings are heavily downloaded by machines began avoiding words algorithms read as negative, Bloomberg reported. The same research, published in the Review of Financial Studies, found executives even sound more upbeat on calls when their investors use AI. “There’s a change in tone and sentiment,” co-author Alan Zhang of Florida International University says.
Some IR teams turned the tools on themselves early. Sun Life Financial’s IR group ran its prepared remarks, and peers’ transcripts, through a tone analyzer after hearing how BlackRock used its Aladdin platform to gauge call sentiment, IR Impact reported. The goal was to brief executives on words to use or avoid.
AI is now converging with the mention markets themselves. Firms such as Octagon AI publish model probabilities alongside live Kalshi prices, including for Amazon’s earnings mention markets. Academic teams are testing how to feed prior calls and news into language models to forecast which words will be spoken, research Mayew and Zhu report.
If machines get good at predicting what a CEO will say, mention prices will increasingly reflect the company’s own public record. That is consistent with the Duke finding. Much of what the market “knows,” the IR department published first.
Winners, Losers and the Operator
Even if prices carried more signal, acting on them is fraught with risk. Every mention contract settles with one winner and one loser. In a market dominated by retail traders, the authors warn, responding to prices could look like management picking sides and alienate part of the retail base.
The risk is not hypothetical. On Coinbase’s third-quarter 2025 call, CEO Brian Armstrong closed by reciting “Bitcoin, Ethereum, blockchain, staking, and Web3,” words that were the subject of an $84,000 betting market across Kalshi and Polymarket, Bloomberg Law reported. The stunt showed how easily the people being bet on can settle the bet.
Control over outcomes is also murkier than it looks. CEOs triggered most of the winning “yes” bets in the Duke sample, followed by CFOs. But two contracts resolved on words spoken by conference call operators, who usually aren’t company employees. In Q&A, an analyst’s question can force an executive to say a word management deliberately left out of the script. However, that was rare in the sample.
Then there is insider trading. Executives sometimes talk privately with analysts and call operators before the call, opening several channels for leaks. The authors suggest companies consider updating quiet-period policies to address prediction markets explicitly. The stakes are rising as the Commodity Futures Trading Commission, which regulates Kalshi, sought public comment this year on the costs and benefits of these markets.
A Cheap Signal With Expensive Strings
For IR professionals, the takeaway is a trade-off. Kalshi prices are public, continuously updated, and free, making them an appealing gauge of retail interest compared with paid platforms for collecting retail questions. However, the Duke data says they mostly echo what a company has already said.
The authors offer one blunt option for firms that decide the costs outweigh the benefits, which is to pre-commit to saying the contracted words on the call. A certain outcome, they note, would collapse the market.
Mayew and Zhu caution that their sample covers early contracts on one platform and some of the largest companies in the economy. Whether the findings hold for smaller firms, or on Polymarket, remains an open question.
For now, the best predictor of what a CEO will say next quarter is still what the IR department wrote last quarter.
About the Author
Evan Pondel, IRC, is founder of E72 Communications and a member of the NIRI Learning and Publications Advisory Committee; email epondel@e72comms.com.