The Beneish M-score: a screen for manipulation, and its limits
A screen flags Lowe's with an M-score of +7.3, far above the warning line. Take the score apart, find what drives it, and decide whether it means anything.
FinanceReading the accounts like an analystAbout 10 min: 10 readingUpdated
General education, not financial advice.
- Calculate the eight Beneish ratios and the M-score from a 10-K
- Read the score as a probability with known error rates, not a verdict
- Find which ratio drives a high score and whether a business event explains it
Builds toward: Judge a company. Read a company's accounts, apply Graham's tests, and make a rough valuation.
Reading the accounts like an analyst: lesson 10 of 10
- Reading the accounts like an analyst
- The three statements and how they link 10 min
- Common-size statements and trends 9 min
- Profitability ratios: margins, ROA and ROE 9 min
- Efficiency ratios and the cash conversion cycle 9 min
- DuPont analysis: where ROE comes from 10 min
- Reading the cash flow statement like an analyst 9 min
- Profit against cash: accruals and earnings quality 11 min
- When is a sale a sale? Revenue timing and channel stuffing 10 min
- The expense side: capitalised costs, reserves and big baths 8 min
- The Beneish M-score: a screen for manipulation, and its limits 10 min
Pairs with Messod D. Beneish, "The Detection of Earnings Manipulation", Financial Analysts Journal 55(5), 1999, pp. 24–36
Mission
From: the head of risk at an asset manager Due: tomorrow, half a page
Our quarterly screen has flagged Lowe's with a Beneish M-score of about +7.3, among the highest on our screen and miles above the −1.78 warning line. Do I need to escalate this? Recalculate it from the 10-K for the year to 30 January 2026, tell me what's driving it, and whether it means anything.
The documents (Lowe's Companies, Inc., Form 10-K for the year to 30 January 2026; millions of dollars; "t" is the latest year, "t−1" the year before):
| t | t−1 | |
|---|---|---|
| Net sales | 86,286 | 83,674 |
| Cost of sales | 57,401 | 55,797 |
| SG&A | 16,791 | 15,682 |
| Depreciation and amortisation | 2,194 | 1,972 |
| Receivables, net | 1,090 | 94 |
| Total current assets | 20,955 | 20,358 |
| Property, net | 18,362 | 17,649 |
| Total assets | 54,144 | 43,102 |
| Cash | 982 | 1,761 |
| Total current liabilities | 19,463 | 18,757 |
| Current maturities of long-term debt | 2,431 | 2,586 |
| Long-term debt | 37,490 | 32,901 |
| Income taxes payable (1) | 23 | 491 |
(1) Current accrued income taxes, from the XBRL data filed with the 10-K.
What it is
The toolkit for this job.
The model. Messod Beneish studied 74 US companies that had manipulated their earnings (found through SEC enforcement actions and press reports) against 2,332 other companies, using data from 1982 to 1992. He combined eight ratios, each comparing year t with year t−1, into a score:
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
| Ratio | What it compares | A high value suggests |
|---|---|---|
| DSRI | receivables ÷ sales, t over t−1 | sales booked but not collected |
| GMI | gross margin, t−1 over t | worsening margins |
| AQI | share of assets other than current assets and property, t over t−1 | costs being deferred |
| SGI | sales, t over t−1 | growth pressure |
| DEPI | depreciation rate, t−1 over t | slower depreciation |
| SGAI | SG&A ÷ sales, t over t−1 | rising overheads |
| TATA | accruals ÷ total assets, year t | profit not backed by cash |
| LVGI | (long-term debt + current liabilities) ÷ total assets, t over t−1 | rising leverage |
Accruals for TATA follow Sloan's balance-sheet definition: the change in current assets less cash, minus the change in current liabilities less current debt and income taxes payable, minus depreciation and amortisation.
The cut-off. Beneish converted the score into a probability. If missing a manipulator is thought 20 to 30 times as costly as wrongly flagging an honest company, the best cut-off is a score above −1.78. At that cut-off his model wrongly cleared 26% of the manipulators in the sample used to build it, and 50% in a later test sample, while wrongly flagging 13.8% and 7.2% of honest companies. It is a screen that tells you where to look, and it is wrong often in both directions.
What it can't see. The ratios are mechanical. Acquisitions, big changes in business mix, or a big-bath year in t−1 can push ratios far from 1 for innocent reasons. Beneish himself notes that goodwill from acquisitions can raise AQI.
Why it matters
The M-score is widely used in screens and investment research; in Beneish's own holdout tests it identified about half of the manipulating companies before their manipulation became public. But a number that prints "+7.3" next to a company's name can do damage if nobody checks what's behind it. The skill is to rebuild it, find the ratio that drives it, and say plainly whether it means anything.
Worked example
An imaginary company, Penrose: receivables rose from 100 to 180 while sales rose from 1,000 to 1,200.
- DSRI = (180 ÷ 1,200) ÷ (100 ÷ 1,000) = 0.15 ÷ 0.10 = 1.5.
- Its contribution to M is 0.920 × 1.5 = 1.38, against 0.92 if receivables had grown with sales: on its own it raises the score by about 0.46.
- If every other ratio were 1 and TATA 0, M would be −4.84 + 1.38 + 0.528 + 0.404 + 1.070 (SGI 1.2 × 0.892) + 0.115 − 0.172 − 0.327 ≈ −1.84: just below the cut-off. One ratio can move a company across the line.
Common mistake
Reporting the score without its parts. An M-score is a sum of eight contributions; a high one is almost always driven by one or two of them. Always show the contributions and ask what happened in the business to produce them.
Key terms
- M-score
- Beneish's weighted sum of eight ratios, used as a screen for earnings manipulation
- DSRI
- days' sales in receivables index
- AQI
- asset quality index: change in the share of "softer" assets
- TATA
- total accruals to total assets
- false positive
- an honest company the screen flags
- false negative
- a manipulator the screen misses
Practise the key terms
Tap a card to turn it over. 6 cards.
| Front | Back |
|---|---|
| M-score | Beneish's weighted sum of eight ratios, used as a screen for earnings manipulation |
| DSRI | days' sales in receivables index |
| AQI | asset quality index: change in the share of "softer" assets |
| TATA | total accruals to total assets |
| false positive | an honest company the screen flags |
| false negative | a manipulator the screen misses |
Task
Write the half-page note for the head of risk:
- the eight ratios and the M-score, recalculated from the documents;
- the contribution of each ratio, and which one drives the result;
- what in the 10-K explains it, with the note you found it in;
- whether to escalate, how confident you are, and what would change your mind.
Model answer
Show the model answer
| Ratio | Value | × weight = contribution |
|---|---|---|
| DSRI | 11.24 | 10.35 |
| GMI | 1.00 | 0.53 |
| AQI | 2.32 | 0.94 |
| SGI | 1.03 | 0.92 |
| DEPI | 0.94 | 0.11 |
| SGAI | 1.04 | −0.18 |
| TATA | −0.040 | −0.19 |
| LVGI | 0.88 | −0.29 |
| M (with −4.84) | +7.34 |
Workings for the two that matter: DSRI = (1,090 ÷ 86,286) ÷ (94 ÷ 83,674) = 0.01263 ÷ 0.00112 = 11.24. AQI = [1 − (20,955 + 18,362) ÷ 54,144] ÷ [1 − (20,358 + 17,649) ÷ 43,102] = 0.2738 ÷ 0.1182 = 2.32. Accruals = (597 + 779) − (706 + 155 + 468) − 2,194 = −2,147; TATA = −2,147 ÷ 54,144 = −0.040.
Lowe's M-score: a false alarm from two acquisitions · Source: Form 10-K, year to 30 January 2026
No escalation needed on this evidence. The score of +7.3 comes almost entirely from one ratio: receivables rose from 94 million to 1,090 million while sales grew 3%, making DSRI 11.2 and adding about 10.3 points to the score. The 10-K's acquisitions note explains it: the purchase-price allocations for Foundation Building Materials (bought October 2025) and Artisan Design Group (June 2025) include 1,114 million of receivables, more than the whole increase, from businesses that sell to trade customers on credit; the 10-K says sales on Lowe's branded and private-label cards aren't reflected in its receivables, because Synchrony Bank extends that credit. The same deals added goodwill and intangible assets (from 588 million to 9,853 million), which pushes AQI to 2.3.
With receivables growing only in line with sales (DSRI of 1) the score would be about −2.1, below the −1.78 line; neutralising the acquisitions' effect on asset quality too (AQI of 1) takes it to about −2.6, and sales growth also includes acquired sales; accruals are negative (cash ran ahead of profit). Beneish's own tests found the model wrongly flags 7–14% of honest companies, and this flag has a documented cause.
What would change my mind: receivables growing faster than the acquired businesses' sales next year, a rising allowance for credit losses, or a rise in AQI not explained by acquisitions.
How the expert gets there. They rebuild the score from the filing instead of trusting the screen, list the eight contributions, and see that one ratio supplies nearly all of it. Then they look for a business event that would move that ratio, and find it in the acquisitions note. They recalculate without it and state the model's error rates.
What a reviewer would flag:
- Escalating on the headline number without decomposing it.
- Dismissing the flag without finding the cause in the filing.
- Forgetting that removing the acquisitions' effect is itself an estimate (the acquired businesses were owned for only part of the year).
Practice
Transfer: a grey-zone score. Dell Technologies' 10-K for the year to 30 January 2026 gives an M-score of about −2.07 on the same method (income taxes payable not being shown separately, its change is taken as zero): below −1.78. A colleague says: "Below the line, so nothing to see." Open Dell's 10-K. In three sentences: do you agree, and what would change your mind?
A model answer
Not quite: below the line isn't a clean bill, and two ratios deserve a look, though the 10-K explains them. Receivables grew 71% on sales up 19% (DSRI 1.44) and gross margin fell from 22.2% to 20.0% (GMI 1.11), as the mix moved towards AI-optimised servers (the MD&A says "financing receivables and working capital were primarily affected by increased demand for our AI-optimized servers offerings"); accruals are negative because payables rose even faster. The model misses about a quarter to a half of manipulators at this cut-off, so the score clears nothing on its own. What would change my mind: receivable days still rising once the AI-server growth slows, or the margin falling further without a mix explanation.
How the expert gets there: they read a score near the line as "look at the parts", not as a pass.
What a reviewer would flag: treating −2.07 as proof of quality; ignoring the payables that make accruals look good.
The other view
Some researchers and practitioners argue that the M-score's coefficients, estimated on companies from the 1980s and early 1990s, fit today's companies poorly: service, software and platform businesses have different balance sheets, and acquisitions are far more common. Beneish and co-authors have argued that the score still carries information about future returns (Beneish, Lee and Nichols, 2013). Either way, the parts tell you more than the total.
In short
- 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.
- A score above −1.78 is Beneish's cut-off at 20–30:1 error costs; it is a screen, not a verdict.
- Its error rates are large in both directions.
- Always show the eight contributions and find the driver.
- Acquisitions and changes in business mix produce false alarms: check the notes.
Check yourself
Answer in your head first, then open the answer. Nothing is saved.
Receivables ÷ sales was 10% last year and 13% this year. What is DSRI, and how much does it add to M compared with a DSRI of 1?
Show answer
1.3; it adds 0.920 × 0.3 ≈ 0.28.
Why can an acquisition push the M-score up?
Show answer
It adds goodwill and intangibles (raising AQI) and may add receivables and sales unevenly (moving DSRI and SGI), none of which needs any manipulation.
A company scores −2.5. What can you conclude?
Show answer
Only that the screen doesn't flag it; the model misses a large share of manipulators, so the parts and the notes still need reading.
Further reading
- Messod D. Beneish, "The Detection of Earnings Manipulation", Financial Analysts Journal 55(5), 1999, pp. 24–36 (working-paper version, June 1999: model, Table 3; cut-offs, Table 5).
- Messod D. Beneish, Charles M. C. Lee and D. Craig Nichols, "Earnings Manipulation and Expected Returns", Financial Analysts Journal 69(2), 2013, pp. 57–82.
- Lowe's Companies, Inc. and Dell Technologies Inc., Forms 10-K for the years ended 30 January 2026 (SEC EDGAR).
- On this site: Profit against cash: accruals and earnings quality.
Finished? Mark it done to fill in your leaf on the Finance map.