Survivorship bias: 815 tickers vanished in nine months

If you build a stock universe from today's ticker list and then backtest the last nine months with it, you are testing a market that did not exist. Every company that failed, merged or was taken private during that window is missing, and it is missing precisely because of what happened to it. Everyone knows this as survivorship bias. Almost nobody knows the size of it. We counted.

1. What we did

We keep unadjusted daily bars for every US symbol that traded. A ticker appears on a day if it traded that day, so a ticker that stops appearing has either been delisted, renamed, or has simply gone quiet. To separate those, we took the tickers present in all of the first five sessions of the window and found the ones absent from all of the last five. Then we looked every one of them up in the vendor's ticker reference, which carries a delisting date and a security type. Note the top line already hides the problem. The universe grew, from 11,713 to 12,539. A count of the list on two dates tells you nothing about what left it.

2. How many left

One ticker in fourteen was gone within nine months.

3. Checking that they really are gone

A ticker can drop out of our data for boring reasons, so before drawing conclusions we checked each one against the vendor's reference. Almost all of them were genuinely delisted. This matters: it means the disappearance is a real market event and not a hole in our recording.

4. Not all of them were companies

The raw count overstates the thing people worry about, because a large part of what delists was never common stock in the first place. Warrants, units, rights, preferred and notes together are 176 of the 815. These expire or convert on a schedule; their disappearance is an event on a calendar, not a failure. Closed funds account for another 244. Common stock is 375, less than half the headline count. So the honest version of the number is smaller than 7.2%. It is also not the number that matters.

5. The number that matters: were they tradable

A ticker that delisted while trading two hundred dollars a day was never on anyone's screen. The question is how many of the vanished names were still liquid enough to trade on their last day. Requiring a prior close of at least $1 and at least $1,000,000 of turnover on the day: 295 names were on screens, passing an ordinary liquidity filter, and then stopped existing. And they did not vanish at the edges of the window: the median one last qualified 45% of the way through it, so it traded normally for months first.

6. What it costs a backtest

Counting tickers is the wrong unit for a backtest. What a backtest consumes is symbol-days. So we counted every qualifying symbol-day in the window and asked how many of them belong to tickers that no longer exist. If you assemble your universe from today's ticker list, that is what disappears from the test. Roughly one day in fifty-seven. Small enough to feel harmless, which is the problem with it.

7. Why it is not harmless: they were different

If the missing days were a random sample, 1.75% would not matter much. They are not random. Here is what the two groups looked like on the days they qualified. The names that vanished traded at about half the price of the ones that survived, on a little over half the turnover, and with a daily range slightly below theirs. That combination is the day trading universe. The cheap and thinner end of the market, moving about as much, is exactly where the missing days are concentrated, and exactly where most short-term strategies do their shopping. So the bias does not spread itself thinly across your test. It removes a specific kind of name from a specific part of the universe, and the part it removes is the part you were most likely to be trading.

8. What to do about it

The fix costs nothing if you have per-day data: take the universe from the day being tested, not from today. The reason it stays broken in so many tests is that a current ticker list is the easiest file to obtain, and nothing about the result looks wrong afterwards. Limitations Security types come from one vendor's reference data, and the field is imperfect. We found notes typed as common stock, so we classified by instrument name as well and used the name where the two disagreed. Some misclassification will remain, which affects the split in section 4 but not the totals in sections 2, 3, 5 and 6. Tickers that were renamed rather than delisted appear here as one symbol vanishing and another appearing. We did not attempt to link them, so a rename is counted as a disappearance, which overstates section 2 slightly. The 21 with no delisting record are the most likely candidates. The window is 181 sessions, 2025-12-02 to 2026-08-21. Delisting activity is not constant, and a period with more failures would produce a larger number, not a smaller one. The tradability cut of $1 and $1,000,000 is our own, chosen to match an ordinary screener; a stricter cut would leave fewer than 295 names and a looser one more.