Survivorship bias in backtesting: a worked example

What is survivorship bias in backtesting?

A backtest needs prices for every stock the strategy would have bought. Most free price sources only keep companies that are still listed. Build a test on them and every trade in a company that later disappeared is silently skipped.

The same happens when the list of stocks to test is today's index members, or today's list of tickers. The test then knows something no trader knew at the time: which companies would still be around.

How much does survivorship bias change a backtest?

Here is one case from our own research. The strategy buys US stocks after company insiders report buying their own shares on SEC Form 4. Same rules, same costs, same test. Only the price data changed.

With survivors only, the strategy passed our development test. With the missing companies back, it failed. Nothing about the idea changed; the first test simply never saw half of the trades.

Where did the missing companies go?

Later SEC filings show what happened to 6,761 of the purchases that had no price history in the survivors-only data: 63% of the companies were acquired, 25% deregistered and 9% went bankrupt.

So survivorship bias is not only about failures. Many of the missing companies were bought out. The problem is that the data, not the strategy, decided which trades counted. Putting the dropped trades back pulled the average down.

Does survivorship bias always make a backtest look better?

Usually, but any stock list picked with hindsight biases a test, in either direction. In October 2026 an error in joining two data files left one of our stock lists holding only companies that were later delisted. That list returned +4.9% a year in 2005 to 2020, so it was biased against the strategies we were testing.

We re-ran every rule on the stocks that were in the S&P 500 at each date, and published the correction with the audit of 18 strategies from X. The verdicts did not change.

How to avoid survivorship bias in a backtest

1. Build the stock list as it was on each date. Use the index members or listed stocks at the time, not today's list.
2. Use prices that include delisted companies. Our research used a database with 33,051 delisted US common stocks.
3. Match the company, not just the ticker. Tickers get reused. We checked each price series against the price the insiders themselves reported paying.
4. Give delisted positions an exit. When a company leaves after a bankruptcy or deregistration filing, we take an extra −30% at exit, a common research convention; acquisitions get none.
5. Screen out impossible prices. One corrupt price made an activist-stake strategy look like +2.35% a month. We drop any position with a single day above +300% or below −90%.
6. Fix all of these rules before you look at any result. Otherwise the choices drift toward the answer you hoped for.

In the insider-buying test above, the −30% exit applied to only 3 of 6,122 trades. Almost all of the drop came from trades the survivors-only data never contained.

Do forward track records have survivorship bias?

Not if they are recorded as they happen. A forward track stores each pick and its prices when they occur, so a company that later disappears stays on the record with the prices it actually had. That is why an Autrady track needs neither a holdout nor a database of delisted companies: it is out of sample from the first pick.

A seller's own record can still be survivorship-biased in a different way: by dropping picks, or newsletters, that went badly. See how to check a stock picker's track record.

Does Autrady's self-serve audit include delisted companies?

Not yet. The self-serve audit of uploaded signal files uses prices without delisted companies, so a pass there is labelled provisional. Treat any pass on survivors-only prices, ours included, as a reason to look closer, not as proof. The study above shows why.

For how long a record needs to be once the data is clean, see skill or luck: how long a track record must be.