Survivorship Bias: What a Winning Track Record Leaves Out

Every track record you are shown has been through a filter before it reaches you. Not necessarily a dishonest one: a strategy list, a fund table, a leaderboard of copyable traders and a page of funded-account results are all populations that something removed entries from before publication.

Survivorship bias is the name for what that removal does to the average. The version explained on most pages is about delisted stocks, which is a real problem and almost never the reader’s problem. What follows is the version that applies to a record someone is showing you now.

What Gets Removed Before You See the Record

A sample is what remains after every process that could remove something from it has finished. When those processes remove members for a reason connected to performance, the average of what is left is higher than the average of what existed.

That is the whole mechanism. It is not a bias in anyone’s judgement and needs no intent. A signal service that quietly stops publishing its worst month, a copy-trading page that lists only accounts still open, a strategy vendor whose forum thread ends when the strategy stopped working: each leaves a record whose remaining entries are the ones that lasted.

The distinguishing feature is that the filter is invisible in the output. A drawdown figure, a return figure and a trade count all describe entries that are present. Nothing in the table describes the entries that were removed, and nothing in the arithmetic can recover them.

This is why the fault cannot be fixed downstream. A more careful calculation on a filtered population produces a more precise description of the survivors. The correction has to happen at the point where the sample is assembled, which is usually a point you have no access to.

So the useful question about any published record is not how good the numbers are. It is: what was the rule that decided which entries appear here, and could a poor result have failed that rule?

Why the Delisted-Stock Version Barely Applies to Currencies

Almost every explanation of survivorship bias is an equities one. A stock index drops constituents that went bankrupt, were acquired or fell below a threshold. A database of current index members therefore contains no failures, and a strategy tested on it inherits an advantage that was never available.

That story is correct and it transfers badly to a retail forex or contracts-for-difference account. Major and minor currency pairs do not go bankrupt and are not removed from a broker’s symbol list for having performed poorly. A pair that a broker stops quoting is normally a liquidity, regulatory or entity decision rather than a performance one.

What does transfer is selection in the records rather than in the price data. The strategies, signal sellers, copy-trading leaders and funded traders that remain visible are the ones that survived long enough to still be visible. That population is filtered on exactly the thing you are trying to measure.

The practical consequence is that a forex reader should stop looking for the bias in the chart and start looking for it in the list. Price history for a major pair is usually complete. The list of people whose results you are comparing almost never is.

Survivorship Bias and Overfitting Are Different Faults

These two are routinely listed as neighbouring biases, and treating them as siblings hides that they break in different places and are repaired by different means.

Overfitting is a defect in how a rule was fitted. The data was complete, and the rule was tuned until it described that particular sample closely enough to describe noise as well as signal.

The remedy is methodological: reserve data the rule never touched and test on that. Splitting in-sample from out-of-sample data is covered in our guide to how a backtest is set up, and this page does not repeat it.

Survivorship bias is a defect in what the sample contained before any fitting happened. It is present in the data the moment you receive it, and it is not reduced by holding some of that data back. Reserving a portion of a filtered population gives you a clean test on a population that was already wrong.

Where It Enters a Strategy Someone Sold You

A strategy offered with a record attached carries two separate selection steps, and only the first is obvious.

The first is inside the strategy. If the rules were arrived at by trying many variants and keeping the one that tested best, the surviving variant is the winner of a contest whose losers are not shown. Its result describes the best outcome of that search rather than the expected outcome of the rule.

The second is above the strategy. The vendor also produced other strategies. The ones that failed early were withdrawn, renamed or never marketed, so the catalogue you are browsing is the surviving subset of a larger set. Both steps push the visible average up, and neither leaves a trace in the figures.

This is where a published drawdown number stops meaning what it appears to mean. It is the worst decline of a record that was kept, not the worst decline of the approach. How to read such a figure, what it does and does not describe, and how it interacts with recovery is handled in our page on reading a drawdown figure in someone else track record; this page states only the selection half.

The same reasoning applies to record length. A short record is not merely less reliable, it is more likely to have been selected for, because a filter that removes poor performers removes them early. Our page on how long a record has to be covers the statistical side of that question.

The Funded-Account Population You Cannot See

Funded-account programmes make the mechanism unusually literal. Participants pay to attempt an evaluation, a proportion pass, and the traders who then appear as funded are by construction the subset that passed.

Any statistic drawn from that visible group therefore describes the survivors of a selection step, not the population that entered it. That remains true whether the evaluation is easy or hard, and it is a structural feature of the arrangement rather than a criticism of it.

Two further filters usually sit on top. Accounts that breach a rule after funding stop appearing, so the visible group thins again over time in a performance-related way. And the results promoted publicly are chosen from among those still active, which is a third selection on the same axis.

None of that says a programme is unsound. It says the published outcomes cannot be read as the outcome of participating, because the denominator, everyone who paid to try, is not in the figure.

How these challenges are structured is covered on our page about how a funded account challenge works. The same reasoning applies to a copy-trading leaderboard, where following another trader sits alongside a ranking that lists only accounts still open.

What the Past-Performance Rules Actually Require

There is a regulatory answer to the cherry-picked window, and it is more specific than the general warnings that appear at the foot of marketing pages. In the United Kingdom the FCA Handbook sets conditions that a firm must satisfy whenever a communication contains an indication of past performance.

The condition that bears directly on selection is one of coverage. The figures have to reach back five full years, and where the investment has been on sale or the service running for less time than that, they have to span its entire life instead. Every stretch reported has to be a whole 12-month period, never a part-year cut to taste.

A window chosen because it looked good is exactly what a complete-12-month-period requirement is designed to prevent.

The remaining conditions in the same rule are worth knowing because their absence is informative. The indication must not be the most prominent feature of the communication, and the reference period and the source of the information must be clearly stated.

A prominent warning has to sit alongside the numbers, saying plainly that they describe what already happened and that a past record does not reliably indicate future results. If the figures are in a currency other than sterling, that currency must be stated with a warning about currency movement. And where the numbers are quoted gross, the firm has to disclose what commissions, fees and any other charges take out of them.

Simulated past performance, which is what a backtest produces, is treated separately and anchored to something real. It has to rest on the genuine record of one or more investments or indices that either match the investment being promoted, come substantially close to matching it, or sit beneath it.

Most of the conditions above then bind that genuine record too, and a prominent warning must label the figures as simulated and deny them any reliable bearing on future performance.

A further rule provides that an indication of future performance must not be based on or refer to simulated past performance at all.

These rules bind firms communicating with clients, not an individual posting results. Their value to a reader is as a checklist. When a record arrives without a stated reference period, without a stated source, in incomplete periods, or gross of costs, you are looking at something a regulated firm could not have presented that way.

The Figures Every Page States and No Source Supports

Research for this article covered six pages on the topic. Between them they state a strategy failure rate, a count of collapsed funds, a compound annual return, two Sharpe figures, a minimum trade count, spread and slippage ranges, a challenge fee and a profit split.

Not one of those figures is attributed to a source that can be checked. Several are attributed to other pages on the same site. One page states more than thirty numeric claims and cites nothing external, while funnelling the reader into its own funded-account product. Another reports a return and a ratio from a single backtest the author ran.

That pattern is not incidental to the subject, it is an instance of it. A percentage describing how many traders fail is a statistic about a population whose members mostly stopped being observable, which is the exact measurement problem the page is supposedly explaining.

So no such figure appears here. Where a number would ordinarily go, this page names the mechanism instead, and where a figure was needed it came from the rulebook cited in the sources note below rather than from another article.

Who This Page Is Not For

If you are testing a rule on the complete price history of a major currency pair, the classic survivorship problem is largely absent from your data and your attention is better spent on how the test itself is constructed.

If your question is whether a specific vendor is honest, this page cannot answer it. The mechanism described here operates identically whether a record was filtered deliberately or by ordinary attrition, and it offers no way to tell those apart from the outside.

And if you are looking for a threshold, a minimum number of trades or a length of record that makes a track record trustworthy, no such threshold follows from any of this. A minimum sample size can be derived for one strategy from the scatter of its own results, as the page on expectancy works through, but that arithmetic assumes the sample was never filtered to begin with. A selected sample does not become representative by being larger.

Frequently Asked Questions

What is survivorship bias in trading?

It is the distortion that appears when the records you can see have already had poor performers removed from them. Because the removal is related to performance, the average of what remains is higher than the average of everything that existed, and nothing inside the visible figures reveals the gap.

Does survivorship bias affect forex the way it affects stocks?

Not in the same place. The equities version comes from delisted and acquired companies missing from an index database, and currency pairs do not delist that way. In forex the selection happens in the records: strategies, signal services, copied traders and funded accounts that are still visible are the ones that lasted.

Is survivorship bias the same as overfitting?

No. Overfitting is a fault in how a rule was fitted to a sample and is addressed by testing on data the rule never saw. Survivorship bias is a fault in what the sample contained before any rule was fitted, so reserving part of that sample does not help.

What do the past-performance rules require a figure to show?

Under the FCA Handbook a firm showing past performance has to reach back five full years, or across the entire life of the service where it has run for less, and every stretch it reports must be a whole 12-month period. It must also state the reference period and the source, attach a prominent warning to the figures, and spell out what charges take out of them wherever the numbers are quoted gross.

How can a reader tell whether a track record is a filtered sample?

By asking what rule decided which entries appear, rather than by examining the entries. If a poor result could have caused an entry to be withdrawn, renamed, closed or simply left off the page, then the record is a sample of survivors and its average describes them alone.