Survivorship bias

By Jarosław Wasiński • Published: Apr 2, 2026 • Updated: Oct 11, 2026

Distortion from assessing only elements that remain available while omitting those that disappeared earlier. Examples include testing current instrument membership, ranking active accounts and narratives about successful traders. Define the cohort, analytical unit, period and treatment of closed or missing observations in advance. A missing record does not automatically mean a loss, but excluding it can alter conclusions. Success statistics need the denominator appropriate to their stated question; survivors’ results do not automatically describe everyone who attempted it.

Example. One hundred hypothetical accounts start with equal capital and no later cash flows. Seventy remain at period end with an average +10% return; thirty closed earlier with an average −40% and do not trade afterwards. The full-cohort mean is (70 × 10% + 30 × (−40%))/100 = −5%. Studying survivors alone gives +10% and answers a different question.

Define the cohort before observing outcomes

Specify which accounts, systems or instruments were eligible initially and how later entrants are handled. Disappearances can reflect losses, a firm closing, a name change or removal from a platform. A missing record does not automatically imply a particular loss, but cannot simply be discarded without assessing its effect. The example uses an arithmetic mean with explicitly equal initial weights.

Bias in data and narratives

Backtesting today’s index members omits instruments delisted earlier. A ranking of active copy-trading providers can exclude closed profiles. Interviews with successful traders are not representative statistics for all participants. See Brown, Goetzmann, Ibbotson and Ross: Survivorship Bias in Performance Studies on selection; a particular study does not establish a universal Forex loss rate.

Reduce the distortion

Use historical membership and identifiers that include closed elements. Record entry and exit dates and actual final settlement. For unknown outcomes, describe a range of possible effects rather than invent a result. Also check backfill: adding an attractive pre-submission history is a different selection mechanism. Chronological OOS cannot repair a database from which unsuccessful elements have already disappeared.

Match the denominator to the question

An account-count-weighted mean is different from a capital-weighted mean. Active participants’ results do not automatically describe new entrants. Cohorts with different starting dates also have different exposure periods. Each analysis can be meaningful when its scope is clear; the error is changing the question after seeing an attractive result. State whether the unit is a person, account, programme or individual attempt.

Assess published statistics

Ask about denominator, success criterion, period and missing observations. A result can legitimately concern active profiles, but cannot be generalised to the full cohort without justification. In the course: strategy testing and Forex myths and facts. Source checked on 11 October 2026.

FAQ

Did every missing account lose?

No. Establish the facts; missing data need explicit uncertainty treatment rather than automatic classification.

Does a large sample remove this bias?

No. Thousands of selected survivors can still answer a different question from outcomes for every member of the defined cohort.

← Back to the full glossary