How CoinSpectrum ranks exchanges
173 centralized exchanges, one 0-100 score, and every number behind it stated. What is measured, how it is weighted, what is checked on chain, and what gets a venue demoted.
Start from what a venue cannot type
Almost every figure an exchange ranking runs on is supplied by the exchange. Volume is a number it reports. The order book is served by its own matching engine. Even a proof-of-reserves page is, at first, a page. Rank exchanges on those alone and the ranking measures a willingness to type large numbers.
Two things cannot be typed. The first is traffic: the Tranco list counts how often the internet's resolvers ask for a domain, and no exchange can enter a value into it. The second is time: a venue founded in 2011 has survived every cycle since, and no venue founded in 2024 can claim that. Those two signals decide half the score. Everything else corroborates.
Nine signals, weights that sum to 100
Each signal scores 0 to 5. Its points are the signal divided by five, times its weight.
| Signal | Weight | How it is scored |
|---|---|---|
| Web rank | ×20 | Tranco. Rank 1,000 scores 5; every tenfold worse rank costs 1.5. A domain outside the global top million scores the floor rather than being skipped, because that absence is itself a reading of traffic. |
| Age | ×10 | Years since founding, 0.35 per year, reaching 5 at about fourteen years. A twenty-year-old venue is not twice as trustworthy as a fourteen-year-old one; the scale plateaus. |
| Liquidity size · transparency | ×10 | Dollars resting on the thinner side of the book within 1% of mid — a large order eats one side — summed across the venue's ten largest markets and read from its own books: $50K scores 1.5, $500K 2.5, $5M 3.5, $50M 4.5, $160M the full 5. Where those markets were not read directly, the BTC book's thinner side within 0.1% of mid stands in, scaled ×10, and a sum known only through CoinGecko never scores above 2.5. Capped by how much of the book anyone can read: the whole book or ±1%+ keeps the full signal, a 100-level book caps at 4, too shallow at 3, depth known only through CoinGecko at 2.5, no book 0. Transparency is a ceiling, never a bonus. |
| Reliable volume | ×10 | The reported volume discounted by the reliability ratio: at 10× the top-ten norm, a tenth of the claim is scored. $10K scores 0, $1B scores 5. A claim with nothing behind it scores 0. |
| Reliability ratio | ×10 | The ratio itself: claimed volume against the venue's own traffic, 1× the top-ten norm. At or under 1× the full 5, falling log-linearly to 0 at 25×, where the venue is also flagged. A claim with nothing behind it scores 0, and so loses twice — the volume it inflated and the ratio it broke. |
| Liquidity / volume | ×10 | Liquidity standing near mid on the venue's ten largest markets as a share of their daily volume, against the top-ten norm. At or above it 5, each halving below costs a point, a thirty-second of the norm scores 0. |
| Proof of reserves | ×10 | Eight outcomes, from an open audit on chain down to a report of a page nobody can see, which scores 0. Detailed below. |
| KYC | ×10 | Verification only above a limit scores 3, none at all 2, verification required 1, on the 0-5 scale — KYC decides at most 6 of its 10 points. Required is the industry default and not a finding against the venue; a venue that asks nothing of anyone sits under one that asks above a threshold. |
| CoinGecko trust | ×10 | The aggregator's 0-10 exchange grade, halved. A second opinion on liquidity and traffic, not ours; it also stays a contradiction at 6/10 or lower. Skipped where the feed returns none. |
Web rank, liquidity size, volume and age are scored continuously on log scales rather than in fixed bands, so a book of $9M genuinely outscores one of $3M and a venue at Tranco #6,000 outscores one at #21,000, instead of both pairs tying at a band ceiling. The weighted sum is normalized over the signals that could actually be measured.
Missing evidence is not evidence, with two exceptions
Where a figure is genuinely absent — a book the venue's API refused to serve, a founding year nobody recorded, a trust grade the aggregator stopped returning — the signal is skipped and the score normalizes over the rest. A venue is never marked down for what could not be checked.
Two absences are readings, not gaps. A Tranco rank outside the top million says the domain has no traffic to speak of, and scores the floor. And a book that was read and found empty within 0.1% of mid is a measurement of an empty book: it scores 1, the continuous scale's own floor, while a venue that publishes no book at all scores 0. Hiding the book must never tie with showing an honest empty one.
Reserves: the open audit
A proof-of-reserves page is a claim. Published addresses turn it into something anyone can verify at any time, and "anyone" includes this site. Where a venue publishes its reserve addresses in a machine-readable form, every build reads them on chain — Bitcoin, Litecoin, Dash, Ethereum, the ERC-20 dollar tokens and the pegged tokens on BNB Smart Chain — and compares what they hold with what the venue says it owes its users. A coin is judged only when the chains read here carry nearly all of what the venue lists for it; the rest is reported as partly readable and decides nothing. Binance and OKX publish a monthly file of every wallet address, so their Bitcoin and Litecoin are read the same way, against liabilities counted at the snapshot; a gap inside a stated tolerance is movement since then, not a shortfall.
If every checkable asset covers its liabilities, the venue takes the top of the reserves scale. That sits above a signed third-party attestation on purpose: with an attestation you still have to trust the signer; with published addresses you do not have to trust anyone. The venue's own "we hold X" figure is only a snapshot and is allowed to go stale; where it exceeds the chain, the chain figure is shown beside it and used. Only a material shortfall — the chain holding less than the venue says it owes — counts against a venue: the floor of the scale, and a wash-suspect flag, because it is the venue's own numbers against public evidence. Privacy coins and rows listed without an address cannot be checked and never decide the outcome either way.
The full scale, top to bottom: open audit on chain; a signed attestation, level with addresses published but not readable in a given build; a verified page; nothing known either way; a page the chain contradicts, level with publishing nothing at all; and, at zero, a report that a page exists with no page to show — an unverifiable claim of reserves is the shape a fabricated one takes. AnonKYC was the first venue to reach the top tier, Binance and OKX followed with their monthly address files (every open audit, in full): its Bitcoin address holds exactly what its file says it owes, and the file's larger holdings figure is shown beside the chain number as stale.
The reliability ratio: liquidity against Tranco
A venue's claimed volume is measured against its own traffic, never against anyone's absolute size. A small exchange with small honest numbers sits at the same ratio as Binance. What the ratio catches is a venue reporting more turnover than the people visiting it could produce.
Every build takes the ten best venues by this score that trip no wire and measures what they turn over per site visit — claimed monthly volume divided by an independent monthly-visit estimate. That gives a median and a band, the range honest venues occupy. Across every clean venue with a Tranco rank, the build also fits how volume falls off with rank. Each exchange is then measured on both: its own dollars per visit against the median, and its own volume against what its rank predicts. The smaller of the two ratios is taken, so a single broken traffic estimate cannot convict a venue — a Philippine exchange whose visit count arrived as a few hundred a month reads as thousands of times the norm on visits and under ten on rank, and it is the ten that counts.
Inside the band nothing happens. From five times the median a venue loses ten points. From twenty-five times it is flagged like any other contradiction. From a hundred times it is flagged twice. A large claim with no traffic estimate and no Tranco rank behind it at all counts as unbounded. The low side never demotes: traffic far above what the volume implies is brokerage-style flow or under-reporting, not fraud. The reference figures are published in the data file, and each row's tooltip states its ratio in one line.
Liquidity against volume: could the volume have happened?
The reliability ratio compares a venue's claim with its traffic. Liquidity capacity compares it with its own order books. Every build takes each exchange's ten largest markets and reads each pair's book — from the venue's own API wherever the registry can name the pair, through CoinGecko only where a venue refuses automated reads — and sets the pair's daily turnover against the liquidity standing near mid. An honest market turns that over tens to a few hundred times a day. A washed market turns it over thousands of times, because the trades were never matched against standing orders; and the signature is one small pair, not the whole venue. Thresholds are calibrated on the ten best venues by score, the venue's figure is the share of its top-ten volume in pairs beyond them, and the worst pair is named in the tooltip. Half of the volume beyond the book, read from the venue's own API, is a contradiction — a share known only through CoinGecko lowers the signal but never flags —. The same reads also give the venue's liquidity as a share of its volume across those markets — ten dollars traded on one of liquidity is 10%, a hundred on one is 1%, weighted by where the volume trades so a deep BTC book cannot cover for a washed listing — which against the top ten's median is the liquidity / volume signal: at or above the norm the full five points, one off per halving below it, so the smaller the share the less the venue scores. Depth is the venue's own data, so none of this can promote past what was measured — and transparency about the book sits inside the book signal as a ceiling, not a bonus: the venue that lets anyone read its whole book can score the full signal, the venue that shows a hundred levels cannot, and the venue whose depth reaches us only through a third party is capped at the middle.
Contradictions demote
On top of the composite, each of the following costs a venue 15 points and drops it beneath every venue that trips one fewer. They accumulate. All but the last three apply only to venues claiming more than $50M a day; a small exchange reporting small honest numbers contradicts nothing.
- Claims high volume from a domain with no measurable Tranco traffic.
- Claims a top-ten volume position while drawing a fraction of the traffic every comparable venue draws.
- Publishes no public order book at all, so nobody outside the venue can check the volume.
- Reports a day of turnover thousands of times larger than what rests on its book.
- Has book depth that is really one or two large orders rather than a spread of many.
- Carries a CoinGecko trust grade of 6/10 or lower.
- Has more than half its reported volume discounted by CoinPaprika's adjusted-volume model.
- Publishes reserve addresses that hold less than it says it owes, checked on chain.
- Runs a reliability ratio of 25× or more.
- Has half or more of its top-ten volume in pairs that could not have executed against the book they show.
Being unable to read a venue's book is deliberately not on that list. Several exchanges refuse automated requests outright; that is a limitation on this side, not a fact about them.
A row, worked
Kraken, on the build this was written against: web rank #2,173 scores 4.49 of 5, which is 18 points of 20. Founded in 2011, age scores the full 10. $167M resting within 1% of mid across its ten largest markets, on a venue that serves its whole book, scores 4.72, for 9.4. $1.8B of volume at a reliability ratio of 0.7× is scored in full, for 10, and the ratio itself, inside the band, for another 10. Its liquidity, 5.9% of its volume, sits above the top-ten norm, for 10. A signed reserves attestation scores 4.5, for 9. KYC required scores 1, for 2. CoinGecko's 10/10 scores 5, for 10. That is 88.4 of 100 measured, no contradictions, and a final score of 88, an A.
GroveX, on the same build: a domain at Tranco #464,585, a claimed $943M a day, no public order book, a reliability ratio of 441×. Web rank scores the floor, liquidity size scores 0 for a book nobody can read, the ratio scores 0 and the volume is scored at a 441st of its claim. The composite normalizes to 34; five contradictions subtract 75; the row sits at zero, at the tail, and the tooltip says why in five clauses.
What the score is not
It is not a guarantee of safety. A venue can be widely visited and still lose its customers' money; a venue can be small and honest. What the score says is how much of a real business exists behind the ticker, and whether the numbers the venue reports about itself agree with the numbers the world reports about it. Those are the questions a ranking can answer.
The Score column is off by default on the CEX leaf and can be turned on in profile settings as a number, a letter grade or a filled bar. Every row's tooltip breaks its number down signal by signal, and the methodology states every weight and threshold in one place. The reasons are written down so a reader can disagree with them directly.
Weights and thresholds are current as of the build of 14 September 2026 and move only when a stated editorial reason requires. The worked figures move with every six-hourly refresh. The argument does not.