Average P/E ratio: 21.16 on one count

On 2026-09-07 the average P/E ratio in the US market, taken as the median of the published distribution, was 21.16 across a cohort of 1,994 companies. An independent recount over a different population, on the same data and the same day, puts it between 20.75 and 20.94.

The usual answer to what a normal price to earnings ratio is comes from a book. Graham put the defensive investor's ceiling at 15 times earnings in 1973, and that number is still quoted. It is worth knowing what the market actually does, and it is worth knowing that the answer moves depending on who you count.

Two measurements, both taken on 2026-09-07, both from the same underlying data:

• The published distribution puts the median at 21.16, across a cohort of 1,994 companies.
• Counting every listing that held a P/E in the screener pool that day, without deduplicating companies and without excluding anything, put the median between 20.75 and 20.94, across 2,359 listings.

Neither is wrong. The gap is what a cohort definition is worth, and this page is mostly about that.

The published distribution

This is the authoritative figure, and it carries its cohort rule with it.

Cohort size 1,994. The rule, quoted from the response: coverage tier VERIFIED, one row per company, exact zeros excluded as coerced missing data, 156 excluded as not applicable because they are equity real estate investment trusts, winsorised at the 1st and 99th percentiles, minimum cohort 30. The winsorisation clamped the sample to between 1.32 and 640.0 before the breakpoints were computed, and clamped values are kept rather than dropped, so the cohort size is the cohort that produced the numbers.

By sector, with each cohort's size:

The independent recount

I recomputed the same statistic from the public screener without using the distribution at all, by partitioning each sector's P/E range into intervals until every one returned an exact count, then narrowing the interval holding the median by bisection. The population is deliberately different: every listing in the pool that holds a P/E, no deduplication, no tier filter, no exclusions, no winsorisation.

Two things in that table are worth pausing on.

The first is how flat it is. There is no peak. From 10 times to 45 times, every band of comparable width holds between 125 and 185 listings. A P/E of 12 and a P/E of 40 are both perfectly ordinary in the sense that has an equal number of companies standing next to them.

The second is where Graham's ceiling falls. Measured exactly, 766 of the 2,359 listings carry a multiple below 15, which is 32.5 percent. A rule written to describe a conservative minority still describes a third of the profitable market.

Where the two disagree, and why

Side by side, the published cohort median against the recounted interval.

Four sectors agree to within a quarter of a point. Four differ by one to two and a half points. Two differ by a lot. Every one of those differences has a named cause and none of them is a mistake.

Real estate is the big one, and it is the applicability rule. The published cohort excludes 156 equity real estate investment trusts from the P/E entirely, because the ratio is not meaningful for a business whose reported earnings are struck after a large non-cash depreciation charge on property. The recount includes them, because a numeric filter runs on the stored value and a stored value exists. Forty-four companies against 153 listings is that exclusion, and the published figure is the one to use.

Communication services is deduplication and tier. Sixty-four companies against 88 listings, and this is a sector with a high share of dual-class structures where one company carries two listings at the same earnings.

The rest is the cohort rule doing exactly what it says. One row per company rather than one per listing; only companies whose filed statements are on record; exact zeros treated as missing rather than as a value; and winsorisation at the first and ninety-ninth percentiles before the breakpoints were taken. Every one of those shifts a median slightly, and they do not all shift it the same way.

Both medians describe a real population. The published one describes companies whose accounts we hold and for which the ratio is meaningful. The recounted one describes every listing that has a number in the column, including the ones a careful reader should not compare on it.

What to take from a median at all

The sector, not the market. The lowest published sector median and the highest differ by a factor of nearly three. Comparing two companies' multiples across that gap is mostly comparing two sectors.

The cohort, not just the number. A median with no cohort rule attached is not a fact you can check. The difference between 11.42 and the interval containing 24.50 for real estate is entirely a cohort decision, and it is a decision worth understanding before quoting either figure.

Not a target. A company trading at half its sector's median is not therefore cheap; the market may be reading its earnings as temporary. The median is a description of where a company sits among its peers, which is the beginning of a question rather than the answer to one.

What this leaves out

Both measurements are from 2026-09-07, on prices stamped 2026-08-26, so every ratio in both was about two weeks old when it was read. The published distribution carries a build date of 2026-09-07 and its own staleness flag.

Both samples are profitable listings or companies only. A P/E is withheld when trailing earnings are not positive, so neither median says anything about the majority of US listings that had no positive earnings to divide by.

The recounted medians are stated as the interval that provably contains them rather than as a point, because that is what the method supports: the public endpoint reports an exact count only below its ceiling, so the figure was found by narrowing intervals rather than by sorting a list.

The recount is per listing and unweighted. A company with two share classes counts twice, and a hundred billion dollar company counts the same as a fifty million dollar one, so it is not an index-level P/E and will not agree with one.

The published cohort excludes equity real estate investment trusts from this metric and the recount does not. Where the two disagree most, that is why, and the published figure is the one that reflects a deliberate judgement about where the ratio means something.

Nothing here is a view on any company or any sector.