Forensic scores
Beneish M flags 1 in 10 to 1 in 4 firms
On 2026-09-07 the Beneish M-Score's 90th percentile across 3,054 US companies was -1.07 and the 75th was -2.21, so its one cutoff sits between them.
By Javier SanzPublished
Founder of ValueScreener. Built Ninety Nine, a securities services provider, and sold it to a tier-one global bank; ran operations and internationalization at Alpaca, the brokerage API.
Who writes thisEditorial processLinkedIn (opens in a new tab)X (opens in a new tab)
On 2026-09-07, across 3,054 US companies with filed statements on record, the 90th percentile of the Beneish M-Score was -1.07 and the 75th was -2.21. The model has one published cutoff, -1.78, above which its own label is "possible manipulator", and it sits between those two percentiles.
So the share of the cohort on the flagged side of the line is more than a tenth and less than a quarter. That is the number this page exists to publish, and it is the number that should change how the flag is read.
| Percentile | M-Score |
|---|---|
| 5th | -3.96 |
| 10th | -3.29 |
| 25th | -2.82 |
| 50th | -2.54 |
| 75th | -2.21 |
| 90th | -1.07 |
| 95th | 1.05 |
From GET /api/public/distribution/beneish_m, build dated 2026-09-07. One
deduplicated company per row, drawn only from companies whose filed statements
we hold. Percentiles are the endpoint's own values rounded to two decimals, halves away from zero. More negative is better on this score, which is the opposite of
Altman's Z-Score and
Piotroski's F-Score, the other two forensic figures
in this set.
What a one-in-eight hit rate means
Accounting manipulation serious enough to be established after the fact is rare among listed US companies. Any screen firing on more than a tenth of the market is therefore not identifying manipulators. It is identifying the accounting patterns that manipulators share with a much larger group of ordinary companies.
That is not a defect in the model and it was never hidden. The M-Score is a detection model built to be sensitive, on the reasoning that missing a real manipulator costs more than looking twice at an honest one. A sensitive test run across a whole market produces mostly false positives whenever the thing it is looking for is rare, and no amount of care in the model changes that arithmetic.
The practical consequence is a rule about what the score may be used for. It can order a shortlist. It can tell you which set of statements to read first. It cannot conclude anything about a company on its own, and a page that named companies on the strength of it would be asserting something the model does not support.
This page names no company, and the product publishes no list of flagged companies. The score is shown on a company's own page, next to the reason it is what it is, and it is not turned into an accusation anywhere.
Why growth trips it
The formula is eight indices, each a ratio of this year to last:
M = -4.84 + 0.920 DSRI + 0.528 GMI + 0.404 AQI + 0.892 SGI
+ 0.115 DEPI - 0.172 SGAI + 4.679 TATA - 0.327 LVGI
SGI is sales growth: this year's revenue over last year's. Its coefficient is positive and it is the second largest of the eight. A company that grew revenue 40 percent contributes more to its own M-Score than a company that grew 5 percent, before anything else about it is considered.
DSRI is days sales in receivables, this year against last. A company that grew fast and whose receivables grew with the revenue moves that term up too, for the same reason and with no accounting judgement involved anywhere.
TATA is total accruals to total assets and it carries by far the largest coefficient, 4.679. That term is doing most of the discriminating, and it is the one closest to the underlying question of whether reported profit is turning into cash.
Put those together and the profile the model likes least is a fast-growing company with rising receivables and profit that has not yet arrived as cash. Some of those companies are manipulating. Most of them are growing.
The sectors where the flag is common
The 90th percentile of the score is past the -1.78 cutoff in six of the eleven sectors and below it in five.
| Sector | Companies | Median | 90th | 90th past the cutoff |
|---|---|---|---|---|
| Healthcare | 600 | -2.42 | 0.37 | yes |
| Technology | 581 | -2.48 | -0.62 | yes |
| Industrials | 508 | -2.55 | -1.66 | yes |
| Consumer Cyclical | 413 | -2.60 | -1.93 | no |
| Energy | 213 | -2.71 | -1.86 | no |
| Consumer Defensive | 176 | -2.55 | -1.35 | yes |
| Basic Materials | 176 | -2.57 | -0.98 | yes |
| Communication Services | 164 | -2.63 | -1.98 | no |
| Utilities | 94 | -2.59 | -2.21 | no |
| Financial Services | 83 | -2.41 | -2.01 | no |
| Real Estate | 45 | -2.53 | 0.35 | yes |
The medians barely move: every sector sits between -2.41 and -2.71, inside a range of 0.29. The tails move a great deal. That is the signature of a model whose ordinary output is stable and whose flag is driven by the specific companies at the top of each sector's distribution rather than by the sector.
Where the score is withheld entirely
740 companies were excluded from this cohort as not applicable: banks, capital markets firms, insurers, equity REITs and mortgage REITs. Beneish's study excluded financial firms, and several of the eight indices have no meaning without a gross margin, a classified balance sheet and a revenue line that behaves like revenue. For those companies the product shows no score and a reason, rather than a number computed out of whatever inputs happened to survive.
The figure is also withheld where only one fiscal year is on file, because every one of the eight indices compares two consecutive years.
The stored value is clamped to a range of -10 to 5. The upper clamp is visible in this build: the cohort's 99th percentile is exactly 5.00. At least one percent of these companies are therefore sitting at the ceiling, and for them the stored figure is the clamp rather than the computed score.
What this leaves out
These are cross-sectional percentiles from one build, dated 2026-09-07, on US listings only, at the verified tier, one company per row. They describe the shape of a score across a market. They are not a view on any company, and no company is named on this page.
A score above the cutoff is not evidence of manipulation, of fraud, or of any wrongdoing. It is the output of a statistical model that flags an accounting pattern, and the pattern has many innocent causes, the most common of which is rapid growth.
The share of the cohort past the cutoff is stated as a range, more than a tenth and less than a quarter, because that is what the published percentiles support. The exact count is not stated here because this product's public screener reports at most fifty matching rows to an anonymous caller, and a count nobody can reproduce is not a fact.
The scores are computed from this product's implementation of the model, which takes revenue, receivables, gross profit, current assets, property, total assets, depreciation, selling and administrative expense, long-term debt, current liabilities and cash from two consecutive filed years. Another implementation reading different line items from the same filings will produce a different score, and Beneish's original coefficients were estimated on a sample from an earlier accounting era.
Percentiles are winsorised at the 1st and 99th before they are computed, clamped to the observed edge rather than dropped, so the cohort size is the cohort that produced them. That clamp sat at -7.74 and 5.00 in this build.