Survivorship bias is the error of studying only the companies, funds or securities that still exist, so the ones that failed, were delisted or were acquired are missing. It makes past performance look better, and failure look rarer, than they were.
Where it appears
It appears in a back-test run on today's index members, in an average built from current customers that leaves out those who left, and in an adoption or hiring series that counts only companies still in a vendor's coverage. In each case the weak are removed before the count starts, so what remains looks strong.
An illustrative case: 100 listed companies are followed for a year, and 12 are delisted before it ends. A test that starts from the 88 still listed omits the 12, and any of them that fell before delisting no longer pulls the average down.
How to avoid it
- Test on the list as it stood on each date, not on today's.
- Keep delisted and acquired securities in the data, with a status and an end date.
- Reconcile counts: names at the start, plus entrants, minus exits, should equal names at the end. With 100 at the start, 15 entrants and 12 exits, expect 103.
- Report how many names were dropped, and why.
A vendor's history has the same weakness when a company leaves its coverage and its past rows go with it. Ask whether a provider keeps the history of companies that stopped being covered.
In Fokals data
The company index keeps one record per equity listing and keeps delisted listings, marked with a status, so joins to history stay intact. Check the list before any test:
select status, count(*)
from listed_securities
group by status;The weekly Market Series include families for new listings and for delistings. The use case for back-tests covers how to build the list as of a date.
Related terms
- Look-ahead bias: using information from after the test date.
- Point-in-time data: data stored as it stood on each date.
- Same-store cohort: a fixed group that keeps coverage changes out of a series.
- Security master: the reference table that should keep delisted securities.
Frequently asked questions
What is an example of survivorship bias in investing?
A fund database that drops funds when they close shows only the funds that lasted, so its average return overstates what investors in the whole group earned. The same happens when a strategy is tested on the companies in an index today: those removed after falling are absent, and the test never suffers their losses.
How do you avoid survivorship bias in a back-test?
Build the list as it stood on each test date from a source that keeps delisted, merged and bankrupt companies with the dates they ended. Check the number of names at the start against entrants and exits to the end. Where a source's record of exits is incomplete, say so in the results and treat them as an upper bound on what the strategy earned.
Is survivorship bias a type of selection bias?
Yes. Selection bias is any process that makes the sample differ from the population you want to describe. Survivorship bias is the case where the process is survival, so only entities that lasted enter the sample. The remedy is the same: start from everything that existed at the beginning and follow it to the end.
The queries and code on this page are examples to adapt. Test them in your own environment before you rely on them.