Arithmetic audit of 44 companies already analysed. Forecasts and figures as of September 9, 2026. This does not measure whether the analyst will be right: it measures whether any combination of revenue and margin exists that produces the profit being promised.
The question, and why it is not prediction
Guessing earnings better does not make money. We measured it on a sample frozen in January 2025 and checked nineteen months later: analysts got the earnings roughly right, and yet getting the number exactly right had almost no relationship with what the share price did. One company delivered eight times what was asked of it and fell 54%.
The whole calculation is this:
| Step | What it is |
|---|---|
| Profit required | what it promises to earn per share × the shares it will have |
| ÷ 0.79 | because tax has to be paid: a uniform 21% for every company |
| ÷ its best-ever margin | the highest that company has ever reached |
| = the revenue it demands | compared against what it bills today |
The result falls into three bands:
| Band | Threshold |
|---|---|
| Comfortable | needs less than 40% more revenue |
| Stretched | between 40% and 150% more |
| Impossible | more than 150% more, at the best margin in its history |
Four bugs found and fixed before trusting the result
The first time this test ran, five of six companies came out impossible. When a filter flags almost everything, you audit the filter, not the market. It was broken, for four separate reasons.
1. Two different sets of accounts were being compared
A company publishes two profits: the one in its audited accounts, and the one it calculates itself after stripping out costs it considers exceptional. The analyst forecast sits on the second. The historical margin was coming out of the first. Dividing one by the other means nothing.
| Real example, same quarter and same company | |
|---|---|
| Analyst estimate | $1.93 |
| Estimate on the official accounts | $1.36 |
| the second is 70.5% of the first | |
The signal that gave it away: the engine calculated that Broadcom needed a 106% operating margin. No company can keep more money than it bills. That was not an absurd forecast: it was a badly framed division.
The fix: build a comparable margin from the accounts themselves — operating profit plus share-based payments to employees plus the wear on what has been bought — and set the ceiling there.
2. Share-based pay is published cumulatively, not quarter by quarter
Rebuilding it as standalone quarters loses three out of every four.
3. The first tag with data was taken, not the most recent
4. The fourth quarter was rebuilt by calendar year
The fourth quarter is not published on its own: it has to be subtracted from the full year. Delimiting it by the calendar fails for every company whose fiscal year does not end in December.
How we know it is right now: Dell comes out at 151,197 million of revenue and 14,252 of operating profit, exactly what the daily market data system gives by a completely different route. Workday also matches to the million.
The result across the 44
| Band | How many | Meaning |
|---|---|---|
| Comfortable | 18 | the forecast works with the margin they already have |
| Stretched | 7 | demands between 40% and 138% more revenue |
| Impossible | 1 | not even at the best margin in its history |
| Not judgeable | 7 | the bottom-line profit does not come from operating |
| No data | 5 | foreign issuer, short series, or no revenue yet |
| Not applicable | 6 | banks and financials |
The ones that demand the most
Ordered by what the forecast demands. «Its margin» is the one it has now on the comparable measure; «its record», the highest it has ever reached.
| Company | Bills today (M$) | Its margin | Its record | Promises to earn | Must bill (M$) | Growth | |
|---|---|---|---|---|---|---|---|
| AMD | Advanced Micro Devices | 41,305 | 25.8% | 27.0% | $15.73 | 122,432 | +196% |
| NVDA | Nvidia | 302,969 | 67.8% | 67.8% | $15.70 | 721,610 | +138% |
| AVGO | Broadcom | 75,465 | 65.7% | 69.2% | $19.28 | 171,304 | +127% |
| AMAT | Applied Materials | 30,837 | 32.1% | 33.6% | $18.67 | 56,966 | +85% |
| ALNY | Alnylam | 4,411 | 23.8% | 23.8% | $10.08 | 7,300 | +66% |
| AGYS | Agilysys | 330 | 23.0% | 23.0% | $3.20 | 506 | +53% |
| AGX | Argan | 1,188 | 16.1% | 16.1% | $15.88 | 1,783 | +50% |
| KRYS | Krystal Biotech | 440 | 57.3% | 57.3% | $9.47 | 632 | +43% |
| ADSK | Autodesk | 7,790 | 35.9% | 35.9% | $14.18 | 10,653 | +37% |
| ORCL | Oracle | 67,357 | 40.2% | 45.7% | $10.99 | 89,408 | +33% |
| INOD | Innodata | 252 | 21.4% | 22.0% | $1.63 | 329 | +31% |
| CAT | Caterpillar | 74,729 | 18.1% | 21.1% | $32.42 | 91,013 | +22% |
| WDAY | Workday | 10,155 | 26.9% | 27.2% | $13.23 | 12,315 | +21% |
The comfortable end
A negative growth figure means it already bills more than would be needed.
| Company | Bills today (M$) | Its margin | Its record | Promises to earn | Must bill (M$) | Growth | |
|---|---|---|---|---|---|---|---|
| DXCM | DexCom | 4,969 | 26.3% | 26.3% | $3.12 | 5,786 | +16% |
| ADMA | ADMA Biologics | 512 | 48.3% | 48.3% | $0.94 | 566 | +10% |
| BKNG | Booking | 28,241 | 35.6% | 40.2% | $12.38 | 30,159 | +7% |
| DASH | DoorDash | 15,891 | 13.9% | 14.8% | $4.70 | 16,776 | +6% |
| ACMR | ACM Research | 1,038 | 15.8% | 25.6% | $3.23 | 1,051 | +1% |
| NXPI | NXP Semiconductors | 12,615 | 36.1% | 49.2% | $18.18 | 12,033 | -5% |
| ROP | Roper | 8,280 | 40.9% | 42.2% | $24.33 | 7,365 | -11% |
| DELL | Dell | 151,197 | 10.2% | 10.2% | $30.91 | 124,355 | -18% |
| RDDT | 2,779 | 40.7% | 58.9% | $7.06 | 2,256 | -19% | |
| KNSA | Kiniksa | 754 | 17.3% | 22.5% | $2.05 | 568 | -25% |
| DXPE | DXP Enterprises | 2,139 | 10.1% | 10.1% | $7.66 | 1,514 | -29% |
| CARG | CarGurus | 967 | 29.8% | 47.6% | $3.02 | 620 | -36% |
| TSLA | Tesla | 103,619 | 7.9% | 18.9% | $2.27 | 61,389 | -41% |
The only impossible one
Advanced Micro Devices. To earn the $15.73 being asked of it, it needs 26,316 million of profit, and for that, revenue of 122,432 million. Today it bills 41,305.
Every margin was tested, including ones it has never touched:
| If its margin were… | It would have to bill | Growth |
|---|---|---|
| Today's, 25.8% | 129,266 M$ | +213% |
| Its historical record, 27.0% | 123,194 M$ | +198% |
| 30% (never seen) | 111,039 M$ | +169% |
| 35% (never seen) | 95,176 M$ | +130% |
| 40% (thirteen points above its record) | 83,279 M$ | +102% |
The most demanding after that
The four at the top of the list are the four AI chip names. That is not a coincidence: it is the only place where the forecast asks for two consecutive years like the last one.
And two cases that look identical are worth separating:
| What is asked of it | What it has just done | |
|---|---|---|
| Nvidia | grow 138% | grew 131% last year — it is asked to repeat |
| AMD | grow 196% | grew 50% — it is asked to quadruple the pace |
What the other extreme says
To earn the $2.27 being asked of Tesla, with the margin it had in 2023, it would be enough to bill 41% less than it bills today. That forecast is not merely possible: it has room to spare.
And one that changed when the fourth quarter was corrected: Oracle moved from stretched to comfortable, at +33%. It is one of the few large caps where arithmetic is not the problem.
The seven that cannot be judged
It does not mean their forecast is impossible: it means the test does not apply. When a company's bottom-line profit looks nothing like what it earns from operating, it is built on something that is not the business, and it cannot be projected from it.
| Company | Earns from operating | Bottom-line profit | ||
|---|---|---|---|---|
| SNOW | Snowflake | loses money operating | — | |
| CORT | Corcept | 6 M$ | 55 M$ | 8.6× |
| WDC | Western Digital | 4,453 M$ | 9,424 M$ | 2.1× |
| MRVL | Marvell | 1,561 M$ | 2,640 M$ | 1.7× |
| GOOGL | Alphabet | 147,628 M$ | 244,205 M$ | 1.7× |
| AMZN | Amazon | 93,712 M$ | 135,281 M$ | 1.4× |
| ALAB | Astera Labs | 273 M$ | 369 M$ | 1.4× |
A bottom-line profit at double what is earned from operating comes from a tax refund, from a disposal, or from a revaluation of investments. None of those three repeats next year. That is a warning in itself, independent of any forecast.
And what falls outside the test's reach:
| Companies | Reason |
|---|---|
| ASML · NESR | foreign issuers: they do not publish US quarterly accounts |
| CDNL · FCFS | series too short to rebuild twelve months |
| RVMD | no revenue yet |
| ATLC · BOW · BWB · ENVA · ESQ · MCB | financials: in a bank an operating margin measures nothing |
What we would do
1. Only one of the forty-four has a forecast that does not work under any arithmetic. And not even with a margin thirteen points above its record would doubling revenue be enough.
2. The four most demanding are the four AI chip names. That is where consensus is stretched, and practically only there.
3. Eighteen come out comfortable, six of them with room to spare. That is what makes the rest of the result believable.
4. Seven have their profit contaminated by something that is not the business. That finding is independent of the forecast and stands on its own.
How it was done, and its limits
| Input | Where from |
|---|---|
| Revenue, operating profit, net profit | structured data from the US registry, rebuilt quarter by quarter |
| Share-based pay and wear on purchases | cash flow statement from the same registry, correcting for cumulative publication |
| Record margin | historical maximum over four-quarter windows, same measure |
| Forecasts | latest published estimate for the next financial year |
| Tax | a uniform, declared 21%. Using each company's real rate breaks down when one of them has a tax refund |
The limits of this test, stated plainly: it does not predict the price, it does not know the business — it knows nothing of patents, contracts or approvals — and it is deliberately conservative. An impossible forecast says the multiple calculated with it is false, not that the share is going to fall.
Forecasts change every week. This snapshot is from September 9, 2026.