Survivorship Bias

Survivorship bias is the systematic overstatement of historical performance that results from studying only the entities that survived to the present. A backtest run on today's index members, or a fund study run on funds still open today, has silently conditioned on survival, and survival is correlated with good returns. The universe you test on has been curated by the very outcome you are trying to measure, so the result is biased upward before a single trade is simulated.

Where the names go

Companies leave an index or a database for reasons that are overwhelmingly bad news: bankruptcy, insolvency, delisting for failing listing standards, or being acquired at a distressed price. Mutual funds close after a run of poor performance. If your data vendor only carries currently-listed securities, all of those failures have been erased from history. Backtesting a "buy the S&P 500" rule on the current 500 members over the last 20 years never holds Lehman Brothers, Enron, Bear Stearns, Washington Mutual, or the hundreds of smaller names that went to zero, because they are not in today's list. The strategy is implicitly "buy the companies that we now know did not fail," which is not a strategy anyone could have traded.

The magnitude

The bias is large enough to reverse conclusions. Classic estimates:

• Equity-index survivorship inflates measured returns by roughly 1–4% per year, depending on universe and period; small-cap and emerging-market universes are worse because failure rates are higher.
• Mutual-fund studies (Elton–Gruber–Blake; Brown et al.) find survivorship adds on the order of 0.5–1.5% per year to average fund returns, enough to make a losing average-fund look like a winner.
• For individual delisted equities, the terminal delisting return is severe: bankruptcies frequently realize $-30\%$ to $-100\%$ in the delisting month. Databases that simply drop the security (recording a missing value rather than the loss) omit exactly the worst return in the name's life.

Formally, if $\mu$ is the true mean return of the full universe and $\mu_S$ is the mean over survivors, the bias is

$$\text{Bias} = \mathbb{E}[R \mid \text{survive}] - \mathbb{E}[R] = \mu_S - \mu \;>\; 0,$$

because $\Pr(\text{survive})$ increases with realized return, survival and return are positively correlated by construction. The same conditioning inflates the Sharpe Ratio (the losers that would have added left-tail volatility are gone) and understates Max Drawdown.

Worked example

Take a universe of 1,000 small-cap stocks. Over ten years, 200 go bankrupt with an average total return of $-90\%$; the surviving 800 average $+120\%$. The true equal-weighted decade return is

$$0.8 \times 120\% + 0.2 \times (-90\%) = 96\% - 18\% = 78\%.$$

A survivorship-biased backtest on the 800 survivors reports $120\%$, a $42$-percentage-point overstatement, roughly $3.6\%$ per year of phantom return, and it entirely hides the tail risk that a real investor bore. Worse, momentum and quality signals that appear to work in the survivor sample may owe their edge to implicitly avoiding the delisted names, an edge that vanishes when the failures are restored.

How to fix it

• Point-in-time universes. Reconstruct index membership as it was on each date, adding and removing names on their actual inclusion/exclusion dates. Trade only the securities investable at time $t$.
• Include delisted securities with delisting returns. Use a database (e.g. CRSP-style) that retains dead tickers and records the terminal delisting return, not a missing value. Never forward-fill or drop a bankrupt name, book the loss.
• Guard against ticker reuse. Delisted tickers get reassigned; join on a permanent security identifier, not the symbol, or you will splice a dead company onto a live one.
• Watch the funnel, not just the survivors. When studying strategies, track the entire cohort that entered the universe, including everything that later left.

This is closely tied to Look-Ahead Bias: using current membership is a form of using future knowledge (who survived) at a past decision point.

Failure modes

• Backfill bias, a cousin: a vendor adds a security to its database and backfills its history only for names that turned out successful (common in hedge-fund databases where funds report voluntarily after a good start).
• Self-selection in fund data: poor performers stop reporting, so the live database looks better than reality.
• Delisting recorded as missing rather than as a large negative return.
• Corporate-action splicing through ticker reuse.
• Assuming the bias is negligible in liquid large-caps, even the S&P 500 turns over roughly 20–30 names a year, and the churn is not random.

In interviews

A standard prompt: "You backtest a strategy on the current members of the Russell 2000 and get a Sharpe of 1.8, what is wrong?" The answer is survivorship bias: the universe conditions on survival, so both the mean return and the Sharpe are inflated and the drawdown is understated. Be ready to estimate the magnitude (1–4% per year for equities, more for small caps and emerging markets), to explain why the fix is a point-in-time universe with delisting returns booked as losses, and to connect it to look-ahead, choosing today's constituents is itself a use of future information. A crisp closing point: survivorship bias makes almost every naive long-only backtest look better than it truly was.