Our data feed showed Alphabet at 16.9 times earnings. The real figure, measured on the operating business, was about 34.
Nothing was broken. The arithmetic was correct, the earnings were as reported, and the screen would have sorted it straight into any "cheap quality" filter you care to build.
The reason it was wrong is that $6.26 of the $9.11 of earnings per share — about 69% — came from revaluing stakes in private companies. Not cash. Not recurring. Not search, and not cloud.
A screener is a very fast way to be confidently wrong, and it fails in four distinct ways. Here they are, with a case for each.
The denominator is not what you think
Right ratio, wrong earnings
This is the Alphabet case, and it is the most common failure by a distance, because "earnings" is an accounting output that absorbs whatever happened that year.
Paper gains on stakes in private companies are real in the sense that they are properly recorded. They are unreal in every sense that matters to a buyer of the shares: they do not repeat, they produce no cash, and they will reverse if the valuations do.
Strip them out and the multiple roughly doubles. The company did not change. The question you were asking did.
Before dividing by earnings, read what is in them. Most of the time it takes two minutes and changes nothing. Occasionally it doubles the answer.
The model was built for a different animal
| The screen says | On | Why it is a category error |
|---|---|---|
| Bankruptcy score 1.08 | Chubb | The Altman-Z was built in 1968 on 66 manufacturers. It reads an insurer's reserves — money held against future claims — as dangerous leverage. It is the business model. |
| Return on capital 3.5% | Chubb | It divides by a capital base that includes the float. The float is other people's money; counting it as capital the shareholders supplied makes the return look a third of what it is. |
| EV/EBITDA of 12 | Chubb | Enterprise value subtracts cash and adds debt — both meaningless for a company whose business is holding financial assets. |
| P/E, return on equity, Altman-Z | IIPR | For a property trust, depreciation is charged on buildings that are not losing value. Earnings are therefore fiction; the industry reports AFFO for exactly this reason. |
The pattern is always the same: a ratio encodes an assumption about what a business looks like, and then it is applied to a business shaped differently. Nothing warns you. The number comes out, formatted to one decimal place, looking exactly as authoritative as a number that means something.
Regulated utilities have their own version of this, severe enough that we gave it its own article.
The data itself is wrong
This one is the hardest to catch, because there is nothing conceptually odd to notice. The feed simply reports the wrong number and every ratio built on it inherits the error silently.
Our own feed reports net property and equipment of $0.5 million for a real estate investment trust that owns 111 buildings. Anyone computing price-to-book, asset turnover or return on assets from that line gets a result that is not wrong by a little.
In September we discovered that our Estimates tab was showing Microsoft's forward price-to-earnings ratio as 20.1× when the correct figure was 26.2×.
The cause was dull and instructive. Our data provider returns quarterly estimates from the furthest future backwards — so code that asked for the first page expecting the next four quarters was quietly summing quarters from 2028. It only misfires when the provider publishes more than about ten future quarters, which is why it went unnoticed.
Of the thirty largest companies we then tested, thirteen were affected. And the error had a direction: because more distant estimates are higher, it always made the company look cheaper — in Microsoft's case by 23%.
It is fixed. We are telling you because it is the exact failure this article is about, and because a research site that only ever finds errors in other people's numbers is not being straight with you.
The number is right, and it is the wrong question
The subtlest one. Nothing is broken, nothing is miscategorised, the data is clean — and the metric is simply not the thing that decides the outcome.
Our Oracle report found revenue growing 30% and cloud infrastructure 121%. Both accurate, both genuinely impressive, and both beside the point next to free cash flow of minus $23.7 billion, debt of $156 billion and a BBB− rating.
A screen sorted on revenue growth puts that company near the top. A screen sorted on free cash flow does not return it at all. Same company, same quarter, same data — opposite conclusions, and only one of them is about whether the equity survives a bad year.
How to use one without being had
A screen is a search tool. The moment it becomes a judgement tool you have delegated your thinking to whoever chose the formula, on data neither of you has checked.
The single highest-value habit in this article. A bankruptcy score built on manufacturers cannot tell you anything about an insurer's reserves — so do not look at it, rather than looking at it and discounting it.
Not every number — two. Revenue and one balance sheet line. If those match the annual report, the feed is probably fine for that company. If one is off by an order of magnitude, nothing derived from it is usable.
Errors are far more common than mispricings, and as our own bug showed, they are not evenly distributed — they skew towards making things look cheap. Cheapness is the signal that most deserves a second look at the plumbing.
A company whose earnings were mostly paper gains, an insurer four ratios misread, and a fast-growing business consuming cash.
Each one states which metrics we refused to use, and why.
Frequently asked
Why can't you trust a stock screener?
Because a ratio can fail in four separate ways and the screen looks identical in all of them: the denominator may not be what you think it is, the model may have been built for a different kind of business, the underlying data may simply be wrong, or the number may be perfectly correct and answer a question that does not matter here. Alphabet's feed showed a price-to-earnings ratio of 16.9 when the figure on operating income was about 34.
Why is the P/E ratio useless for some companies?
Because earnings include things that are not the business. Roughly $6.26 of Alphabet's $9.11 of earnings per share came from revaluing stakes in private companies — about 69% of the reported figure, and not cash, not recurring, and not from search or cloud. For a REIT the problem is different again: depreciation of buildings that are not losing value makes earnings meaningless, which is why the industry reports AFFO instead.
Is the Altman-Z score reliable?
Only for the kind of company it was built on — 66 manufacturers, in 1968. It reads heavy fixed assets funded with debt as distress, so it flags insurers, utilities and property trusts as dangerous by construction. Chubb, one of the best underwriters in the world, screens with a bankruptcy score of 1.08.
Do stock screeners have wrong data?
Regularly, and it is the hardest failure to catch because nothing looks odd. Our own feed reports net property and equipment of $0.5 million for a REIT that owns 111 buildings. And we found a bug in our own forward price-to-earnings calculation that displayed Microsoft at 20.1 times instead of 26.2 — the error made 13 of the 30 largest companies we tested look cheaper than they were.