The Beneish M-Score is the academic standard for spotting manipulated earnings. Built from eight ratios in a company's financial statements, it flags a firm as a likely manipulator when the score rises above −1.78. It caught real frauds early, and business schools still teach it.
We computed it for every company in our universe with complete statements — 430 firms in the latest build, up from 344 when this article first appeared. 39 tripped the alarm, or 9.1% of the computable set, against 24 (7.0%) before. NVIDIA is one of them, at −1.16 — the 19th of 39 flags, almost exactly the median flag of −1.20. Its own score has not moved: every one of its eight inputs matches the first edition to rounding. What changed is the denominator, and the company the flag keeps.
The eight inputs, and what each one measures
The score is a weighted sum of eight ratios, each built from two consecutive annual financial statements. The model compares this year with last year and asks, in eight different ways, whether the numbers moved the way a manipulator's numbers tend to move. The weights and the −4.84 intercept come from Beneish's original estimation; the site applies them exactly as published, with no adjustment.
- DSRI, the receivables index (weight 0.920). Receivables this year as a share of sales, divided by the same ratio last year. A value above 1.0 means receivables grew faster than sales — the classic sign of revenue pulled forward or booked before it is real. A value near 1.0 means receivables kept pace with sales.
- GMI, the gross-margin index (weight 0.528). Last year's gross margin divided by this year's. A value above 1.0 means the margin deteriorated. Manipulators often show weakening margins before they resort to aggressive accounting, so the model treats a falling margin as a warning.
- AQI, the asset-quality index (weight 0.404). The share of total assets that is not current assets or plant and equipment, this year versus last year. A rising value means more of the balance sheet sits in soft, capitalised items — the kind of assets whose value is easiest to overstate.
- SGI, the sales-growth index (weight 0.892). This year's sales divided by last year's. Growth itself is not manipulation, but in Beneish's sample the manipulators grew faster than the honest firms, so the model adds points for fast growth.
- DEPI, the depreciation index (weight 0.115). Last year's depreciation rate divided by this year's. A value above 1.0 means the depreciation rate fell — which can mean assets are being written off more slowly, flattering earnings.
- SGAI, the SG&A index (weight −0.172). Selling, general and administrative expense as a share of sales, this year versus last year. A value below 1.0 means administrative costs grew more slowly than sales, which the model reads as a small point in the company's favour — hence the negative weight.
- TATA, total accruals to total assets (weight 4.679). Accruals — the gap between reported earnings and cash flow — scaled by assets. This carries the largest weight by far, because accruals are where manipulation most directly shows up: earnings that do not turn into cash. A value of 0.084, as NVIDIA shows, is modest; the model's most accrual-led flags sit much higher.
- LVGI, the leverage index (weight −0.327). Total liabilities as a share of assets, this year versus last year. A value below 1.0 means leverage fell, which the model reads as a point in the company's favour.
Two properties of this list matter for everything that follows. First, every input is a change — this year against last year — so the score is a statement about movement, not about a level. A company can be large, profitable and entirely honest and still score high if its ratios moved sharply. Second, the inputs are computed from the latest complete pair of annual statements for each company, so the pairs are not all from the same year. That is why the flag list mixes 2026, 2025 and much older base years, a point we return to below.
NVIDIA, input by input
The score is a weighted sum, and NVIDIA's decomposition shows where the push comes from. Sales growth (SGI 1.65) contributes +1.48 points, the largest single push: revenue rose from $130.5 billion to $215.9 billion in one fiscal year. Asset quality (AQI 1.52) adds +0.61, and receivables (DSRI 1.01) add +0.93 — but read that one carefully: a DSRI near 1.0 means receivables grew almost exactly in line with sales, which is the opposite of the classic channel-stuffing pattern. Margin deterioration (GMI 1.06) adds +0.56, as the gross margin eased from 75.0% to 71.1%; accruals (TATA 0.084) add +0.39; and slower depreciation (DEPI 1.06) adds +0.12. Two inputs pull the other way: administrative costs grew more slowly than sales (SGAI 0.79, −0.14) and leverage fell (LVGI 0.83, −0.27). The total is −1.16, above the −1.78 line.
| Input | Value | Weight | Contribution |
|---|---|---|---|
| DSRI (receivables vs. sales) | 1.01 | 0.920 | +0.93 |
| GMI (gross margin, prior / current) | 1.06 | 0.528 | +0.56 |
| AQI (asset quality) | 1.52 | 0.404 | +0.61 |
| SGI (sales growth) | 1.65 | 0.892 | +1.48 |
| DEPI (depreciation rate, prior / current) | 1.06 | 0.115 | +0.12 |
| SGAI (SGA vs. sales) | 0.79 | −0.172 | −0.14 |
| TATA (accruals / assets) | 0.084 | 4.679 | +0.39 |
| LVGI (leverage) | 0.83 | −0.327 | −0.27 |
| Intercept | — | — | −4.84 |
| M-Score (fiscal 2026 vs. fiscal 2025) | — | — | −1.16 |
Checking the arithmetic
A reader should be able to reproduce the headline number from the table, so here is the addition. Start with the intercept, −4.84. Add the six positive contributions: DSRI +0.93, GMI +0.56, AQI +0.61, SGI +1.48, DEPI +0.12 and TATA +0.39, which together add +4.09 and bring the running total to −0.75. Then subtract the two negative contributions, SGAI −0.14 and LVGI −0.27, which together remove 0.41 and bring the total to −1.16. The table and the headline agree.
The same arithmetic shows how much each input matters. Erase sales growth entirely — set SGI to 1.0, so its contribution falls from +1.48 to +0.89 — and the score drops to −1.74, still above the −1.78 line. That is the article's original point, and it is worth pausing on: even a company with no growth at all would still trip this screen on NVIDIA's other seven inputs. Erase accruals instead (TATA to zero) and the score falls only to −1.55. Erase the receivables push (DSRI to 1.0, contribution +0.92 instead of +0.93) and almost nothing changes, because a DSRI of 1.01 is already nearly neutral — the +0.93 looks large only because the weight 0.920 is applied to a value near 1.0, and any company with receivables roughly in line with sales gets about +0.92 from this input alone. That is a property of the formula, not a finding about NVIDIA: the DSRI term contributes roughly +0.92 to every company whose receivables track sales, and the intercept of −4.84 is what keeps most companies below the line.
The practical consequence is that the flag is a pile-up. No single input carries the accusation, and no single input's removal clears it. That cuts both ways: it makes the flag harder to dismiss as one ratio's quirk, and harder to treat as a smoking gun, because the case rests on eight mildly unusual readings arriving together.
What −1.78 is, and what it is not
The threshold −1.78 is a classification cut, not a probability. Beneish chose it to separate his sample of known manipulators from a matched sample of honest firms: above the line, a company resembled the manipulators more than the honest firms on these eight dimensions. It does not mean a flagged company has a particular probability of having committed fraud, and it does not mean an unflagged company is clean.
The base rate is the reason this distinction matters. In the latest build, 39 of 430 computable companies trip the screen — 9.1%. Actual earnings fraud is far rarer than one company in eleven. So even if the model were perfectly calibrated, most flags would be false alarms: companies whose ratios moved the way manipulators' ratios move, for entirely legitimate reasons. The model's value is not that it convicts; it is that it narrows a universe of hundreds of companies to a few dozen worth a closer look, and it shows its working so the reader can judge whether the resemblance is meaningful.
That is also why the score is reported as a number, not a verdict. A company at −1.16 (NVIDIA) and a company at 44.02 (Crown Holdings, on a 2014 base year) both sit above the same line, but they are not the same kind of flag. The first is a pile-up of mild readings on current financials; the second is an extreme value driven by a corporate event a decade ago. The threshold treats them identically. The reader should not.
The denominator moved
The October 30 fundamentals rebuild widened the fundamentals cache and recomputed every lens: Beneish computable firms rose from 344 to 430, an increase of 86, and flags rose from 24 to 39, an increase of 15. The flag rate rose from 7.0% to 9.1% — but that rise partly reflects coverage, not conduct: companies that were previously uncomputable entered the set, and a rate computed over a different set is not directly comparable to the old one. The honest statement is that 39 of today's 430 computable companies trip the screen, and that NVIDIA is one of them, as before.
Coverage is not conduct
The rebuild deserves a closer look, because it is easy to misread. When the computable set grows from 344 to 430, two things change at once: the number of companies the screen can judge, and the composition of that set. The 86 companies that entered were not a random sample of the market — they were the companies whose statements were previously incomplete or unmapped, and which the wider cache could now parse. If those newly computable companies flag at a higher rate than the original 344, the overall flag rate rises even if no company's behaviour changed at all.
The arithmetic makes the point. The original set flagged 24 of 344 (7.0%). The new set flags 39 of 430 (9.1%). The 15 additional flags came from the 86 newly computable companies, which means the new entrants flagged at 15 of 86, or 17.4% — more than twice the rate of the original set. That is not evidence that the market got more manipulative in a month. It is evidence that the companies the cache could not previously parse are different from the ones it could: smaller, less covered, more likely to have the volatile ratios that trip a screen. A wider net catches more of everything, including more false positives.
This is a general property of screens, not a flaw in this one. Any time the universe changes, the rate changes with it, and the change in the rate says as much about the universe as about the companies in it. The correct way to read the rebuild is as two separate statements: the screen now covers 430 companies instead of 344, which is a genuine improvement in coverage; and 39 of those 430 trip the screen, which is a statement about the current set, not a trend. Comparing 9.1% with 7.0% as if they were two readings of the same population is the error to avoid. The next rebuild will change the denominator again, and the rate will move again for the same reason.
The other 38 flags
NVIDIA is not an extreme flag; it is a typical one. Of the 39 flags, only 2 score above zero; 15 sit between −1.0 and zero, and 22 sit between the −1.78 line and −1.0. The median flag is −1.20, a hair below NVIDIA's −1.16. Growth is the common thread but not a universal one: 35 of the 39 grew revenue year over year, 13 grew by 50% or more, and 9 more than doubled — yet 4 flags shrank or stood still, including Wolfspeed, whose revenue fell by nearly two-thirds (SGI 0.36) and which flags on other inputs entirely. When each flag is sorted by its largest positive contributor, sales growth leads 17, receivables lead 12, accruals lead 5, and asset quality leads 5. The model has more than one way to cry wolf.
| Rank | Company | M-Score | Base year | SGI |
|---|---|---|---|---|
| 1 | Crown Holdings (CCK) | 44.02 | 2014 | 34.85 |
| 2 | Spectrum Brands (SPB) | 0.56 | 2023 | 0.93 |
| 3 | Aviat Networks (AVNW) | −0.03 | 2021 | 1.15 |
| 4 | Strata Critical Medical (SRTA) | −0.07 | 2025 | 1.34 |
| 5 | ADMA Biologics (ADMA) | −0.18 | 2025 | 1.20 |
| 6 | Western Digital (WDC) | −0.19 | 2026 | 1.36 |
| 7 | Aehr Test Systems (AEHR) | −0.34 | 2022 | 3.06 |
| 8 | Credo Technology (CRDO) | −0.50 | 2026 | 3.06 |
| 9 | Fabrinet (FN) | −0.51 | 2026 | 1.36 |
| 10 | Haverty Furniture (HVT) | −0.58 | 2019 | 0.98 |
How to read the flag list
The flag list is a reading list with mixed vintages, not a league table. Only 23 of the 39 flags rest on 2025 or 2026 financials; the other 16 rest on older pairs, some years old, because the screen uses each company's latest complete pair of statements and some companies stop supplying a complete pair. The highest score on the list illustrates the problem: Crown Holdings scores 44.02, a number so extreme it looks like a different scale — and it is, in effect: its latest complete pair is fiscal 2014, when revenue jumped nearly 35-fold and the gross-margin index hit 33.8, the signature of a corporate event, not a current fraud signal. The next names are more current: Spectrum Brands at 0.56 (2023), Aviat Networks at −0.03 (2021), and Strata Critical Medical at −0.07 (2025) — the company that topped the first edition's list, now fourth, its own score unchanged. Aehr Test Systems is still there at −0.34 with an SGI of 3.06, but on a 2022 base year.
The vintage problem has a specific cause. The screen needs two consecutive complete annual statements. When a company's filings stop yielding a complete pair — a restatement, a missing tag, a change in reporting — the screen keeps the last complete pair rather than dropping the company. That is a defensible choice (dropping companies would hide them), but it means the list always carries stale entries, and the stale entries are the ones with the most extreme scores, because extreme scores are usually produced by corporate events that also disrupt the data. A reader who sorts by score and starts at the top is therefore reading the stalest flags first.
The practical way to use the list is to filter before ranking. Start with the base year: keep only flags on 2025 or 2026 financials, which leaves 23 of the 39. Then look at the largest contributor for each remaining flag: a flag led by receivables (DSRI) is a different question from one led by growth (SGI) or accruals (TATA). Then ask whether a corporate event explains the move — a merger, a divestiture, a reclassification — because those move these ratios more than any accounting choice. Only after those three filters does the ranking order carry information, and even then it ranks resemblance to a 1990s sample, not likelihood of fraud.
Why the model fires on growth
The model's suspicion has a vintage. Beneish's sample comes from the 1990s, when explosive growth often was cosmetic: companies dressing up their books tended to grow fast on paper first. In 2026, a company selling the world's AI accelerators during a supply shortage can grow that fast with entirely real invoices. The model also notes asset quality (AQI 1.52) and accruals (TATA 0.084) — worth a look, nowhere near proof. And a score above the line is not a probability: it means the company resembles the manipulators in a 1990s sample on these eight dimensions, not that it probably committed fraud.
The hyper-growth names the model cannot distinguish from frauds are still there, and there are more of them now: Credo Technology (SGI 3.06, 2026), Amprius Technologies (SGI 3.02, 2025), Sandisk (SGI 2.75, 2026), AeroVironment (SGI 2.41, 2026), and Applied Digital (SGI 6.48, 2023) all trip the alarm on growth-heavy profiles. None of that is evidence about any of these companies; it is evidence about a model whose growth input cannot tell a boom from a fabrication.
There is a second, quieter reason growth breaks the model. Consider a company with every input at its neutral value — each index at 1.0 and accruals at zero. Its score is −2.48, comfortably below the line. Now let that company double its sales and nothing else (SGI 2.0 instead of 1.0). The growth term alone adds +0.89, and the score rises to −1.59 — above the line. Doubling revenue in a year is, by itself, nearly enough to trip the screen. That is why 17 of the 39 flags are led by sales growth, and why 9 of the flags more than doubled revenue. The model was built in an era when doubling revenue in a year was itself suspicious. In a hardware cycle, it is the normal shape of a boom. The screen has no way to know which era it is in.
Before reacting to any flag, check four things
- The base year. A flag on 2014 or 2021 financials is not a statement about this year.
- The largest contributor. A flag led by receivables (DSRI) tells a different story from one led by growth (SGI) or accruals (TATA).
- Whether the jump is a corporate event. Mergers, divestitures, and reclassifications move these ratios more than any accounting choice.
- Whether the inputs are current. If the latest complete pair is years old, the flag is stale until the data refreshes.
Three limits belong in the open
First, the original model was built excluding financial firms, but the screen applies to every company because no sector breakdown is available in the data — some flags may be sector artifacts. A bank or insurer's balance sheet does not look like a manufacturer's: its "assets" are largely financial instruments, its receivables are loans, and its accruals behave differently. The model was never estimated on such firms, so a flag on a financial company is a weaker signal than the same score on an industrial company, and the reader should discount it accordingly.
Second, the computation approximates where the data runs out: securities are treated as zero inside asset quality, accruals use net income instead of continuing-operations income, and long-term debt falls back to total liabilities minus current liabilities when the direct figure is missing. Each approximation bends the score in a known direction. Treating securities as zero understates the soft-asset share, which pulls AQI down and makes the score slightly lower than a full computation would give — the screen is, on this point, conservative. Using net income instead of continuing-operations income folds one-off gains and losses into accruals, which can push TATA up or down depending on the year; a company with a large discontinued-operations loss will show higher accruals than the model intends, and a company with a large one-off gain will show lower. The long-term-debt fallback is the least consequential of the three, because the leverage index enters with a small weight (−0.327) and the fallback tracks the true figure closely for most companies. None of these approximations is hidden — they are stated in the data note on every flag — but they mean the published score is an estimate of the model's score, not the model's score itself, and two companies with the same published score may have different true scores.
Third, the screen uses each company's latest complete pair of statements, so base years differ across the list. None of this makes the screen useless; it makes the flag a starting question with a visible method, not a conclusion.
How we use it
A red flag is a question, not a verdict. Our Red Flags section lists every flagged company with the model's inputs visible, precisely so readers can do what the model can't: read the context — and check the base year before reacting to a rank. When a screen cries wolf on the most scrutinized company on earth, the correct response isn't to short it — it's to check why it fired, and whether the inputs behind the score are current.
What would change this reading?
Not a single quarter of results — the screen uses annual pairs — but the next fundamentals rebuild: whether NVIDIA's next annual pair keeps it flagged, whether the current-vintage share of the flag list rises above today's 23 of 39, and whether stale pairs like Crown Holdings' 2014 set ever refresh into the present. If NVIDIA's receivables ever start outpacing its sales (DSRI climbing well above 1.0) or its accruals jump toward the 0.4 region where the list's most accrual-led flags sit, the flag would deserve a harder look. Until then, it remains what it was in the first edition: a famous detector, firing, on a company whose growth it was never built to understand.
The tool is useful. Blind faith in it is not. Verify everything.
Revision — October 31, 2026: census refreshed after the October 30 fundamentals rebuild, from 344 companies (24 flagged) to 430 companies (39 flagged). NVIDIA's score and all eight inputs are unchanged from the first edition; the flag-list ranking changed — Strata Critical Medical, formerly first, now ranks fourth — and the sections on NVIDIA's decomposition and the flagged set's composition are new. No other figure from the first edition changed.
Data: SEC company financial statements (companyfacts) via the site's fundamentals cache, latest annual pairs as of the October 30, 2026 build; M-Score computed with the original eight-variable model; 430 companies had complete inputs. Method approximations: securities treated as zero in asset quality; accruals use net income; long-term debt falls back to liabilities minus current liabilities where missing. Not investment advice.