The Beneish M-Score, Explained: Why a High Earnings-Quality Score Usually Isn't Fraud

A high Beneish M-Score sounds alarming, but for a fast-growing company it is often just growth showing up in the math. Here is what the model's eight ratios measure, where the threshold sits, and why an elevated score is a prompt to dig deeper rather than a verdict.

What the Beneish M-Score measures

The Beneish M-Score is a statistical model, published by Professor Messod Beneish in 1999, that estimates how likely a company is to have manipulated its reported earnings. It does not measure whether a business is good, cheap, or growing. It asks a narrower question: do the relationships between this year's numbers and last year's numbers resemble the patterns seen historically among firms that were later found to have overstated profits? The output is a single score. A less-negative score points to a stronger resemblance to that manipulation pattern; a more-negative score points away from it.

The model is deliberately built from ordinary financial-statement lines — receivables, sales, gross margin, depreciation, accruals — so it can be run from public filings (SEDAR+/EDGAR/SEDI) without any inside information. That accessibility is the point: a screen anyone can compute from a 10-K or an annual report. Investors often treat it as an early-warning filter that raises a hand for closer reading, not as a conclusion.

The eight variables, and what each one detects

The full model combines eight indices. Each index is a ratio-of-ratios: it takes a financial relationship this year and divides it by the same relationship last year, so a value near 1.0 means "no change" and a value above 1.0 means the relationship moved in a particular direction. Here is what each one is watching for, and — importantly — the direction that pushes the overall score up (toward the flag line).

Two of the eight — SGAI and LVGI — carry small negative weights, meaning higher values nudge the score slightly down. That is an empirical result, not a moral one: when Beneish fit the model, those variables happened to correlate the opposite way from the simple intuition. The takeaway is that not every "index above 1.0" is bad, and the eight variables are not equally important.

The model expression: how the pieces combine

The eight-variable score is a weighted sum. Conceptually, it is a constant plus each index multiplied by a fixed coefficient:

The single most influential term is TATA, with a coefficient of 4.679 — an order of magnitude larger than most of the others. In plain terms, the model cares enormously about how much of reported profit is accrual (booked but not yet collected in cash) versus real cash flow. DSRI (0.920) and SGI (0.892) come next: receivables outrunning sales, and fast sales growth. The remaining indices are refinements. Reading the coefficients tells you where to look first when a score is elevated — almost always, the accruals line is doing the heavy lifting.

A worked example: computing DSRI, GMI, and SGI

Take an illustrative company (round numbers, entirely hypothetical). Last year it reported sales of $2,000M, cost of goods sold of $1,200M, and receivables of $200M. This year: sales of $2,400M, COGS of $1,536M, and receivables of $300M.

• DSRI. Receivables were 200/2,000 = 10% of sales last year, and 300/2,400 = 12.5% this year. DSRI = 12.5% ÷ 10% = 1.25. Days-sales-in-receivables rose 25%: customers are being carried longer relative to what is being sold.
• GMI. Gross margin was (2,000 − 1,200)/2,000 = 40% last year, and (2,400 − 1,536)/2,400 = 36% this year. GMI = 40% ÷ 36% = 1.111. Because GMI compares prior to current, a value above 1.0 means margins shrank.
• SGI. Sales growth is simply 2,400 ÷ 2,000 = 1.20 — 20% growth.

Now fold those into the full model. Suppose the other five indices are unremarkable: AQI = 1.00, DEPI = 1.00, LVGI = 1.00, SGAI = 0.95, and accruals run at 3% of total assets (TATA = 0.03). The contributions are:

The result, roughly −1.86, sits just below the −1.78 flag line — close, but not flagged. Now watch the accrual sensitivity: if the same company's accruals were 5% of assets instead of 3% (TATA = 0.05), the TATA term jumps from +0.140 to +0.234, and the score rises to about −1.77 — now just over the line. A two-point-of-assets change in accruals, with everything else held constant, is enough to tip the verdict. That is the practical meaning of TATA's outsized coefficient, and a reminder to always inspect the accrual line before reacting to a borderline score.

The −1.78 threshold, and what "flagged" really means

For the eight-variable model, the conventional cut-off is −1.78. Because the scale runs negative, "above the threshold" means less negative. A score of −1.5 or −1.0 is above −1.78 and therefore inside the flag zone; a score of −2.5 or −3.0 is well below it and looks unremarkable. The threshold is a probability boundary, not a bright line — a firm at −1.79 and a firm at −1.77 are, for practical purposes, the same. Investors usually treat scores hugging the line as "worth a second look" rather than as any kind of verdict.

The most cited real-world illustration is historical: a group of Cornell business students computed elevated M-Scores for Enron before its collapse, using only public filings. That anecdote is often used to show the screen's value as an early filter — while Enron is a long-settled, adjudicated case, the lesson is about the method, not about labelling any company still trading today.

The 5-variable and 8-variable variants

Beneish also published a stripped-down five-variable version that drops SGAI, LVGI, and TATA and keeps DSRI, GMI, AQI, SGI, and DEPI:

Its flag threshold is different — around −2.22 — so the two versions are not interchangeable, and comparing a five-variable score to the −1.78 cut-off is a common error. The five-variable model exists mainly because the three dropped variables (especially the accrual and leverage terms) can be harder to compute cleanly from some datasets. The eight-variable model is generally regarded as the more complete tool, largely because it retains TATA, the accruals term that does most of the work. Running our hypothetical company through the five-variable formula gives about −2.47 — below its own −2.22 line, consistent with the eight-variable result of not being flagged.

Why high-growth, high-quality companies score elevated

The M-Score's biggest quirk is that several of its "danger" signals are also completely normal features of a healthy, fast-growing business. A company scaling quickly will often show sales growth (high SGI), receivables rising as it books more orders (high DSRI), and accruals building as it invests ahead of cash collection (high TATA). Every one of those pushes the score up — even when the accounting is entirely honest.

This makes an elevated M-Score not an accusation. It is best read as: "these numbers move the way manipulators' numbers moved, and also the way many aggressive growth companies' numbers move — figure out which story fits." A high score is frequently growth-driven, a common false positive rather than evidence of wrongdoing. The correct response is to open the filings and read: Are receivables rising because of a new enterprise-sales motion with longer payment terms? Are accruals up because of deferred revenue or a genuine build in capacity? The score points you to the question; the footnotes answer it.

Using the M-Score with Altman Z and Piotroski F

The M-Score answers only one of three very different questions, and it is strongest when read alongside two other classic screens that answer the other two.

Because they measure independent things, they triangulate. A name with a high M-Score, a weak (low) Z-Score, and a low F-Score is showing three unrelated yellow flags at once — a much stronger signal to slow down and read than any one screen alone. Conversely, a high M-Score paired with a robust Z-Score and a high F-Score often points back toward the growth false-positive explanation. Investors typically use the trio as a triage layer over a universe, then do the real work by hand on whatever surfaces.

Common mistakes and edge cases

• Applying it to banks, insurers, or other financial firms. The model was estimated on non-financial companies. Financial-sector balance sheets have entirely different accrual and leverage structures, so the score is not meaningful there.
• Reading one year in isolation. Every variable is a year-over-year change. A single acquisition, divestiture, spin-off, or accounting-standard change can distort the indices for one year and produce a spurious flag. Look at the trend across several years, not a lone reading.
• Broken comparables. The model assumes clean, like-for-like prior-year figures. Restatements, mergers, or a change in reporting standards break the "this year vs last year" comparison the indices depend on.
• Confusing the two thresholds. −1.78 is for the eight-variable model; roughly −2.22 is for the five-variable model. Judging a five-variable score against −1.78 (or vice versa) misclassifies companies.
• Treating a flag as proof. The score is a probability screen with a real false-positive rate. It identifies resemblance, not intent, and never substitutes for reading the filings.
• Ignoring coefficient direction. Because SGAI and LVGI carry negative weights, assuming "every index above 1.0 is bad" leads to misreading which factors actually drove a score.

Frequently asked questions

What is a bad M-Score?

For the eight-variable model, a score above −1.78 (that is, less negative) is the conventional flag for a higher likelihood of earnings manipulation. For the five-variable model the cut-off is around −2.22. "Bad" is relative — scores well below the threshold (−2.5, −3.0) look unremarkable, while scores hugging or crossing the line warrant a closer read rather than a conclusion.

Does a high M-Score mean a company committed fraud?

No. The M-Score is a statistical resemblance screen, not a finding of wrongdoing. Many honest, fast-growing companies score high because growth naturally lifts receivables, sales, and accruals — the same lines the model watches. A high score is a prompt to investigate the filings, not evidence of manipulation.

Why do growth stocks so often flag?

Rapid sales growth (SGI), receivables rising with new business (DSRI), and accruals building ahead of cash collection (TATA) all push the score up, and all three are ordinary consequences of scaling a business. That overlap between "growing fast" and "manipulating" is the model's main source of false positives.

How is the M-Score different from the Altman Z-Score?

They answer different questions. The M-Score estimates the likelihood of earnings manipulation; the Altman Z-Score estimates distance from financial distress or bankruptcy. A company can be financially sound (high Z) yet show an elevated M-Score, or vice versa. They are complementary, not substitutes.

How reliable is the Beneish M-Score?

In the original research it correctly flagged a large majority of the manipulators in its sample, but with a meaningful false-positive rate — it also raised a hand on a share of non-manipulators. It is best understood as a filter that reduces where you look, not a detector that decides. Its accuracy also depends on clean, comparable year-over-year data.

Can I use the M-Score on any company?

It is intended for non-financial companies with stable, comparable reporting. It is unreliable for banks and insurers, for firms in the middle of major restructuring or M&A, and for any period where an accounting-standard change breaks the prior-year comparison. Outside those cases it applies broadly, since it is built from standard filing lines.

How Quintarthai helps

Quintarthai computes the Beneish M-Score deterministically from public filings (SEDAR+/EDGAR/SEDI) across its US and Canadian coverage, showing the underlying eight variables so an elevated reading can be traced to the specific line — receivables, margin, or accruals — that drove it, and presenting it alongside the Altman Z-Score and Piotroski F-Score so the three screens can be read together. Every figure is sourced and reproducible, and the platform is an educational research tool: it surfaces the data and the model output for your own analysis, and does not provide investment advice or recommendations.