Research · calibration v0
Does the survival rating carry information?
Partly. It sorts assets by the risk of a severe drawdown and by the chance of a later downgrade. It does not sort them by return, and below the top band it barely sorts them at all. Everything behind that sentence is on this page, including the parts that argue against the product.
Study generated 2026-09-19. 3,092 observations across 204 coins (46 skipped for insufficient history), spaced 30 days apart over 730 days of daily history. Each observation is scored with the production formula on the data available at that date — point in time, no lookahead. Measurements, not promises.
What holds, and what does not
What holds
Assets rated Active go on to see a severe drawdown less often than the rest: 16.6% of Active observations fell more than 30% within the next 30 days, against 20.1% for Weakening, 22.3% for Distressed and 22.6% for Dormant.
Active also holds its rating better than Weakening: it was revised down within 90 days in 31.5% of cases against 46.9%. Sorting by the risk of a bad outcome is the whole of the credit-rating claim, and on that claim the top band earns its place.
What does not
The three lower bands do not separate from one another. Median 90-day returns run −12.7%, −14.1% and −12.9% — a spread narrower than the noise in samples this size, and not in rating order.
The score carries essentially no information about 30-day returns. Its rank correlation with the next 30 days is 0.0004. Not weak — absent. Anyone using the score to rank assets by their next month of returns is using it for something it cannot do, and we do not present it that way.
Read the limitations before the tables
The first one below is the largest. Every band on this page looks better than it should, because the universe is the top 250 by market cap today: assets that fell out of that list over the two years are simply absent from the sample. The effect is strongest exactly where the rating claims most — the low bands, whose worst outcomes are the ones most likely to have dropped out.
- Universe is today's top 250 by market cap: coins that fell out of it are absent, which flatters every band (survivorship).
- All-time high is the running maximum inside the 730-day window, not the true peak; drawdown is understated for coins that peaked earlier.
- Observations are 30 days apart and overlap in their 90-day windows; counts overstate independent evidence.
- Regimes are split only by the sign of Bitcoin's trailing 30-day return.
- Outcomes are price returns; rank survival and delisting are not measured in v0.
How to read the columns
- n — observations in the bucket. Because observations are 30 days apart and their 90-day windows overlap, n overstates how much independent evidence there is.
- Median 30d / 90d return — the middle price return over the following 30 and 90 days. A median, not an average, so one recovery cannot carry a bucket.
- Positive at 90d — share of observations whose price was higher 90 days later.
- 30d drawdown worse than −30% — share that fell more than 30% from the observation price at some point in the following 30 days. This is the column the rating is built to sort.
- Lower / higher band at 90d — share whose rating was in a worse or better band 90 days later. A rating already at the top cannot be revised up and one already at the bottom cannot be revised down; those cells are struck out rather than printed as zero.
By band at observation
The headline table. Across all 3,092 observations the median 30-day return was −3.3% and the median 90-day return −8.5%, so every band should be read against a market that was falling on average.
| Band at observation | Median 30d return | Median 90d return | Positive at 90d | 30d drawdown worse than −30% | Lower band at 90d | Higher band at 90d |
|---|---|---|---|---|---|---|
| Activen 1,239 | −0.9% | −1.9% | 41.3% | 16.6% | 31.5% | — |
| Weakeningn 657 | −7.5% | −12.7% | 34.7% | 20.1% | 46.9% | 20.1% |
| Distressedn 1,041 | −4.3% | −14.1% | 35.4% | 22.3% | 15.1% | 15.6% |
| Dormantn 155 | −2.4% | −12.9% | 34.2% | 22.6% | — | 13.5% |
The drawdown column is ordered as the rating intends. The return columns are not: Distressed and Dormant medians sit between Weakening's, which is the clearest single piece of evidence that the lower three bands are not distinguishing between assets. The band-revision columns are not comparable across rows either — Weakening and Distressed have bands both above and below them, Active and Dormant do not.
By market regime
Observations are split by the sign of Bitcoin’s trailing 30-day return at the observation date — a crude split, and the only one v0 makes. It matters more than the rating does.
Bitcoin rising over the prior 30 days
| Band · Bitcoin rising | Median 30d return | Median 90d return | Positive at 90d | 30d drawdown worse than −30% | Lower band at 90d | Higher band at 90d |
|---|---|---|---|---|---|---|
| Activen 595 | −0.8% | −0.4% | 44.7% | 20.0% | 29.6% | — |
| Weakeningn 282 | −5.3% | −4.8% | 45.7% | 15.2% | 39.0% | 27.0% |
| Distressedn 450 | −10.6% | −17.3% | 34.0% | 31.6% | 19.1% | 20.7% |
| Dormantn 58 | −25.6% | −36.4% | 10.3% | 53.4% | — | 10.3% |
In a rising market the bands do sort, and steeply: severe drawdowns run from 20.0% for Active to 53.4% for Dormant, on 58 Dormant observations.
Bitcoin falling over the prior 30 days
| Band · Bitcoin falling | Median 30d return | Median 90d return | Positive at 90d | 30d drawdown worse than −30% | Lower band at 90d | Higher band at 90d |
|---|---|---|---|---|---|---|
| Activen 644 | −1.1% | −3.7% | 38.2% | 13.5% | 33.2% | — |
| Weakeningn 375 | −10.9% | −18.3% | 26.4% | 23.7% | 52.8% | 14.9% |
| Distressedn 591 | −1.7% | −12.2% | 36.4% | 15.2% | 12.0% | 11.7% |
| Dormantn 97 | 12.3% | −0.8% | 48.5% | 4.1% | — | 15.5% |
And in a falling market the order breaks. Dormant returned a median 12.3% over the next 30 days — the best figure in the table — and saw a severe drawdown in 4.1% of cases, the lowest. On 97 observations that is thin evidence, but it is not noise we can wave away: an asset that has already fallen a long way and trades little has less left to give up when the market turns down. The rating does not know that, and we are not going to pretend the table says otherwise.
By score decile
The bands are thresholds laid over a continuous score. Looking at the raw score in tens shows where the information actually sits.
| Score at observation | Median 30d return | Median 90d return | Positive at 90d | 30d drawdown worse than −30% | Lower band at 90d | Higher band at 90d |
|---|---|---|---|---|---|---|
| 0–9n 11 | 1.7% | 18.0% | 81.8% | 9.1% | — | 9.1% |
| 10–19n 378 | −2.8% | −15.5% | 33.3% | 24.9% | 29.9% | 8.2% |
| 20–29n 510 | −5.4% | −14.6% | 34.7% | 23.5% | 7.6% | 13.7% |
| 30–39n 574 | −5.4% | −11.9% | 35.2% | 17.9% | 30.7% | 19.7% |
| 40–49n 380 | −8.4% | −13.7% | 35.5% | 21.3% | 36.1% | 26.3% |
| 50–59n 302 | −5.4% | −10.3% | 36.8% | 19.5% | 57.6% | — |
| 60–69n 266 | −7.6% | −11.8% | 29.7% | 27.8% | 42.5% | — |
| 70–79n 241 | −0.4% | −2.9% | 38.6% | 16.6% | 23.2% | — |
| 80–89n 207 | −0.0% | −0.1% | 44.4% | 8.2% | 15.0% | — |
| 90–99n 223 | 0.3% | 0.8% | 61.4% | 7.2% | 7.2% | — |
The relationship is not monotonic. It appears only at the top: the 90–99 decile is positive at 90 days in 61.4% of cases with severe drawdowns at 7.2%, and the two deciles below it are the next best. Across the bottom six deciles the outcomes are interchangeable. The 0–9 decile has n 11 and its flattering figures should be ignored entirely — it is a handful of observations on a handful of coins.
By volume-decay deduction
One factor on its own: the points deducted for the fall in trading volume against the busiest sustained stretch in the window. It is worth 25 of the 100 points in the score.
| Volume-decay deduction | Median 30d return | Median 90d return | Positive at 90d | 30d drawdown worse than −30% | Lower band at 90d | Higher band at 90d |
|---|---|---|---|---|---|---|
| no decayn 160 | 0.3% | 0.8% | 65.0% | 9.4% | 13.1% | 0.0% |
| decay 1–9 ptsn 217 | −0.8% | −0.9% | 43.8% | 10.6% | 23.0% | 0.0% |
| decay 10–19 ptsn 924 | −4.8% | −9.2% | 37.4% | 19.6% | 34.2% | 8.1% |
| decay ≥20 ptsn 1,791 | −4.8% | −12.2% | 34.4% | 21.6% | 26.1% | 13.4% |
This single input orders the outcomes more cleanly than the composite score does — positive at 90 days falls from 65.0% at "no decay" to 34.4% at "decay ≥20 pts", with severe drawdowns rising in step. That is a finding against the current weights, not for them, and it is the first thing a v1 of the formula has to answer. Note also the distribution: 1,791 of 3,092 observations sit in the worst bucket, so the factor rarely discriminates in practice.
Rank correlation
Spearman rank correlation between the score at observation and what followed. A value of 0 means the score orders assets no better than chance; 1 would mean it orders them perfectly.
| Score against | Spearman ρ | Reading |
|---|---|---|
| 30-day return | 0.0004 | No relationship. The score is not a return signal over a month, and nothing on this site offers it as one. |
| 90-day return | 0.0517 | Near zero. Directionally right, far too weak to act on. |
| Worst 30-day drawdown | 0.1755 | The strongest of the three, and still modest. A higher score goes with a shallower worst fall. This is the one relationship the rating can claim. |
A correlation of 0.175 is small in absolute terms. It is presented here as what it is: a weak but consistent ordering on downside risk, measured on a sample that flatters it.
What we are doing about it
- The thresholds between bands are held to these tables. Any change to a weight or a cut-off is recorded in the study with the evidence that prompted it.
- v1 of the study has to fix the universe. A point-in-time top 250 — including assets that later left it — is the single change that would most alter these numbers, and almost certainly downward.
- Outcomes in v0 are price returns only. Rank survival, delisting and liquidity withdrawal are the outcomes a survival rating should actually be measured against, and they are not measured yet.
- Volume decay outperforming the composite is a live question about the weights, not a footnote. It is open.
The formula these tables test
Every input, weight and known blind spot is published on the methodology page, which also carries a short form of the table above.
CoinGraph is for information only and is not investment advice. Ratings are opinions formed from public market data, not statements of fact about any project, team or asset. The figures on this page describe what followed past observations in one sample over one two-year window; they are not a forecast, and no band is a claim about what any particular asset will do next. Market data is provided by CoinGecko.