How the numbers are made

A wrong number is worse than no number, so some are refused outright:

• Negative shareholders' equity makes return on equity, price to book, and debt to equity meaningless. They show "n/m" (not meaningful), never a figure.
• Missing statements produce "n/a", never 0. A zero is a real measured value that would be percentile-ranked; writing zeros into gaps would silently poison the scores.
• A company with no interest expense has no interest coverage ratio; it shows "n/a" rather than infinity or a placeholder.
• Some US-listed companies report in another currency. An indicator that would mix USD market data with statements in that currency needs a reliable exchange rate, and without one we do not recompute it. The drawer says "recompute unavailable without FX rate" rather than quietly showing a mis-converted figure.
• A figure with no meaning for the kind of company that filed it is refused, and the gate is the class of filer rather than its sector. No enterprise multiple is printed for a bank: borrowings and cash are what a bank sells rather than how it funds itself, so the multiple measures nothing. No price to earnings is printed for a property trust: the depreciation charged against those earnings is on buildings that mostly appreciate. Each one renders "not meaningful" and carries the sentence explaining why, and every surface shows that same sentence.
• Not meaningful and not available are two different answers and we never collapse them. A bank's operating margin is perfectly well defined and we do not hold the line it needs: that is "n/a", and looking harder would help. A bank's current ratio cannot exist, because an unclassified balance sheet has no current split: that is "n/m", and looking harder would not. There is a third way to lose that difference and it does not look like a blank at all: an exact 0 arriving where a line was simply absent passes every check, so the published distributions discard an exact zero rather than rank it.

A ratio can be wrong for a company in three ways and each needs a different answer. The formula can be undefined, and then there is no number and printing one would be a lie: it is refused, as above. The formula can be sound while its scale is set by the business model, and then a percentile against the whole market is a category error: the figure is shown and ranked against its own company class instead. Or both can be sound while the reading is not obvious, and then the number carries a sentence next to it.

It is decided per figure and never per company, which is why a bank keeps an ordinary price to earnings while losing its current ratio, and why an insurer keeps the debt to equity a bank loses: an insurer's issued debt is ordinary debt, and a bank's deposits are its raw material rather than borrowing.

At full level the drawer checks the stored value against the one it just computed. More than 2% apart, or more than 0.5 points apart for the three score models, and the number carries a mismatch badge. Usually the cause is basis drift, the stored figure being trailing-twelve-month while the recomputation reads the latest annual statements, and the drawer says when that is what happened. A mismatch with no such explanation is a defect in our data, and you are meant to see it.

For a verified company: income statements, balance sheets and cash flow statements for every fiscal year the filings contain, ten years of them on screen; our own daily price archive, five years deep; trailing-twelve-month ratio snapshots; a profile with the sector, the exchange and the listing currency; and a market quote. Every statement row keeps its fiscal year, period, filing date and reported currency, which is what the drawer cites later. A series is as long as the filings make it. We do not pad it, interpolate it or extend it.

The same ratio can be built two ways: from the last four quarters of flow against the most recent quarter-end position, or from one fiscal year of flow against that year's own closing position. They are not the same measurement and the name does not distinguish them, so we do. Where a published trailing-twelve-month figure exists the pipeline prefers it, and where none exists it recomputes from the newest filed year. Which of the two produced any particular value is on that figure's source trail, and the screener reports a basis for every field it filters on: as of a filing, as of the market, or from a vendor.

The two clocks are the reason it matters. Anything built on a quote changes when the market changes and the business has done nothing; anything built on the accounts changes when a document is lodged. That is why a company page carries the price date and the compute date as two separate facts, and never as one.

Three figures left on 11 September 2026: the data vendor's discounted cash flow value, the margin of safety measured against it, and Forward P/E on analyst consensus. None of them could be replayed from a filing, which is the whole test. A saved screen or a guardrail written on one still runs, and tells you which line it dropped.

Every night, in this order (UTC):

Weekly: the market calendar and the listed universe on Sunday, split and dividend histories on Saturday and Sunday.

A company page carries two dates, never one. When the price was taken, and when the figures were computed. A recompute is not a fresh quote and collapsing the two into one date would let a stale price hide behind a recent calculation. A price more than a week old says so on the page.

Each kind of data is fetched and validated on its own. A failed or malformed response is thrown away and what was already stored stays exactly as it was: an error never overwrites a good statement. If an input an indicator needs is missing, the indicator is stored empty. Never zero, never an estimate.

For a bank, an insurer, a broker or a mortgage REIT the forensic trio was never defined, invested capital is not a meaningful base and neither is enterprise value. Five of the inputs are gone, so no score is published for those companies and a sentence stands in its place. The temptation the rule resists is renormalising: computing the score over the inputs that survived and rescaling it to look like the same score. That would be a different statistic under the same name, plotted on the same axis as everybody else's, and it is worse than a blank, because a blank tells the reader something is missing.

The composite applies the same rule to its pillars. A pillar with no indicator we could rank is left out and the others are averaged on their own weights, which is still the same measurement while the pillars that scored carry most of the weight. Below 80% of it, no composite is published and the company has no rank: averaged over what survived, a company read on its risk pillar alone would carry that one percentile as its whole score.

The same discipline applies to the cohorts a percentile is taken over. Fewer than 30 comparable companies with the figure on file and the percentile is refused rather than computed, because a percentile over a handful of companies is noise wearing a number.

It scores nothing, not zero, and there is no total. We do not publish a partial score. A test we cannot measure is not a test the company failed, and scoring the two the same way is a claim about a business made on evidence we do not have. The tests that did not run are named on the company's page with the reason for each.

There are two reasons and they mean different things. Either the test is undefined for that kind of filer, as with a bank, which reports no gross profit and has no invested capital, so margin stability and return on capital do not exist for it rather than being withheld; five of these six tests do not run on a bank. Or we simply do not hold the figure yet, and the next filing may fix that.

A ratio to net income is refused in a loss year too. A loss divided by a loss comes out positive, so a company burning cash would clear a cash-conversion test it is failing in every sense that matters.

From "Value Investing: The Use of Historical Financial Statement Information to Separate Winners from Losers", Journal of Accounting Research. Nine binary accounting tests, one point each, separating a business that is improving from one that is deteriorating. All nine, as we compute them:

1. Positive return on assets
2. Positive operating cash flow
3. Return on assets increasing year over year
4. Operating cash flow exceeds net income (the accruals test)
5. Leverage decreasing (long-term debt to total assets)
6. Current ratio increasing
7. No new shares issued (diluted weighted-average share count did not grow)
8. Gross margin increasing
9. Asset turnover increasing

We read 7 to 9 as strong, 4 to 6 as middling, 0 to 3 as weak. Where the prior year is missing, the year-over-year tests score zero and the drawer says that is what happened.

From "Financial Ratios, Discriminant Analysis and the Prediction of Corporate Bankruptcy", Journal of Finance. A weighted blend of five ratios, all scaled by assets or liabilities, built to estimate distress risk:

Z = 1.2 (WC/TA) + 1.4 (RE/TA) + 3.3 (EBIT/TA) + 0.6 (MktCap/TL) + 1.0 (Rev/TA)

Altman estimated the model on industrial firms. We do not run it on banks and insurers, whose balance sheets it was never built for. They show "not applicable" rather than a score that would mislead you.

From "The Detection of Earnings Manipulation", Financial Analysts Journal. Eight indices that tend to move when earnings are being manipulated, combined as M = -4.84 plus the weighted indices below. It requires two consecutive fiscal years of statements. A score above -1.78 flags a possible manipulator; below it, manipulation is unlikely on this model.

• One constant growth rate for the whole explicit stage; there is no fade path from high growth to terminal growth.
• One discount rate for every year and both stages. You enter it directly; the workbench does not compute a WACC for you.
• The terminal value is a Gordon growth perpetuity, so the discount rate must exceed the terminal growth rate; a large share of the result usually sits in this single term, and the workbench shows you that share.
• The starting free cash flow defaults to the latest annual cash flow statement; it is a single base year, not a normalized or cycle-adjusted figure. You can and should edit it.
• The equity bridge subtracts net debt from the most recent balance sheet, with no seasonal or intra-year adjustment.
• The share count is the current diluted count, held fixed; future dilution and buybacks are not modeled.
• No mid-year discounting convention, no scenario weighting, no sector-specific adjustments.

Save a set of assumptions as a named version and it keeps the complete inputs, the per-share value they produced and the time you saved them. The panel shows which assumptions differ from the ones on screen now. Six months later you can see what you thought and what you changed.

Every valuation on this site is one you run on your own assumptions: the workbench above and the earnings power value page beside it. We publish no fair value or price target of our own for any company. One of ours stood beside the price from 1 to 3 October 2026 and was withdrawn.

A guardrail can be tested against the years a company has already filed, so you can see whether it would ever have fired before a single night has passed. That is only honest when the value for an old fiscal year is built with the same formula as the value shown for the latest one. If the old value has to be assembled from whichever line items happen to be tagged that year, the series being tested is not the series being shown: it carries the same name and it is a different statistic, and a threshold against it produces something that looks like evidence and is not.

So the test is offered on plain ratios whose two inputs appear on the face of almost any annual report, and refused everywhere else with the reason. Two kinds are refused. Anything that needs a quote cannot be rebuilt from an annual report, which carries no share price, and our own price archive is five years deep against ten years of accounts. And a composite, a score or an aggregate whose definition is a choice cannot be rebuilt one year at a time under one formula. The tool also reports how many fiscal years it evaluated and over what range, so a ten-year claim about a company with four years on file cannot be made by accident.