Asked about survivorship bias, most equity data vendors offer the same reassurance. The delisted companies are in there, and the failures are included too. The claim is usually true, and it can be verified in about a minute by looking up Enron, Washington Mutual, or Lehman Brothers, each of which appears with years of daily price history attached.
That check is satisfying but shallow, because the presence of a row for a bankrupt company says almost nothing about whether a backtest can lose money on that company the way an investor did in 2007. That depends instead on three decisions buried further down the vendor’s pipeline:
- What return gets recorded on the day a stock actually leaves the market.
- Whether the universe being tested is the one that existed on the date of the test, with the membership and effective dates that applied at the time, rather than the one that exists today.
- Whether splits and spinoffs have been applied to the companies they actually happened to.
A vendor can pass the easy test about delisted companies being present in the database and still fail all three of these, which is why the old advice to make sure the data includes delisted stocks has lost most of its diagnostic value. Survivorship bias now tends to enter through the pipeline rather than the database, by way of decisions that rarely come up in a sales conversation.
Delisting returns and what gets recorded on the way out
Companies rarely disappear cleanly, and they disappear in two opposite directions that tend to be truncated in the same careless way. A struggling stock is usually halted, drifts into the murkier corners of the market on a trickle of volume, and then its price history simply stops, generally just before the collapse toward zero that a bankruptcy or a forced delisting produces. A company being acquired moves the other way, popping on the announcement and drifting toward the agreed sale price as the deal closes, so that when the history stops before the pop, an investor’s real gain never enters the record either.
In both cases the company remains in the database with years of records attached, and what goes missing is only the final data point, the one that captures what happened to an investor still holding the stock when it went away. Faced with a series that simply ends, most backtesting tools value the position at the last known price and carry on, with no warning and no flag in the output, which amounts to a silent assumption that the shares were sold at a price nobody could have traded at.
The Shumway delisting return correction
Tyler Shumway measured this in 1997 using CRSP, the historical record of the US stock market built at the University of Chicago’s Booth School of Business and treated by academic finance as the gold standard. In The Delisting Bias in CRSP Data he found that terminal returns, meaning the number recorded for the day a stock finally leaves the market, were routinely blank for companies delisted for performance reasons, and he proposed filling that gap with roughly −30 percent for NYSE and AMEX stocks. Two years later, working with Vincent A. Warther on Nasdaq companies in The Delisting Bias in CRSP’s Nasdaq Data and Its Implications for the Size Effect, he arrived at a considerably worse figure of about −55 percent. For many companies removed from an exchange the correct answer is close to a total loss rather than the flat line most databases quietly record. For a strategy that buys struggling companies an error of that size can account for the entire result.
Why the error is asymmetric across strategies
The missing endings cluster among small, thinly traded, cheap stocks, which is exactly where value and deep-bargain strategies prefer to fish. The opposite failure, in which generous buyout premiums never show up, does not offset it, because the two errors land on different kinds of strategy. Missing endings inflate the results of a book that holds failing companies, and the understatement caused by missing premiums falls instead on books that hold acquisition targets, which are rarely the same book.
The useful question to put to a vendor is therefore narrower than the familiar one. Instead of whether delisted stocks are included, it is what number gets recorded on the day a stock is delisted for poor performance, and how often that field is populated rather than left blank, broken down by reason and by year. Vendors call the label a delisting code, and CRSP’s version is the one most of the industry treats as the reference standard, though CRSP itself passed from the university to Morningstar in February 2026. An answer of *100 percent populated in every year* deserves skepticism, since real-world record keeping has gaps, particularly before the early 2000s and on smaller exchanges. The only reliable check is to have a vendor pull up a named company and show the return it recorded on the way out. Sears Holdings and Bed Bath & Beyond both serve well, as each was delisted after a Chapter 11 filing that ultimately left common shareholders with nothing.
Point-in-time universes and index membership history
The second decision concerns which companies were actually available to hold on any given morning in the past. The industry term for getting this right is a point-in-time universe, meaning the set of securities that were investable on each historical date, reconstructed with the index membership and effective dates that applied at the time. Everyone agrees in principle that a backtest should work this way, and the principle gets violated constantly anyway, largely because today’s index roster is a free download while a full history of who was in the index and when takes real effort, and often real money, to obtain.
A backtest of the S&P 500 running back to 1995 on today’s roster of 500 companies is, in effect, a strategy that buys stocks partly on the strength of their having grown large and durable enough to still be sitting in the index decades later.
Survivorship bias is often paired with look-ahead bias, and the two are related without being identical. Look-ahead bias uses information that was not knowable on the date of the signal, such as a restated earnings figure or a closing price applied to a trade earlier in the same session. Survivorship bias removes companies from the record altogether. Building a universe from today’s index roster manages both at once, since it drops the failures and also encodes knowledge of which companies were going to still be there decades later.
The versions that survive a code review are quieter. Most of the industry sorts companies into sectors using GICS, maintained jointly by S&P Dow Jones Indices and MSCI, a system that gets reorganized periodically and whose reorganizations are applied backward through history. In 2018 GICS created an entirely new sector, Communication Services, and moved Alphabet, Facebook (which would not be renamed Meta until 2021), Netflix, and Disney into it. A portfolio constructed today to be balanced across sectors, using that current map, ends up balanced against a grouping that did not exist and that nobody was trading against in 2005.
Something similar happens when a vendor extends coverage to a new company. Its history gets backfilled to the point where it started trading, and the companies worth adding are disproportionately those that survived long enough to become worth adding. The bias enters through the vendor’s coverage decisions rather than through anything the analyst did.
One related trap deserves mention. Tickers get recycled and handed to unrelated companies, so matching records on ticker alone can stitch two histories together into a single series that no investor ever held. The defence is a persistent security identifier that follows a company through ticker changes, mergers and delistings, rather than a ticker string that means different things in different decades.
Corporate actions, reverse splits and spinoffs
Adjusting historical prices for corporate actions, mainly splits and spinoffs, sounds like plain accounting, and it turns into a survivorship problem mainly because of which companies the errors happen to.
Why reverse splits concentrate in failing companies
Reverse splits are performed almost exclusively by companies whose share price has fallen far enough to threaten their listing and which buy themselves time by consolidating shares. That is the same pool of struggling companies whose final returns tend to go missing in the first place. Missing a 1-for-10 reverse split invents a 900 percent gain in a single day for a company circling the drain, and one error of that kind can manufacture a decade of fake outperformance in precisely the corner of the results an analyst hoped was genuine.
Spinoffs and the phantom loss
When a healthy company spins off a division, its share price falls by roughly the value distributed to shareholders, and recording that fall as a plain loss instead of a distribution books a large, fictitious loss on a healthy business on an arbitrary date. Pulling up the combined return for Abbott and AbbVie across the date they separated takes about 10 minutes and is probably the single most useful check in this article.
How to test a backtest for survivorship bias
Fixing all of this properly tends to cost money and time that may not be available, though establishing whether it is worth fixing takes roughly an afternoon.
The test is to run the strategy twice, once across the full universe and once restricted to the 100 largest companies by size at each point in time. That group of giants is close to immune to all three problems described above, since very large companies rarely delist, and when they do it is usually through a well-documented, publicly announced buyout that every data provider handles correctly.
The two runs are commonly compared using the Sharpe ratio, a measure of return relative to how bumpy the ride was, where roughly 1.0 counts as solid and anything above 2.0 starts to look too good. Suppose, purely as an illustration, that the run on the giants comes back near 0.4 while the full universe comes back near 1.9. A gap on that scale is roughly what needs accounting for.
Some of the difference could be genuine, since smaller companies do sometimes carry real extra return as compensation for their extra risk, a pattern known as the size effect. That explanation deserves less benefit of the doubt than it usually receives, because Shumway and Warther found that once missing delisting returns were properly filled in, the size effect vanished completely on Nasdaq while continuing to hold on the NYSE and AMEX. A famous and widely cited market pattern turned out, on at least one exchange, to be largely an artifact of the data.
Decile analysis and the total-loss test
Splitting the universe into 10 equal-sized groups by size, which researchers call deciles, shows how performance changes from decile 1, the largest, through to decile 10, the smallest. A real effect builds gradually across the range, whereas a data problem stays flat through deciles 1 to 8 and spikes only in 9 and 10, where the missing endings and botched reverse splits concentrate. The second check is harsher. Mark every company that left the database for a negative reason as a total loss, whatever the data actually shows. A strategy that still clears the bar under that assumption has settled the question, and one that does not has, at least, revealed what better data would be buying.
It is also worth measuring what share of the strategy’s trading activity came from companies that eventually failed. If that share is negligible, little of this matters much. If it accounts for a meaningful fraction of the holdings, the backtest has mostly been measuring how bankruptcies unfold rather than whatever signal the strategy was supposed to capture.
Three questions to put to a market data vendor
None of these three problems is exotic or hidden, and a careful vendor documents every one of them somewhere. That is precisely why the reassuring line on the sales page is the wrong place to look for answers. It settles the one question that can be answered in a five-minute call, whereas the questions that carry real risk require a data dictionary and a few coverage checks.
Three requests are worth making. The first is a breakdown of delisting outcomes by reason and by year. The second is confirmation that index membership history came straight from the index provider with genuine effective dates. The third is a walkthrough of a single spinoff, showing that it was handled correctly.
Whichever vendor ends up supplying the data, the comparison against the giants and the harsh total-loss test remain worth running independently, since a vendor can answer only for its own data while the pipeline that puts money at risk belongs to the analyst. The bias that survives into a set of results is the one nobody thought to check, and checking it costs an afternoon.
Frequently asked questions
What is survivorship bias in backtesting?
Survivorship bias is the distortion that arises when a backtest runs on a universe containing only the companies that still exist today. Because delisted companies are disproportionately failures, leaving them out inflates historical returns and understates drawdowns. The modern version of the problem is subtler, since the failures are usually present in the database but their final returns, their historical index membership, or their corporate actions are not handled correctly.
Does including delisted stocks fix survivorship bias?
No. Including delisted stocks settles only whether the rows exist. It says nothing about what return was recorded on the day each company left the market, whether the historical universe was reconstructed as it stood at the time, or whether splits and spinoffs were applied to the right companies. A dataset can contain every delisted company and still produce badly inflated backtest results.
What is a point-in-time universe?
A point-in-time universe is the set of securities that were actually investable on each historical date, rebuilt using the index membership and effective dates that applied at the time rather than today’s roster. Using a current index list across past dates introduces both survivorship bias and look-ahead bias, because the list encodes which companies were going to survive.
What delisting return should be used for a company delisted for poor performance?
Shumway’s 1997 work on CRSP proposed roughly −30 percent for NYSE and AMEX stocks, and the 1999 follow-up with Warther arrived at about −55 percent for Nasdaq. For many companies removed from an exchange for performance reasons the realistic answer is close to a total loss, which is far from the flat line most databases record when the price series simply stops.
The views, analysis and figures are the author’s own and do not necessarily reflect those of algoseek. This content is for educational purposes only and is not investment advice.
