Lesson 8 of 9
Survivorship and selection
A screen on positive P/E sees 2,075 of 5,206 US listings. The 3,131 it cannot see are missing for a reason, and the reason biases the result upward.
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
- Survivorship bias is the special case of selection bias where the missing observations are the ones that failed.
- The missing rows are systematically worse than the visible ones, so almost every biased result points the same way: too good.
- The check is to count what is absent and ask why, before reading any statistic computed on what is present.
Every dataset is the set of things that made it into the dataset. When the process that decided which things did is related to the thing being measured, the result is biased, and the direction is almost always flattering.
What is visible to a screen on price to earnings
- Listings with a positive P/E2,075
- Listings with an indicator row but no positive P/E3,131
Counting what is missing
On 16 September 2026 the active US equity pool held 5,521 listings, of which 5,206 carried an indicator row. Of those, 2,075 had a positive price to earnings ratio: 39.9 percent.
The other 3,131 are not a random remainder. The single largest reason a listing has no positive P/E is that it did not report a profit, and the same store makes that visible directly: of the 4,640 listings with a computed return on equity, 1,988 had a negative one. That is 42.8 percent of the measurable pool losing money on equity.
So a screen that filters on a valuation multiple has already, silently, excluded most of the loss-makers before the first criterion is applied.
Why this makes every result look better
The harder version: the rows that already left
Everything above is about companies present today. The more damaging exclusion is the companies that are not.
A universe of active listings contains no company that was delisted, acquired at a discount or went bankrupt before today. An investor who was there at the time was exposed to all of them. Any historical statistic computed on today's active list has quietly removed the worst outcomes from the past, which is survivorship bias in its original form.
This product keeps the record rather than deleting it: every deactivation is reversible from an audit table, so what left can be counted. That is the mitigation and it is not a fix. A statistic computed on the active pool is still a statistic about survivors, whatever is archived beside it.
The three questions
Before reading any market statistic, including every one in this course:
How many rows were in the population, and how many reached the statistic. What decided which ones did. And is that decision related to the thing being measured.
If the third answer is yes, the number is biased and you can usually name the direction without computing anything. Missing failures push a result up. Missing loss-makers push a valuation statistic up. Missing delistings push a historical return up.
Check yourself
A screen filters on price to earnings across the US market. What has it excluded before its first criterion?
- Small companies
- Most of the companies that did not report a profit
- Nothing, until a threshold is set
Show the answer
Answer: B.
A company with no positive earnings has no positive ratio, so it cannot pass or fail the filter. It is simply not in the set, and on 16 September 2026 that was 3,131 of 5,206 listings with an indicator row.Which way does survivorship bias usually push a historical return?
- Down, because failures are recorded
- Up, because the companies that failed are the ones missing
- Either way, unpredictably
Show the answer
Answer: B.
The rows that left are disproportionately the bad outcomes. Computing on what remains removes those outcomes from the record, which flatters every statistic built on it.A published cohort's exclusions are all documented. Is the resulting statistic unbiased?
- Yes, documentation removes bias
- No: documentation makes the bias checkable, it does not remove it
- Only if the cohort is large enough
Show the answer
Answer: B.
Publishing the rule is what lets a reader see the selection and adjust. The statistic still describes the selected set, and quoting it as a statement about the whole market is still wrong.Sources
- Counts across active US equity listings, read from this product's own store on 16 September 2026: 5,521 in the pool, 5,206 holding an indicator row, 2,075 with a positive P/E, 4,640 with a computed return on equity of which 1,988 negative.
GET /api/public/distribution/peon ValueScreener, build of 15 September 2026: a cohort of 1,964 companies with equity REITs excluded as not applicable, median 20.37.
What this leaves out
- The 3,131 listings without a positive P/E are not all loss-makers. Some lack a usable filed figure for other reasons, and the return on equity count is given separately rather than derived from them, because the two sets overlap without matching.
- Counts on one date, on listings rather than deduplicated companies.
- This lesson counts what is absent from a present-day pool. It does not attempt to quantify what delisting has removed from any historical series, because that measurement is not made here.
- Nothing here is advice about any security.