Lesson 1 of 9
A sample is not the world
The published median US P/E describes 1,964 companies out of 5,346 with filed accounts on record. Which ones are missing, and why it is never random.
By Javier Sanz
Founder of ValueScreener. Built Ninety Nine, a securities services provider, and sold it to a tier-one global bank; ran operations and internationalization at Alpaca, the brokerage API.
Who writes thisEditorial processLinkedIn (opens in a new tab)X (opens in a new tab)
In one minute
- A population is everything you want to describe. A sample is what you managed to measure, and they are never the same set.
- The gap between them is almost never random, so a summary of the sample is a summary of a selection.
- The honest form of every statistic states its cohort size and what was excluded from it.
Every market statistic is computed on the rows that were available. The rows that were not available are the whole subject of this lesson, because they left for reasons, and the reasons are usually related to the thing being measured.
From the market to the cohort a published median describes
- Active US listings5,521
- Companies with filed statements on record5,346
- Companies in the published P/E cohort1,964
The three sets
Each count below is read when this page is served and prints the day it was taken, because all three move: listings are added and withdrawn, filings arrive nightly, and a company enters the cohort the quarter it reports a profit.
The population is the thing you want a statement about: US listed companies. The active universe holds 5,532 (counted ) listings and 5,514 (counted ) deduplicated companies, which are two different numbers and not two names for one.
The measurable set is where filed statements are on record, which is 5,305 (counted ) companies. The companies outside it are not a random selection of the ones inside it.
The cohort is what the statistic was actually computed on: 2,625 (built ) companies for the published price to earnings distribution, after equity REITs were excluded as not applicable and exact zeros were dropped as coerced missing data.
So the published median of 19.84x (built ) is a statement about the third set. It is a correct number about a defined group, and it is not a number about the market.
Why the missing rows are never random
The largest exclusion here is companies without a usable positive ratio, and the main reason a company has none is that it did not earn a profit. That is not a coincidental absence. The companies that dropped out are systematically different on the exact dimension being measured.
What a sample can and cannot support
It can support a statement about the cohort: among US companies with a positive P/E on that build, the middle one was at 20.37. It can support a comparison with another cohort measured the same way.
It cannot support a statement about the market, because more than half of the market is not in it. And it cannot support a statement about any individual company, because a percentile is a place in a group.
The discipline this course asks for
Three questions, every time a market statistic appears, including the ones this product publishes.
How many rows is it computed on. What was excluded and for what stated reason. And is the exclusion related to the thing being measured. If the answer to the third is yes, the statistic describes a selection, and the sentence has to say so.
Check yourself
A published median P/E is computed on 1,964 of the 5,346 companies with accounts on record. What does it describe?
- The US market
- The 1,964 companies in that cohort, on the date it was built
- Every company with a share price
Show the answer
Answer: B.
A statistic describes the rows it was computed on. Extending it to rows that were excluded, especially rows excluded because they lacked the figure, is the error this lesson is about.Why is the absence of a P/E ratio not a random omission?
- Because the data provider chose which to include
- Because the main reason a company lacks one is that it did not earn a profit
- Because small companies file less often
Show the answer
Answer: B.
The missing rows differ from the present rows on exactly the dimension being measured. That relationship is what makes the exclusion a selection rather than noise.Which of these does a cohort statistic support?
- A forecast for one company in the cohort
- A comparison with another cohort measured the same way
- A claim about companies excluded from it
Show the answer
Answer: B.
Two cohorts built by the same rule are comparable. Neither an individual forecast nor a statement about excluded rows follows from a percentile.Sources
GET /api/stats/universeon ValueScreener, read 16 September 2026: 5,521 active US listings, 5,499 deduplicated companies, 5,346 with filed statements on record.GET /api/public/distribution/peon ValueScreener, build of 15 September 2026: a cohort of 1,964 companies, one row per company, equity REITs excluded as not applicable and exact zeros excluded as coerced missing data.
What this leaves out
- The counts in "The three sets" are read when the page is served and carry the day they were taken. The bar chart and the table are a single reading on 16 September 2026, kept as they were read, so the two will drift apart as listings, delistings and new filings move the first.
- The published distribution counts deduplicated companies while several later lessons count listings. The two are different populations and this course states which one each figure uses.
- Nothing here is advice about any security.