Only 25 S&P 500 Companies Pass the Screen. Nine of Them Are Cases the Method Cannot Judge.

Data as of 18 September 2026. P/E and estimates from Finviz; the accounts from 10-K and 10-Q filings made to the SEC. Prices move in real time.

Last week we ran a simple screen across the Russell 2000: a stock is cheap if its P/E is lower than its estimated earnings growth, two years running. Of 1,944 companies, 115 passed — and when we audited them, only twenty survived. The rest were growing from a near-zero earnings base.

Now the same screen on the S&P 500. It fails for an entirely different reason.

Twenty-five out of five hundred and three

Of the index's 503 companies, 489 can be measured. Fourteen drop out because Finviz has no estimates for them or because they file no US accounts.

Twenty-five out of 503. Five percent — almost exactly the pass rate the Russell 2000 produced. So far the two indices look alike.

Here the problem is the banks

We opened the accounts on all twenty-five. The split:

Nine of the twenty-five cannot be judged, and five of those are banks. Bank of America, Citigroup, Citizens Financial, PNC and Charles Schwab. The method compares required earnings against a company's best-ever operating margin, and a bank has no operating margin to cap: its business is the spread between what it pays for money and what it charges to lend it. The arithmetic does not apply, and saying so is more honest than forcing it.

The other four unjudgeable names each have a specific reason. Crown Castle (CCI) has just sold half the company and its margin "record" comes out at 1,103% — a meaningless number. Fidelity National Information Services (FIS) has net income above operating income, a sign that what it earns is not coming from operating. And Everest Group (EG) and KeyCorp (KEY) have incomplete data in their filings.

The six that clear

Six out of 503. That is 1.2% of the index.

One caveat on the table: Invesco's P/E is not the Finviz twelve-month figure, which Finviz does not publish for that name, but price divided by its last reported earnings. The other five come from Finviz.

NVIDIA is the one worth stopping on

It clears the screen comfortably and its year 2 is possible: it would need a 54.1% operating margin when its record is 65.2%. In other words, it could deliver what is estimated of it without beating its own best-ever margin.

Now the other half of the sentence. To earn the $15.72 a share estimated of it, it would have to bill $742,624M against $302,970M today. That requires revenue growth of 81.8% a year, sustained.

The screen passes it because it is currently growing at 105.9%. Which means: the verdict does not say NVIDIA will do it, it says its current pace is enough to do it if the pace holds. The entire thesis fits inside that condition, and the two should not be confused.

Micron and the cycle

Micron shows year-1 earnings growth of 785.2%. That number comes from a base of $8.29 a share against an estimate of $73.38 — the memory cycle in its good phase.

For year 2 it would have to bill $338,121M against $90,274M today: nearly quadrupling revenue, which requires 187.3% a year. It is growing at 345.7%, so the screen passes it. But passing here means exactly one thing — if the memory cycle holds until its August 2027 year-end, another eleven months. Cycles do not usually do that.

Super Micro clears the screen and burns cash

SMCI demands the least margin of the six: 3.4% against an 11.4% record. On paper the most comfortable name on the list. Its operating cash flow last fiscal year was −$6,810M.

It is a useful reminder of what this screen measures and what it does not. It checks whether the promised earnings are arithmetically reachable, not whether the company generates cash along the way. Two different questions, and both need asking.

The ones that fall short

Western Digital is the most striking. It would have to go from $12,919M to $44,080M of revenue, and do it at a 59.4% margin when it has never exceeded 36%. It is growing at 43.8% a year, and the number still demands 84.6%.

Weyerhaeuser (WY) fails on the other rule. Its required margin is only 38% of its record and it would need to bill less than it bills today — but the earnings being asked of it exceed 90% of what is left after the cost of production. It is a timber company: however much operating margin headroom it has on paper, the profit cannot come out of an income statement that does not stretch that far.

What the in-house panel adds

In parallel we ran the same screen across a panel of fifteen names we follow closely. Two findings the full index does not give.

AMD is the furthest away. It demands a 43.8% operating margin against a 25.8% record — 170% of its best-ever mark. It would have to bill $129,681M against $41,305M today. That matches what we already found when we audited 44 consensus forecasts, where AMD came out as the only impossible name in the sample. There the label was "impossible" and here it is "does not work" — two different thresholds on the same fact: the number demands a margin the company has never had. Two independent screens, the same finding.

And a check that the method is stable. Broadcom appears in both screens with a different starting price ($361.14 against $348.04) and comes out at 95.1% of its record in one and 95.3% in the other. Two tenths of a point apart, on a verdict decided at 95%. That is not luck; it is what should happen if the arithmetic is built correctly, and it is worth showing.

From the rest of the panel: Applied Materials (AMAT) comes out impossible, at 153% of its record. Ciena (CIEN) does not work, at 151%, and neither does Coherent (COHR), at 130%. Arista Networks (ANET) is borderline at 103%. And Dell (DELL) is borderline at 96% — a company we analysed separately.

What had to be read by hand

No automated screen does any of this, and it changes verdicts:

Interest expense. Read from the 10-Qs and annualised for CHTR, FCX, LUV, AVGO and APP.

Minority interests. Freeport-McMoRan has 33% of its profit committed to its Indonesian partner, and Charter 15%. With those applied, Charter moves from "possible" to "borderline" and Freeport from a pass to "does not work". Two verdicts changed by a figure that only appears in the body of the filing.

Contaminated records. NXP Semiconductors needed 2018 excluded, the year it collected $2,000M from Qualcomm for the collapse of their merger: its clean record is the 36.4% of 2022, not the one that year produces. And Western Digital is audited against operating income, because its net income carries gains from its Sandisk stake.

What we would do

The same as with the Russell: not buy a list. What this screen produces is not six recommendations, it is six names whose accounts deserve to actually be opened — and two of them, Micron and NVIDIA, pass only on the condition that their current pace holds to their respective year-ends, August 2027 and January 2028.

The lesson from running the same screen across both indices is that screens fail differently depending on where you point them. In small caps you get fooled by the rebound from zero. In large caps you get fooled by believing a bank can be measured with a factory's yardstick.

In one sentence

Twenty-five of the S&P 500's 503 companies pass the P/E-below-growth screen, but nine are banks and cases where the method measures nothing, four are impossible, and only six can deliver what is estimated of them: and two of those six only if the cycle pushing them today holds to their 2027 and 2028 year-ends.

← All articles