A list is a decision about the past
Every historical test starts by choosing which companies to test on. That choice feels administrative and it is the single most consequential thing in the exercise.
If the list is today's — the current index, the companies your screener returns, the names in a data file downloaded this morning — then membership required surviving until today. Nothing on it was delisted, wound up, merged away or taken private, because if it had been it would not be on the list.
So the test never meets a company that went to zero. The strategy is not avoiding disasters; the disasters were removed before it started. This is survivorship bias, and it does not announce itself, because the resulting numbers are all arithmetically correct.
Fifty companies, ten years later
An illustrative cohort. Switch between the list as it looks today and the list as it actually was.
The 14% is not a false number. It is a true average of a list that was selected for having done well enough to still exist.
Where the missing companies went
They are not exotic cases. A company leaves a list by being delisted — voluntarily or compulsorily, which 01.02.04 covered — by being merged into another, by being acquired and taken private, or by failing outright.
Index membership turns over for its own reasons too. 01.05.04 showed constituents being replaced at each review, so today's NIFTY 50 is not the NIFTY 50 of ten years ago, and testing 'the NIFTY 50' across a decade using today's names is testing a list that did not exist for most of it.
The same hole appears in fund performance tables. A fund that performed badly is often merged into a better one, and once merged its record leaves the table. The surviving category average is then an average of the survivors.
- Companies at the start
- 50
- Delisted, merged or wound up over ten years
- 8
- Average return of the 42 survivors
- +14.0%
- Average return of the 8 that left
- −62.0%
- Average across all 50, as it actually was
- +1.8%
A test on the current index constituents over fifteen years reports a very high return and a mild worst drawdown. Which is the more suspicious of the two?
What a clean version requires
Point-in-time membership. The list as it stood on each date in the test, not as it stands now. For an index this means the historical constituent record; for a screen it means the universe as the screener would have returned it that day.
The dead included. A company that was delisted has a price history up to the day it stopped trading, and that history has to be in the data with its ending recorded — including the case where the ending was worthless.
The ending stated. Delisted at a price, merged at a ratio into another company, or gone to nothing are three different outcomes and cannot be treated alike. A merger is not a loss; a wind-up usually is.
This is genuinely harder than it sounds, and a great deal of freely available data does not support it. The honest response is not to pretend otherwise. A test on a survivor list is worth running, provided its result is reported as what it is — an upper bound assembled with hindsight, not an estimate of what was available.
Which of these datasets carries survivorship bias?
How would a company that failed have entered this test? If there is no answer — no route by which a delisted, merged or worthless company could appear in the data — then the result does not describe a decade of investing. It describes a decade of investing in companies already known to have come through it, which was not a choice anybody could have made at the time.
Survivorship is information wrongly removed. The next bias is the opposite and worse: information wrongly included — facts a test used on dates when nobody could have had them.