Most crypto research is built on a sample that excludes its own failures. Price databases drop coins when they stop trading — CoinGecko’s own documentation states that historical data for inactive listings is unavailable at any subscription tier — so a study assembled from what’s listed today is a study of survivors.
We rebuilt the sample with the failures left in: 1,707 assets that ever entered the top 300 by market capitalisation between 2015 and 2025, of which 549 no longer trade. 3.53 million daily observations. Membership is point-in-time, so an asset counts for the period it was actually large, not because it is large now.
Two things came out of it. One is a number. The other is backwards from what almost everyone assumes.
32% of the top 300 is gone
Of every asset that ever reached the top 300, roughly a third no longer trades. Not down heavily — gone: no price for six months or more, and a final print at or below 5% of its peak.
That count deliberately excludes migrations. Polygon moving from MATIC to POL is not a death; holders were made whole in a new ticker. So is Fantom becoming Sonic. Vendor “inactive” flags count both as deaths — 549 by that measure — and we don’t.
The calmest assets died the most
Group assets by realised volatility into four bands, and the twelve-month probability of disappearing looks like this:
| turbulence band | 12-month death rate |
|---|---|
| Low | 6.43% |
| Elevated | 3.40% |
| High | 2.66% |
| Extreme | 4.90% |
The safest-looking assets died at more than twice the rate of the calmest-but-one, and the relationship isn’t monotone in either direction — it’s U-shaped, with its minimum in the High band.
The obvious objection is that dying assets stop trading, so their prices go flat and their measured volatility collapses — an artefact, not a finding. We tested it. The Low band does contain far more stale quotes: 11.9% of its windows are more than half zero-return days, against roughly 1% in every other band. Removing those windows moves the Low band’s death rate from 7.24% to 6.43%. The inversion survives.
What actually separates them: turnover
Volatility tells you almost nothing about survival. Trading activity tells you a great deal, and the two are nearly independent — the rank correlation between turnover (30-day median volume ÷ market cap) and realised volatility is −0.14.
Twelve-month death rate, by both at once:
| thin turnover | low | mid | deep | |
|---|---|---|---|---|
| Low turbulence | 11.77% | 1.30% | 1.40% | 1.84% |
| Elevated | 5.33% | 1.63% | 1.13% | 1.24% |
| High | 5.53% | 1.31% | 1.34% | 0.96% |
| Extreme | 6.82% | 1.64% | 1.93% | 2.10% |
A 12.2× spread, and neither axis finds it alone. Turbulence on its own separates death rates 2.6×; turnover on its own, 4.8×.
The Low band was never one population. It was Bitcoin-class assets and abandoned tokens sharing a label. A quiet asset that people actively trade is among the safest things in the market. A quiet asset that nobody trades is the single most dangerous, at nearly one in eight disappearing within the year — and to a casual eye the two look identical.
Volatility does something else entirely
It sets the range of outcomes. The 10th-to-90th percentile band of 90-day forward returns:
| band | p10 | median | p90 | width |
|---|---|---|---|---|
| Low | −34.2% | +1.3% | +69.2% | 103 |
| Elevated | −50.0% | −9.8% | +79.5% | 130 |
| High | −59.3% | −17.2% | +102.9% | 162 |
| Extreme | −66.0% | −6.4% | +258.7% | 325 |
Sorted by turnover instead, that width barely moves: 186, 200, 153, 154.
So the two measures answer different questions. Volatility tells you how wide the range of outcomes is. Turnover tells you whether the asset survives to see them.
Worth noting against the usual framing: from Low to Extreme the upside opens 3.7× while the downside opens only 1.9×. High volatility in crypto has not historically meant mostly-downside. It has meant mostly-wider.
Method
Daily closes and market caps from CoinMarketCap, including inactive listings, 2015–2025. Universe is point-in-time top 300 by market cap, reconstructed from daily market caps rather than purchased as a ranking. Volatility is 30-day realised, annualised on 365 days. Turnover is 30-day median volume over market cap. Stablecoins are excluded behaviourally (≥90% of closes within ±2% of the median), not by name — 34 assets. Observations are sampled weekly per asset to reduce overlap. Death requires both 26 weeks without a price and a terminal value at or below 5% of peak. Forward-death windows that extend past the end of the panel are censored, not counted as survival.
What we are not claiming
This is not a prediction model. Thin turnover and death are partly the same event observed at different times — an asset that is dying stops trading. A twelve-month forward horizon separates them, but a proper lead–lag test is outstanding, and until it’s done these are descriptions rather than forecasts.
It is not investment advice, and nothing here says buy or sell anything.
Some contamination remains. Bad supply data produces phantom market caps in the source — one token reports a $126.7bn peak at a unit price of $0.000311 — which distorts both ranking and turnover. A hand review of the largest deaths is still needed to catch any remaining migrations.
Everything past the two pre-registered hypotheses is exploratory. More than twenty tests have been run against this dataset. At that count, roughly one in twenty will look significant by chance, and results not listed above should be treated accordingly.
Things we tested and found nothing in, recorded so nobody repeats them: critical slowing down as an early-warning signal (no effect at 90–360 days); category and sector breakdowns (vendor tags are themselves survivorship-biased — zero dead assets in our sample carry tags like “layer-1”); and behavioural clustering, which found one factor and no sectors at all. Crypto appears to have tiers, not kinds.
Filter Lab measures market structure and does not make recommendations. The underlying observation table — 487,829 rows — is available on request to anyone who wants to check this.