A ranking may show accounts still active, while a course advertisement features selected graduates. Before judging success, identify the starting group and who is absent. Survivorship bias arises when conclusions about the whole group rely on cases that remained observable, omitting those that disappeared earlier.

What survivorship bias is, and where it comes from

Survival-based exclusion, selection of the best days from an ongoing account and self-selection into a survey are different mechanisms, though they can coexist. Do not assume every absent participant lost: an account can close or a person stop publishing for other reasons. An unknown result is neither zero nor a loss. Ask whether missingness is related to the outcome being measured.

Brown, Goetzmann, Ibbotson and Ross’s 1992 paper Survivorship Bias in Performance Studies examines volatility and returns in a sample truncated by survival. Its examples show how selection can create apparent predictability. This concerns performance measurement, not a statistic about today’s forex traders or a reconstruction of an aircraft story.

Where survivorship bias hides in trading

Advertisements, rankings and public posts can select cases differently from the population you want to evaluate. Check the inclusion criteria, period, costs, cash flows and open exposure. No evidence here quantifies how often an algorithm suppresses losses, why an author stops posting or whether every successful account is marketing.

For a historical equity strategy, using only today’s index constituents can omit instruments available earlier. The eligible universe should match the rule at each decision time. A study of one predefined currency pair does not automatically require adding every currency that ever existed. Choosing a pair, period or parameter after seeing results is a related selection/overfitting problem.

Own hypothetical cohort: visible versus complete records
20 visible accounts16 positive and 4 negative after costs: 16/20 = 80% positive.
80 omitted accounts, known in the full register14 positive and 66 negative: absence is not uniformly a loss.
All 100 original accounts30 positive and 70 negative: 30/100 = 30% positive.
Difference50 percentage points; genuine visible results can misrepresent the whole cohort.

Missing results? Report bounds

In a hypothetical predefined group of 20 people, 6 results are positive and 6 non-positive. The other 8 are unknown. The observed subset is 50% positive, but the full-group share can only be bounded from 30% to 70%. There is no need to assume missing people lost or earned zero.

Scroll sideways to see the full diagram

Known results and full-group positive-share boundsTwenty circles represent six positive, six non-positive and eight unknown results. The full-group minimum is 6 of 20, or 30%; the maximum is 14 of 20, or 70%.Full group: 20 people6 positive6 non-positive8 unknown0%50%100%Minimum: 6/20 = 30%Maximum: (6 + 8)/20 = 70%
The shaded range follows solely from known counts. It is neither a confidence interval nor a forecast.

The lowest possible share occurs if all unknown results are non-positive: 6/20. The highest occurs if all are positive: 14/20. The share 6/12 = 50% describes only the subset with available results. Narrower bounds require more data, not a more confident tone.

Assumptions: Counts are fictional. Each of the 20 people has one result over the same specified period, assessed by the same positivity criterion. Exactly 8 results are unknown; no missingness mechanism or distribution is assumed.

How it distorts your expectations and decisions

Fix the population, observation period and outcome criterion before evaluating examples. In the own 100-account cohort, the public subset is 80% positive but the full register is 30%. Without that full register, do not assign the missing 80 results arbitrarily. The diagram gives possible-completion bounds for a different fictional group; those bounds are not a confidence interval.

Own probability model: 16 fictional people each make four independent trials, with success probability 1/2 for everyone and independence across people. A person chosen beforehand has probability (1/2)⁴ = 1/16 = 6.25% of four successes. Across 16 people, the probability that nobody does is (15/16)¹⁶ ≈ 35.61%, so at least one does with probability ≈ 64.39%. A genuine four-for-four winner can therefore be selected without differences in skill. Under these assumptions that person’s next chance remains 1/2. This is not an estimated forex model or proof that all real methods lack an advantage; the psychology section offers related review topics.

Account-loss shares and trade win rates differ

A base rate concerns a defined population and period. ESMA’s announcement of 27 March 2018 cited EU national-authority CFD analyses in which 74–89% of retail accounts typically lost money. That historical range is not a current worldwide figure or personal probability. Its product-intervention framework included provider-specific risk warnings; check the entity, product, current applicable rules and calculation period when reading one today.

A 98% trade win rate cannot be compared directly with the share of loss-making accounts: the units and denominators differ. Own no-cost example: 98 trades gain 1 unit each and two lose 100 each. The trade win rate is 98%, yet the total is 98 −200 =−102 units. An account warning therefore cannot by itself prove that a high-win-rate profile is an outlier, lucky or dishonest.

Different measurements require different questions
Trade win rateWinning trades divided by assessed trades; does not include the size of wins and losses.
Provider account-loss shareAssessed loss-making CFD accounts divided by the covered accounts over the prescribed period.
ESMA historical range74–89% in cited EU retail-CFD analyses in the 2018 announcement, not a 2026 global measurement.
Control questionAre population, period, units, costs and missing-case treatment comparable?

A concrete defence against survivorship bias

Seek a group defined before results, a common period and explicit missing-data treatment. Collecting only spectacular failures is another selected sample, not the missing denominator. Anecdotes can suggest questions but cannot establish a success probability or a universally appropriate early risk level.

Use separate evaluation data that match the strategy’s eligible universe and preserve unsuccessful configurations. Bailey and colleagues’ 2014 paper discusses how testing multiple variants can generate an attractive backtest. Repeatedly checking the same holdout weakens its separation; a holdout does not repair an incomplete universe. Verification of one account’s data also does not explain why it was selected, even if the history is long.

What to do this week — first steps

Review one ranking or advertisement using five fields: starting group, inclusion rule, period, outcome definition and number of missing observations. Mark unknown answers explicitly. Ask whether costs, cash flows and open positions are handled consistently; do not predict that most of ten chosen profiles will fail or treat five failure stories as a representative sample.

State a conclusion that fits the evidence: “this account was positive in this period” differs from “this method makes most participants profitable.” Confirming a displayed record does not settle selection, costs or risk. You can assess evidence without copying any trades or estimating unknown people’s results.

Related reading: overconfidence bias examines confidence versus evidence; herd mentality discusses social influence; realistic goals relates targets to their assumptions; process over outcome separates rule adherence from a single result.