Field report N-95 / L3 · sheet /macro

Boardroom MacroFiled from the territory

AI Startup Samaya Beat Wall Street Earnings Estimates 40 Times Out of 50

Samaya, which raised $43 million in 2025, out-forecast Wall Street consensus on 40 of 50 Q3 earnings calls. The test spanned ten sectors, from Coca-Cola to Carvana.

Grid
N-95 / L3
Surveyed
November 19, 2026
Reading
4 min read
Correspondent
Sophie Lindqvist
I tested an AI startup that claimed it could beat Wall Street earnings estimates. It won 40 out of 50 times
Fig. N-95 / L3I tested an AI startup that claimed it could beat Wall Street earnings estimates. It won 40 out of 50 times — AhmadArdity / pixabay

Field notes

  1. Samaya's AI forecasts beat the analyst consensus for 40 of 50 companies reporting Q3 earnings between October 26 and November 5.
  2. The startup raised $43 million from prominent investors in 2025.
  3. On Palantir, Samaya predicted $2.30 billion in revenue versus actual $2.31 billion; the Street said $2.18 billion.
  4. Its worst miss was Yum Brands: an EPS forecast of $1.42 against an actual $1.56.
  5. Samaya's clients include banks and hedge funds; the model trains only on public data such as filings and transcripts.

An AI startup beat the Wall Street analyst consensus on earnings forecasts for 40 out of 50 major companies in a live two-week test during the Q3 reporting season. Samaya, which raised $43 million from prominent investors in 2025, supplied estimates for companies ranging from Coca-Cola to Carvana before their earnings releases — and in most cases won by a considerable margin.

The experiment, run with Fortune after the magazine challenged the startup to prove its claim, compared Samaya's AI predictions against a Bloomberg-supplied consensus for 50 companies reporting between October 26 and November 5. The sample covered ten sectors, from health care to energy to IT. For each company, Samaya delivered an AI-based estimate for the key metric — EPS (earnings per share, the standard quarterly profit figure) or revenue — at least a day before the release.

What did the numbers show?

The early results set the tone. Samaya forecast Verizon revenue of $34.52 billion, closer to the actual $34.31 billion than the analysts' $34.77 billion. For UPS, its adjusted EPS prediction of $1.65 beat the Street's $1.62 against an actual figure of $1.67.

The analysts struck back on the next two names, S&P Global and PayPal. From then on, Samaya took 36 of the remaining 44 companies. On Palantir, it predicted $2.30 billion in revenue versus an actual $2.31 billion — the consensus said $2.18 billion. On Permian Resources, it hit the actual EPS of $0.69 exactly, while the Street forecast $0.625.

The losses are worth reading as closely as the wins. Samaya's most notable miss came on Yum Brands, where its $1.42 EPS forecast fell well short of the actual $1.56; the analysts came closer at $1.49. It also lost when forecasting how far Amazon and Google would fall short of revenue expectations.

Samaya CEO Maithra Raghu, a former member of the foundational Google Brain AI team, offered an explanation for the pattern: where the consensus won, it typically involved very large companies that are the subject of large volumes of private research. Most of that research was not part of Samaya's training material, she said — the startup relied on publicly available data such as filings, news reports and interview transcripts.

Raghu says Samaya, whose clients include banks and hedge funds, is likely to add analyst notes to its earnings training data in the future. But she flagged a risk: doing so could bias the AI by teaching it to internalize the same desire for consensus that defines many Wall Street analysts. The promise of better data comes with a caveat about what that data carries.

Why does the consensus behave the way it does?

"I think there is safety in numbers," said Sheraz Mian, Director of Research at Zacks, one of the firms that compiles consensus figures. Mian describes a process rooted equally in human nature and tacit corporate pressure.

An analyst may believe a company's quarterly earnings will fall well outside expectations, yet the incentives cut against saying so. The acclaim from a correct unorthodox forecast rarely outweighs the reputational damage of a wrong one. The unspoken incentive is to stick with the herd.

Investor relations teams reinforce the instinct. Companies monitor individual analysts closely and call them quickly if a forecast looks too optimistic — soft pressure that persists even after the SEC's 2000 Regulation Fair Disclosure banned sharing material information, including forthcoming earnings, with analysts. The asymmetry is stark: beats earn stocks a modest boost, while misses are punished disproportionately.

"In the long run, they provide optimistic numbers that get walked down to beatable targets at the end, so a higher proportion of companies end up meeting or beating expectations right at the close," says Paul Hribar, a professor of accounting at the University of Iowa.

Hribar notes that Wall Street has long answered this game with the "whisper number" — a figure closer to what people actually believe than to public predictions. Whether a consensus, or even a whisper version of it, survives the spread of AI forecasting is an open question.

What does this change for the earnings ritual?

The current regime is deeply rooted in human incentives and bias — qualities that insentient models do not share. Two outcomes are possible. AI models may in time internalize those same biases and reflect them in their output, a possibility Raghu explicitly acknowledged. Or AI-based forecasts could break up the cozy analyst-IR consensus altogether.

The caveats deserve equal billing. This was one experiment, 50 companies, a single earnings period. The test does not establish durability, and neither side has published error magnitudes across the full sample.

There is also a commercial question. If the model is this good, why is Raghu building a startup rather than trading? She says the question is "not uncommon" — but her plan for now is to keep working on Samaya's technology. Fortune's 1997 observation that the quarterly earnings number is "the simplest, most visible, most merciless measure of corporate success" still holds. What may be changing is who gets to predict it.

via fortune.com/feed/ (Source)

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Correspondent covering money, funding and market shifts around corporate strategy and chief-executive leadership at CEO Frontier.

483 articles

Filed under

Adjacent sheets

« Previous articleNext article »

End of sheetN-95 / L3 · November 19, 2026