This article describes a bias that affects our own experiments as much as anyone else's. The audit numbers quoted are measurements of our own equity pipeline, published because readers deserve to see the knife cut both ways. Nothing here is investment advice. Decisions are yours.
Ask one question of any crypto backtest: how many of the coins in its universe are dead? If the answer is zero, the test was run on the survivors, and its results are inflated by every collapse it never had to trade through. The answer is almost always zero, because the data most people can get only remembers the living.
How the graveyard disappears
The mechanism is mundane, which is why it is everywhere. Exchange APIs enumerate the pairs that trade today; ask a major venue for its market list and you get the current roster, and ask for the history of a pair delisted in 2022 and you typically get nothing. A developer assembling a backtest universe therefore starts from current listings, reaches backward for history, and produces a dataset that is clean, complete-looking, and quietly missing every asset that failed hard enough to be removed. No fraud is required. The path of least resistance does all the work, and the result is a universe composed entirely of things that, by definition, did not go to zero.
We measured it on ourselves first
Before building this site, we audited the equity research stack of our sibling operation for exactly this bias. The finding was blunt: in a five-year, 6,134-symbol US equity database assembled the industry-standard way from a broker's current listings, exactly zero symbols had their final price bar before 2026. Not one death in five years, in an asset class where a meaningful percentage of small caps delist annually. The universe was survivor-only, and every backtest built on it inherited the flattery.
We then bounded the damage rather than hand-waving it. Assume missing names would have taken candidate slots in proportion to an annual delisting rate, earned the average losing trade's return, and displaced real winners. Under a conservative 4% annual delisting rate, the flagship backtest's 75x multiple bounds down to about 30x. At 8%, closer to reality for small caps, it bounds to 15x. At a pessimistic 12%, to 9x. The true figure lives somewhere inside that band, and the honest lesson is not the exact number but the direction and magnitude: this one data flaw can plausibly cost a headline result most of itself. We published that against our own product. It is the cheapest credibility we ever bought, and it dictated how this site's data layer works.
Why crypto's version is worse
Equities die by merger, acquisition, or slow delisting, and a fair share of those events are neutral or profitable for holders. Crypto deaths are violent and total. Terra's LUNA went from roughly $40 billion of market value to effectively nothing in one week of May 2022, and its original pairs were delisted or renamed across the industry; a dataset built from today's listings cannot see the defining risk event of its own sample window. FTX's exchange token followed within months. Beneath the famous corpses lie thousands of 2021-vintage projects that pumped, bled more than 99%, and were removed from every list a backtester would start from. Academic and industry censuses of dead coins count them in the thousands per cycle; the exact figure matters less than the fact that your dataset contains approximately none of them.
Momentum strategies take the worst of it
Here is the part that makes this site's genre especially exposed. A momentum scanner is a machine for finding the fastest horses, and the coins that later died were disproportionately the ones that once printed spectacular breakouts on spectacular volume, which is precisely the signature momentum systems rank highest. Deleting the graveyard does not remove a random slice of history; it surgically removes the trades most likely to have ended at a stop, a liquidation cascade, or a zero. A momentum backtest on survivors is therefore not mildly optimistic. It is optimistic exactly where the strategy is most dangerous, in the seductive names that were about to stop existing. Our own founding experiment, documented in the transplant write-up, lost money on a survivor-biased universe, which is why we describe even that honest failure as an upper bound.
The vendor gate
The cure is data that remembers the dead, and vendors vary enormously here, so we test before we trust. Our procurement rule is a graveyard gate: before any data source feeds research, it must produce full daily history for a basket of famous corpses, LUNA classic through its final week, FTT through its collapse, and a handful of randomly chosen 2021 mid-caps that later died, and it must be able to enumerate which assets existed and how large they were AS OF any historical date, not just today. A vendor that fails either half is disqualified for research use regardless of price, because every downstream number would inherit the flattery. The same logic gives this site two published universes: the live monitor's current top 100, appropriate for describing today's market, and the research universe, reconstructed point-in-time with the dead included, appropriate for testing anything. Conflating those two universes is the original sin this article exists to prevent, and the split is documented in the monitor methodology.
The one-question test for readers
You do not need our infrastructure to use this. Whenever anyone shows you a crypto backtest, ask what happens in their data when a coin dies. Three answers are possible. "The dataset is current listings" means the number is an upper bound produced on a rigged sample, and you should mentally apply the kind of haircut our bound table implies. "We use point-in-time data including delisted assets, from this named source" is the rare right answer, and it earns the rest of your attention. And no answer at all is also an answer. A vendor whose track record depends on a dataset they cannot describe has told you which kind of operation they run, a theme continued in seven concrete checks. The graveyard is silent, but it is not hidden. You just have to ask. Decisions are yours.