Definition
Survivorship bias arises when a dataset or backtest includes only securities, funds or companies that still exist today and omits those that were delisted, merged, liquidated or closed. Because failures disappear from the sample, average results look better and risks look smaller than they were.
How it works
To avoid it, a test of all listed stocks should use the list of companies that actually traded on each historical date, including those later delisted, along with their final returns. The same issue affects fund databases, where closed or merged funds drop out and flatter the remaining average, and index studies that apply today's constituents to the past.
The bias is largest for small companies, risky assets and long horizons, because more members fail over time. It also hides drawdown: a strategy tested only on the winners never experiences the holdings that went to zero.
Illustrative example
Illustrative: a fictional universe has 10 equal-weighted stocks at the start of a period. Six survive and rise 50% each, while four are delisted after falling 80% each. Survivors alone average +50%, but the full starting universe averaged (6 × 50% + 4 × (−80%)) ÷ 10 = −2.0%.
What to watch for
Survivorship-free data usually requires paid or carefully maintained sources that include delisted securities and their delisting returns. Free price histories of current tickers are a typical cause of this bias, and the overstatement is hard to correct afterwards.