Survivorship Bias: The Backtest Error That Hides Failed Stocks

বাংলায় পড়ুন

Learning objective: recognise survivorship bias in stock research and backtests, build a historical universe that includes failures and interpret results with more realistic limits.

Today's stock list contains companies that survived until today. It excludes many firms that failed, were acquired, delisted or removed from an index. Testing a strategy only on current members allows the researcher to know the winners in advance. This is survivorship bias.

The bias usually makes historical returns look better and risks look smaller, because weak companies disappear from the sample.

1. Recreate the universe as it existed

A backtest should decide which securities were eligible on each historical date using information available then. If testing FTSE 100 or S&P 500 shares, use historical index membership rather than today's constituents throughout the entire period.

Point-in-time data should preserve additions, removals, ticker changes, mergers, bankruptcies and delistings. A company that later failed must remain in the data until its actual exit date.

2. Include delisting returns

The last quoted price is not always the investor's final result. A delisted share may receive cash in an acquisition, convert into another security or become nearly worthless after bankruptcy. Ignoring the final delisting return can materially understate losses.

Data vendors handle these events differently. Document how missing prices, suspended shares and corporate actions are treated. Never silently carry the last price forward forever.

3. Avoid look-ahead in fundamentals

Survivorship bias often appears with another error: using financial data before it was published. A year-end profit figure was not necessarily available on 31 December. The backtest must use the public release date and any later revisions should be handled carefully.

Historical databases can overwrite old figures with restated values. Point-in-time fundamentals preserve what the market could actually know.

4. Compare biased and unbiased results

If possible, run the test twice: once with today's survivors and once with the point-in-time universe. The gap estimates how much the selection error influenced performance.

Review more than annual return. Compare drawdown, turnover, loss frequency, sector exposure and the behaviour during stressed periods. A small return difference can still hide a major risk difference.

Worked example

Suppose a researcher tests a ten-year momentum rule on the 100 companies that belong to an index today. The test reports 12% annualised return and a 16% maximum drawdown.

A point-in-time reconstruction contains 168 companies that appeared in the index during the decade, including removed and failed firms. After adding delisting outcomes and realistic costs, the same rule produces 8% annualised return and a 27% drawdown.

The first result did not prove fraud or bad code. It answered the wrong question: “How would today's survivors have performed in the past?” The second result is closer to: “What could an investor have selected using the investable universe available at each date?”

These figures are hypothetical and illustrate the mechanism, not an expected return.

Common mistakes

• Testing historical rules only on today's index members.
• Deleting securities with incomplete records.
• Ignoring delisting proceeds or bankruptcy losses.
• Replacing missing prices with the last quote indefinitely.
• Using revised fundamentals before their original publication date.
• Assuming a familiar database is automatically point-in-time safe.

Practical checklist

1. How was the investable universe defined on every test date?
2. Are removed, acquired and failed securities included?
3. Does the dataset contain accurate delisting returns?
4. Are ticker and identifier changes mapped correctly?
5. Were fundamentals timestamped by public availability?
6. How are suspensions and missing prices handled?
7. Do costs and liquidity reflect the securities available then?
8. How different is the result when using today's survivors only?

The Money Man view

Survivorship bias is dangerous because the chart can look clean while the sample is wrong. We treat universe construction as part of the strategy, not a data-cleaning detail. A backtest is more credible when it includes the companies that disappeared and explains every assumption around their exit.

Educational information only. This tutorial is AI-assisted and published by The Money Man. It is not personalised investment advice, a recommendation or a promise of returns. Backtests are hypothetical and trading or investing can result in losses.