One of the small costs of running money is that you stop having time to read the people you most want to read. For me that is Matt Levine, whose Money Stuff column runs on Bloomberg, and who has spent a decade explaining financial plumbing better than anyone alive while making it funnier than it has any right to be. I have a folder of his columns I mean to get to. It is not getting smaller.

So this borrows his machinery. A simple model, some actual rules, several worse versions of the model, and the hope that the rules do the joke for you. The idea is to discuss things in a way that does not come across as accusatory, but still allows for the expression that something may be off. We cannot think of any better way to do this than in the style of the maestro.

Unfortunately, the sentence rhythm and the footnotes can be copied. What cannot be copied is the twenty years of structured-products law sitting under every casual aside, or the instinct for which absurdity is the load-bearing one.

Programming note: This is longer than usual. There is a lot of securities law in the middle. It is also focused on US market laws, even though the whisper-impact spread—and the corresponding potential for mischief—is now global. Finally, because we do not have access to Bloomberg’s proprietary fonts or content management system, we cannot perfectly mimic the visual experience of reading Money Stuff. You will just have to imagine the perfectly spaced block quotes and the Arial Black section headers doing the comedic work.

Everything is expectations

From first principles, you might imagine that the way to value a public company is to look at how much money it makes. If you run a company that makes boxes, and you sell $100 worth of boxes, that is good. If you sell $103 worth of boxes, that is better, and your box company shares are worth more money.

Obviously, this is not how it works.

If you are a public company, your value is not determined by your profits. Your value is determined by the derivative of your profits against the expectations of your profits. In the simplest possible model, the math works like this:

1. Analysts write down that you will make $100.
2. You report that you made $103.
3. Three is a positive number, so your stock goes up.

That model had a good run. It was very popular in the 1990s. But the modern version is a bit weirder. Today, you report $103 against a published consensus of $100, and your stock immediately falls 9% in four minutes. A guy on financial television says you had a "massive beat." A guy on a trading desk says the buy-side expectation was actually $105. Both of these people are telling the exact truth. You beat the number on the Bloomberg terminal, and you missed the number inside the hedge fund manager's head.

One of the core themes of this column is that everything is securities fraud. But the earnings-season variant is slightly less ambitious: everything is expectations. You beat earnings, but you missed expectations. You beat consensus, but you missed the whisper number. You beat the whisper number, but you disappointed on forward guidance. There is always another number.

And the problem with the number that actually moves the stock is that it has no author, no timestamp, and no methodology.¹

The official number

Mechanically, the "consensus" estimate is a manufactured product. It sounds like a grand, philosophical agreement among brilliant minds, but actually it is just a data vendor scraping thirty different Excel spreadsheets, applying some proprietary inclusion rules about what the word "adjusted" means, and dividing by thirty.

Because it is a manufactured product, it has manufacturing defects. Here is an academic paper by Larocque, Watkins, and Weisbrod:

We compare the five largest forecast data providers—Bloomberg, Capital IQ, FactSet, I/B/E/S, and Zacks—across a common sample of 94,030 firm-quarters, and find substantial differences in both the forecast and the reported street number, and therefore in the surprise, for the same company in the same quarter. There is a set of announcements where the providers disagree about whether the company beat or missed at all.

Right.

The company reported one objective reality. The market was supplied with five different historical baselines to compare it against.

To be fair, this does not make consensus useless. It is a flawed metric, but at least the factory keeps its receipts. If you want to know why the consensus was $100, you can look up the exact analysts who submitted their math. There is an audit trail.

Which sets up a fun irony. The "whisper number" arrives as a correction to a benchmark that is stale, thin, or mechanically averaged. And sometimes it genuinely is a correction. It just arrives with absolutely zero documentation, to fix a number that has five competing vendors and 90,000 firm-quarters of academic scrutiny.

You can't just do that

There is a standard origin story for how whisper numbers are created. It goes like this:

An equity research analyst builds a massive financial model in October and publishes an estimate of $100. Over the next three months, she learns things. The economy slows down, a supply chain breaks, a competitor goes bankrupt. She does not want to republish a 40-page PDF report every time she learns a new fact, because publishing is tedious and compliance departments are annoying. So her "live" number drifts from her printed number.

Then, the analyst has a conversation: 

Hedge Fund Client: "Your published estimate is $100. What is your actual estimate?" 

Analyst: "Well, honestly, it's $105." 

Client: "Great, I will trade on $105."

This is a very elegant solution to the problem of stale data. It is also wildly illegal.²

In the year 2000, the SEC passed Regulation FD, which made it illegal for companies to selectively tell Wall Street analysts things they weren't telling the public. Then came the Global Research Analyst Settlement of 2003, and Regulation AC, and FINRA Rule 2241. Conceptually, the SEC looked at the elegant solution of whispering the "real" number to favored clients and said, "You cannot do that." If an analyst's published number says $100, and she tells a client she actually believes it is $105, she is certifying a false document.

So the regulated analysts stopped doing it. But the regulatory trap is even funnier from the company's perspective. Imagine you are the Investor Relations officer at the company that makes the boxes. You have read the analyst reports. You know the analyst consensus is $105. You also have the actual financial results sitting on your desk, and you know you are only going to report $103.

Company IR: "All analysts think we are going to make $105. We are only going to make $103. The stock is going to crash."

Lawyer: "You cannot tell anyone that."

Company IR: "But I can see them writing on a mathematically false premise! Can I just call our most allowed analysts and gently walk them down?"

Lawyer: "Absolutely not. That is selective disclosure of material non-public information. You must sit quietly in your office and let them be wrong."

The need for the number did not go away. And risk cannot be destroyed, only moved.

Who is allowed to have feelings?

Because the people in charge of the official numbers are throttled by federal law, the market had to find someone else to invent the numbers. This creates a very strange matrix of legality and market impact.

There is a ladder of people who can say a number out loud before a company reports. The rules get stricter as you climb the ladder, but the relevance gets larger as you descend it.

At the very top is the sell-side equity research analyst. She has a compliance officer, FINRA rules, and the SEC breathing down her neck. If she wants to change her estimate from $1.00 to $1.05, she cannot just say she has a "good feeling" about the quarter. She has to update a massive Excel model, write a rationale, submit it to supervisory review, and publish it to all clients simultaneously. Because she is required to be perfectly diligent, objective, and transparent, she is incredibly slow. Her number is perfectly safe, totally public, and completely stale.

In the middle is a financial journalist or a sales trader. They are constrained, but differently. A journalist cannot invent a number, but she can write: "Traders say the buy-side expectation is closer to $105." She never has to explain which traders, or if they have done any math, or if "traders" is just one guy she plays tennis with. A sales trader can ping a client and say, "The desk feels like $105," as long as he keeps it brief and doesn't cross the line into publishing unregulated research.

At the very bottom is the buy-side analyst, the hedge fund PM, or the crowdsourced internet platform. They have absolutely no obligation to show their work. If a hedge fund manager notices a lot of foot traffic at the mall and decides the number is $105, he can just type "$105" into a Bloomberg chat. He does not need a compliance review to have a feeling. He can change his mind three times before lunch.

This is the great irony of regulating financial estimates. The person carrying the heaviest legal obligations produces the number that no longer moves the stock, because she is not legally allowed to just guess. The people carrying no obligations to explain their math produce the number that does.

Regulation ensures that the only number printed with absolute, rigorous transparency is the one nobody cares about. The market is priced by the people who have the liberty to just trade on vibes.

The inflation of expectations

In 1999, a whisper number meant one very specific thing: it was the quarterly Earnings Per Share for a massive tech company. You could write the names of the people who cared about it on a napkin.

Today, there is a whisper number for everything. There is a whisper number for gross margin. There is a whisper number for capital expenditure, which isn't even a forecast. It is a discretionary decision the board makes, but the buy-side has decided to treat it as a searchable metric. There are whisper numbers for cloud revenue run-rate, for subscription take-rates, for daily active users.

But my absolute favorite is guidance.

In forty minutes, a company is going to announce its own internal forecast for the next quarter. Wall Street analysts have published estimates of what they think the company's forecast will be. And there is now a whisper number for what the forecast of the forecast needs to be.

We have arrived at an unpublished expectation of a published expectation of a future expectation. And your stock can fall 9% because the third one missed the second one while beating the first.

I don't get it! But the supply of these numbers grew because the market moves got bigger, and the market moves got bigger because algorithmic trading systems need a binary trigger to decide whether to buy or sell. The algorithm just wants to trade the stock; it doesn't care if the number it is trading against was hallucinated by a subreddit.

Things happen.

Six different systems

If you say "the consensus is $100," everyone knows what that means. It means a data provider averaged the spreadsheets of thirty identifiable analysts. But if you say "the whisper is $105," you could be describing at least six entirely different production systems:

1. A desk survey. A broker emailed thirty hedge funds, asked for their guesses, and took the median.
2. An options-implied move. The options market is pricing in an 8% move, which tells you exactly how volatile the stock is and absolutely nothing about whether revenue will be $103 or $105.
3. A reverse valuation. A quantitative model says the stock needs 20% growth to justify its current multiple.
4. Alternative data. A satellite company counted cars in Walmart parking lots and produced a highly accurate forecast without talking to a single human being. These things are done, although I do not know how such a number could become the market’s average expectation.
5. One guy. A really smart buy-side analyst who happens to be right a lot.
6. Vibes. "The bar," which is just a measure of market crowding. If a stock is neglected, $103 is a massive beat. If everyone on Twitter owns it, $103 is a catastrophe.

All of these are reported under one label. Ahead of Nvidia's Q3 print, JPMorgan's trading desk circulated numbers that CNBC dutifully published:

Survey says $56.32b & $63.02b

That is a genuine artifact. It was compiled by identifiable people using a stated method (a survey). But here is a professional financial publication writing about a different company on the open internet:

...making the "whisper number" more like $1.60.

Their methodology was: Our target is $1.51, last quarter beat by 6.6%, therefore, the number is $1.60. That is just consensus multiplied by the trailing surprise rate, wearing a name that implies somebody knows a secret. Both of these are "the whisper." There is no Bureau of Whispers keeping the categories apart.

Four clients

Suppose you are an institutional equity salesperson. Published consensus is $100. You ring your four best clients to find out what the buy-side expectation is.

• Client One is a massive long-only fund. They send you a detailed spreadsheet bridging unit economics to revenue. They expect $106.
• Client Two is a hedge fund. They tell you the channel data is seasonally distorted, so they expect $102.
• Client Three is another hedge fund. They expect $105, mostly because they overheard Client One say "$106 is conservative" at an industry dinner.
• Client Four has no math at all, but tells you, "Man, this thing just needs to beat and raise."

What is the whisper number? If you average the three actual numbers, you get 104.3. If you strip out the double-counting of Client One's dinner conversation, you get 104. If you translate "beat and raise" into math, you probably get 105.

So you write this in your morning desk note:

“Our conversations suggest buy-side expectations are around $105.”

This is the most useful sentence anyone will write about the quarter. You know things no mechanical consensus can capture: who has conviction, how people are positioned, and what lines actually matter.

Look at what that sentence does, mechanically. "Our conversations" implies a methodology without forcing you to describe one. "Buy-side" turns three phone calls and a mood into a unified global constituency. "Around" absorbs the math errors. By the end of the sentence, $105 belongs to everybody, even though a quarter of your data was just an echo.

Then the laundering begins. An equity analyst cannot publish an uncertified view, so she stays quiet. But a trading desk is legally allowed to summarize "market color." The desk blasts the $105 out. A financial journalist sees the desk note and writes, "Press reports suggest buy-side expectations are running above consensus." Another broker cites the press report.

Repetition looks exactly like corroboration if you hide the plumbing.

Just make it up

If you are short a stock, and you think the company will report $103, you would really prefer that $103 is viewed as a disaster.

The legal way to achieve this is to genuinely expect $108, show your math, and convince four brokers you are right. The gray-area way is to tell people, "Everyone I speak to needs $108," where "everyone" consists of you, your junior analyst, and a guy who said "sure" while getting into an elevator.

The highly illegal version is that you expect $103, you actively invent a story that serious investors need $110, and you blast it out to journalists and algorithms because you want the stock to tank.

This is not a hypothetical concept. In November 2007, a trader named Paul Berliner sent instant messages to 31 traders claiming Alliance Data's board was meeting to cut a takeover price. The rumor was a complete fiction. The media picked it up, the stock fell 17% in thirty minutes, and Berliner covered his short. The SEC caught him, fined him heavily, and banned him from the industry.

He got caught because he invented a fact about the physical world. (This column does not give legal or investment advice, but you absolutely should not fabricate things that are demonstrably false.)

But a fabricated whisper number is a fact about opinion. "The buy-side is at $110" has no board meeting behind it. The company cannot deny it, because the company has no idea what the buy-side thinks. You cannot subpoena a vibe.

And it gets easier the further down the income statement you go. A fabricated Earnings Per Share whisper is falsifiable within a quarter. A fabricated gross margin whisper is falsifiable never. Somebody is circulating a buy-side whisper for a mid-cap tech company's gross margin, expressed to one decimal place, because a decimal place makes a hallucination sound like an Excel model.³

Time travel

It is very tempting to say the machines "knew the whisper." They did not. There is no secret Bloomberg terminal feed piping the authentic thoughts of the buy-side into the algorithms. The machine just obeys its benchmark. It has a file that says EXPECTATION = 105. It sees 103, and it sells. Any credulity happened three days earlier, when a human decided $105 deserved to be typed into the system.

This is my absolute favorite way a whisper number is generated. "Expectations were higher" is a perfect financial theory, because it cannot be falsified by the press release, and it explains every possible outcome. Stock falls on good results? Expectations were higher. Stock rises on bad results? Expectations were lower. The theory achieves perfect historical accuracy by waiting for history to happen.

The circularity is beautiful: The shares fell, so investors were disappointed. The company beat consensus, so investors must have wanted more than consensus. Therefore, the whisper was $105. We know the whisper was $105 because the shares fell.

By the next morning, there is a whisper number in print that literally did not exist before the earnings print. It was manufactured in good faith by four people trying to be helpful about something none of them understood, on a deadline. Next quarter it becomes history, and the quarter after that, it becomes a data series.

Receipts

None of this means published consensus is better. Buy-side analysts update faster, alt-data is smarter, and the price action actually does carry information. The modest argument is simply that a number capable of moving billions of dollars of market capitalization should probably arrive with a receipt.

Three sentences would do all the work:

1. Desk survey, twelve independent responses, median 105, range 101 to 109.
2. House estimate of 105 based on credit card data. This is not a client survey.
3. We infer the stock requires 105 based on options pricing. We did not speak to anyone.

Notice that the third one, the honest one, is the sentence nobody ever writes.

A well-run bank actually does have these receipts. They have call notes, timestamps, and compliance sign-offs. That doesn't prove the estimate was intelligent—being stupid is mercifully outside the jurisdiction of the securities laws—but it proves the $105 actually existed on Tuesday, rather than being invented on Thursday to explain why the stock went down.

Nobody has asked for these receipts yet. There has never been a massive SEC inquiry into the manufacture of a company-specific whisper number. That tells you either that the system is entirely clean, or that nobody has turned on the lights. From the outside, those two states look exactly the same.

But the price moves keep getting bigger. A lot of people are losing a lot of money in the first sixty seconds of after-hours trading against a benchmark they were never allowed to see. That is a fairly reliable way to eventually acquire a regulator's attention.

When the SEC finally does look, the complaint will probably not use the word "whisper." It will ask who originated a number, what they claimed about its source, what they owned when they claimed it, and who traded after hearing it. The answers are already sitting on a server somewhere, waiting for a question nobody has had a reason to ask.

Things happen

Marvell selloff deepens as investors seek clarity on Google AI deal payoff. Applied Materials forecasts revenue above estimates; shares fall on high expectations. Rubrik shares drop on a raised outlook. Samsung Electronics shares skid as record shareholder returns disappoint. Walmart raised its full-year outlook and had its worst day since 2022. Synopsys shares dip despite an earnings beat and a raised outlook. Fabrinet shares slide despite an earnings beat as margins and cash flow disappoint. Gap taps an industry veteran to revive Old Navy; shares jump. Bank of America on the Marvell selloff: “We ignore this expectation mismatch”.

‍¹ To be fair to Wikipedia, it does not claim the practice is current. It just uses the present tense, which is the same thing.

² The compliant way to distribute a number that might be research is to send it to fourteen people. This is a distribution-list arithmetic problem, and it tells you a great deal about how the rule was drafted and rather less about what it protects.

³ A model can derive 105 automatically from historical beat rates and revisions, which makes the provenance perfectly reproducible without making 105 anybody's expectation. Reproducible and true are different properties, and only one of them is cheap.