Survivorship bias, explained

Q: What is survivorship bias in investing?

Survivorship bias is the error of drawing conclusions from only the things that lasted. In investing it usually means testing a strategy on the companies or funds that still exist today, which silently removes every company that went bankrupt, was delisted, or was acquired after collapsing. The remaining sample is healthier than the real investable universe ever was.

Q: How much does survivorship bias inflate backtest returns?

It depends heavily on the universe and period, but academic studies of mutual fund databases have generally found biases on the order of roughly half a percent to one and a half percent per year. For individual stock universes the distortion can be larger, because delisting returns are frequently severe and are often missing from the data entirely rather than recorded as losses.

Q: What is delisting bias?

Delisting bias is a specific form of survivorship bias where a security disappears from the dataset when it stops trading, and the final loss is never recorded. Research on CRSP data by Shumway found that missing delisting returns meaningfully biased results, and that performance-related delistings warranted a substantial assumed loss rather than a blank.

Q: Does survivorship bias affect index funds?

A real index fund you actually hold does not suffer from it, because it holds losers all the way down until they are removed. The bias appears when someone backtests an index using its current membership list, which credits the index with owning today's winners during years when it did not.

Q: How do I avoid survivorship bias in my own testing?

Use a point-in-time dataset that includes securities which later delisted, with their final returns recorded. Reconstruct universe membership as of each historical decision date rather than using today's list. Where survivorship-free data is unavailable, state the limitation plainly and treat the results as an upper bound rather than an estimate.

The most expensive mistake in backtesting is not a bad formula. It is testing on a list of companies that already won.

Educational background for the strategy builder. Nothing here is investment advice.

The bombers that came back

During the Second World War, the US military studied returning bombers to decide where to add armor. The damage clustered on the wings and fuselage, so the obvious answer was to armor those areas.

The statistician Abraham Wald, working with the Statistical Research Group at Columbia, pointed out the flaw. The military was only measuring planes that made it home. The undamaged areas on those survivors — engines, cockpit — were precisely the places where a hit meant the plane never returned to be measured. Armor belonged where the data showed no holes.

That is survivorship bias in one story: the sample you can see has been filtered by the very outcome you are trying to study.

What it looks like in market data

Take today's list of large US companies, pull fifteen years of price history for each, and test a strategy on it. The test looks legitimate. It is not, because that list was assembled after knowing who survived.

Every company that went bankrupt, got delisted, or was absorbed after a collapse is simply absent. Your strategy is never given the chance to buy them, so it never takes those losses. The sample is not a market — it is a market with the failures deleted.

Where it hides	What gets deleted
Backtesting on current index members	Every company removed from the index after falling apart
Mutual fund or hedge fund performance tables	Funds that closed or merged away after poor performance
“Traits of successful companies” studies	Firms with identical traits that failed anyway
Screens run on a current ticker universe	Names that would have been screened in, then gone to zero
Track records of visible traders	Everyone who blew up and stopped posting

How big is the distortion?

Large enough to change conclusions, and it varies with the universe and the period, so treat any single figure with suspicion.

Fund databases. A body of academic work from the 1990s onward, including studies by Brown, Goetzmann, Ibbotson and Ross, and by Elton, Gruber and Blake, found that removing dead funds from performance samples inflated measured returns by roughly half a percent to around one and a half percent per year, depending on the sample.
Individual stocks. The problem is often worse, because the failure is not merely missing — it is missing at its worst moment. Shumway's work on the delisting bias in CRSP data showed that returns around performance-related delistings were severely negative and frequently absent from the database, meaning naive tests skipped the loss entirely instead of recording it.
Compounding. A one-percent annual overstatement is not a rounding error. Over fifteen years it compounds to roughly 16% of extra phantom wealth, which is plenty to make a mediocre strategy look market-beating.

The relatives you should know

Survivorship bias rarely travels alone. Three close cousins do similar damage:

Delisting bias. The security vanishes from the data rather than being marked down to its real final value. The loss is not understated — it is absent.
Index reconstitution bias. Testing an index with its present membership credits it with holding today's winners during years when it actually held something else.
Look-ahead bias. The general case: using any information in a decision that was not knowable at the time. Survivorship bias is really look-ahead bias applied to the question of who still exists.

How to avoid it in your own testing

Use point-in-time data. A proper dataset records what was knowable on each historical date, including securities that later stopped trading, with final returns recorded rather than blank.
Reconstruct membership historically. Ask who was in the universe on the decision date, not who is in it now.
Record delisting returns explicitly. A dead position should be marked to its real outcome, not dropped from the portfolio.
Sanity-check against a known benchmark. If your universe's buy-and-hold return meaningfully exceeds the real index it resembles over the same window, survivorship bias is a prime suspect.
When you cannot fix it, say so. Survivorship-free data is expensive and often unavailable to individuals. Disclosing the limitation and treating results as an optimistic upper bound is honest; presenting them as an estimate is not.

Where StratPick stands on this

This applies to this site, so it is worth stating plainly rather than burying in a footnote.

StratPick's own universe carries survivorship bias. The dataset covers roughly 200 US stocks and ETFs that are listed today, with about fifteen years of daily split- and dividend-adjusted history from Yahoo Finance. Companies that delisted or failed during that window are not in it. Every backtest on this site should therefore be read as an upper bound on what the same rules would have achieved against the real historical universe.

That is a genuine limitation of a free educational tool, not a detail we would rather you missed. The Robustness League and the since-listed forward tracking exist partly to push back against it: a strategy that only looks good because the failures were deleted tends to decay once it is tracked forward on data nobody has seen yet.

Frequently asked questions

What is survivorship bias in investing?

Drawing conclusions from only the things that lasted. In practice it means testing on companies or funds that still exist today, which silently removes everything that went bankrupt, delisted, or was absorbed after collapsing. The remaining sample is healthier than the real universe ever was.

How much does survivorship bias inflate backtest returns?

It varies with the universe and period. Studies of mutual fund databases have generally found biases on the order of roughly half a percent to one and a half percent per year. For individual stock universes the distortion can be larger, because delisting losses are often missing from the data entirely rather than recorded.

What is delisting bias?

A specific form where a security disappears from the dataset when it stops trading and the final loss is never recorded. Research on CRSP data found that missing delisting returns meaningfully biased results, and that performance-related delistings warranted a substantial assumed loss rather than a blank.

Does survivorship bias affect index funds?

Not a fund you actually hold, since it owns losers all the way down until removal. The bias appears when someone backtests an index using its current membership list, crediting it with owning today's winners during years it did not.

How do I avoid survivorship bias in my own testing?

Use point-in-time data that includes later-delisted securities with their final returns. Reconstruct universe membership as of each decision date. Where that data is unavailable, state the limitation and treat results as an upper bound.