The Coinbase Premium is one of the most-cited numbers in crypto and one of the least defined. Commentary says "the premium has gone negative, institutions are selling" and rarely says what is being divided by what.
The metric is not bad. It describes something real. It just describes something other than what most readings claim. This piece covers what the Coinbase Premium measures, what it does not, and the six checks it needs before it goes anywhere near a backtest.
What it is
The Coinbase Premium is the price gap for bitcoin between two venues: Coinbase, where it trades against dollars, and Binance, where it trades against USDT, Tether's dollar token.
The arithmetic is simple:
(Coinbase price − Binance price) divided by the Binance price, times 100.
Expressed as a percentage that is the Coinbase Premium Index. The same difference in dollars is called the Coinbase Premium Gap. They say the same thing, except the dollar version scales with price level: one percent at $100,000 is twice as many dollars as one percent at $50,000.
The data provider CryptoQuant popularised the metric and publishes it continuously. Coinglass and others maintain their own versions, sometimes against different reference venues, which is why exact values differ by source while the direction usually agrees.
One thing CryptoQuant deserves credit for: the definition is published, venues and pairs included. You can look up what is being calculated. That is more than many widely used metrics offer, and it is the precondition for everything else in this piece.
What it does tell you
Two venues, two order books, two pools of buyers. Coinbase is mostly people wiring dollars from an American bank. Binance is mostly people holding USDT who never touch a US bank at all.
When buying lands harder on one side, that price lifts. Anyone wanting to collect the difference buys on the cheap venue and sells on the expensive one, squeezing the gap shut again. So the premium is the trace of an imbalance in buying pressure, caught before the arbitrage closes it.
That these traces are real rather than noise is well documented. Igor Makarov and Antoinette Schoar, in "Trading and Arbitrage in Cryptocurrency Markets" (Journal of Financial Economics, February 2020), show that cross-exchange price deviations in crypto are large, recurrent, and persist for days and weeks. Deviations are far larger across countries than within one, because capital controls hold up the money that would otherwise close them.
So the defensible statement is: the premium describes where trading is happening. Not where price is going.
What it does not tell you
Six places where the standard readings break.
1. It is a price, not a quantity
A price gap says nothing about how much traded. The same premium can come from one large order or from a thin book where a medium order already bites.
Anyone converting a premium into "institutions bought X BTC" has invented a number. Quantity information lives in volume, not in price.
2. The denominator is not a dollar
This is the most important weakness and the least often mentioned.
Binance quotes in USDT, not dollars. While USDT sits at a dollar this makes no difference. When USDT slips below, the bitcoin price expressed in USDT rises mechanically and the measured Coinbase premium falls — with nobody in the United States having done anything.
This is not only a crisis phenomenon. In June 2023 USDT dropped to roughly $0.977 because its share of a large swap pool had ballooned past 70% against a balanced target near a third. Nothing had changed about the reserves; a lopsided pool simply prices the overweight side down.
In windows like that, the Coinbase premium is measuring the stablecoin, not American demand. Without controlling for it you have a metric with two causes and no way to say which one is firing.
3. The direction of causality is not settled
Makarov and Schoar show something else too: deviations widen during periods of large bitcoin appreciation.
Which makes the premium at least partly a consequence of the move rather than a warning of it. That is exactly the setup in which a regression on forward returns finds something that is not there: measure the past carelessly enough and it looks like prediction.
Anyone wanting to treat the premium as leading has to answer that question first, not last.
4. "Institutional" is an attribution, not a measurement
What is observable is two order books. What is not observable is who stands behind them. An order book does not carry a customer category.
On top of that, large buyers typically avoid the public book entirely and settle through over-the-counter desks — precisely so as not to move the price. What shows up in the premium is, by construction, the portion of trading that did move the price. That is not the same as the bulk of trading.
5. The label outlived the venue
The most common definition refers to the "Coinbase Pro price (USD pair)". Coinbase Pro was retired on 9 November 2023 and folded into Coinbase Advanced Trade.
This is not a data error — it is the same order book under a new name. But it says something about how much maintenance definitions receive once a metric is established. If you adopt someone else's metric, know which venue and which pair it actually computes against, not which one the label says.
6. We found no independent test
For this piece we searched for peer-reviewed work specifically testing whether the Coinbase premium predicts forward returns. We found none.
That is not proof that none exists — it is the state of our search. But it means the metric currently rests almost entirely on vendor documentation and chart reading. For a number that appears in market commentary daily, that is a thin foundation.
The surrounding literature is not encouraging either. Work that systematically evaluates bitcoin return predictability repeatedly finds that active strategies trail simple holding. Anyone introducing a new metric carries the burden of proof against that finding, not the other way round.
The contrast case: the Korea premium
There is a second metric built on exactly the same formula that lives in an entirely different order of magnitude. It explains, in retrospect, why the Coinbase premium is so small.
The Korea premium, colloquially the kimchi premium, compares the bitcoin price on Korean venues such as Upbit and Bithumb, which trade in won, against the global price. Same arithmetic, different denominator.
The scale is not comparable. Researchers at the University of Calgary put the peak at 54.48% in January 2018, with an average of 4.73% between January 2016 and February 2018. The Coinbase premium, by contrast, normally hovers around the zero line. Two metrics with an identical formula, and one reaches levels the other never sees.
The reason is market structure, not sentiment. This is where Makarov and Schoar's finding becomes visible: deviations across countries are far larger than within a single currency area, because capital controls hold up the money that would close the gap. In Korea that means foreigners cannot readily trade on Korean venues, Korean traders cannot readily move money out, and bank accounts must run under a verified real name. Anyone trying to collect the premium hits a wall.
Between the United States and the global USDT market there is no such wall. Which is why the gap closes there in minutes, and why the Coinbase premium is mostly noise.
That is the real lesson. A small Coinbase premium is not a sign of weak American demand. It is a sign that arbitrage works. Reading "US demand is collapsing" out of a value near zero means interpreting a number smaller than the trading fee on either side.
And the denominator problem recurs in Korea. In summer 2019 the premium flipped negative — bitcoin was suddenly cheaper in won than in dollars. The Korean daily Dong-a Ilbo attributed this to the won having lost roughly 3% against the dollar since the end of July. A metric read as a thermometer for Korean demand had changed sign because a currency moved. Same error as USDT inside the Coinbase premium, with a national currency instead of a stablecoin.
Even the headline number is disputed. For that same January 2018, credible sources carry 54.48%, "about 55%", and 47% side by side — depending on which global reference basket and which exchange rate. So the rule holds here too: without a stated reference, the number cannot be rebuilt.
And the regime question applies equally. Ki-Young Ju, chief executive of CryptoQuant, has said Korea accounted for roughly 7.9% of global crypto trading volume in 2017 and under 2% by 2021. A metric whose market share falls to a quarter is no longer measuring what it once measured, however unchanged the formula.
What it is actually good for
None of this makes the metric worthless. It makes it useful for something other than what is usually claimed.
As a context variable, not an entry signal. The question the premium answers is which side of the world the buying pressure is on. That is a regime question, not a timing question. Regime questions belong in filters and overlays — in the decision whether a strategy is allowed to run, not in the decision when it enters.
And once it is a filter, it faces the same chain of tests as any other.
Six checks before it goes into a backtest
1. Compute it yourself. Vendor index series may have been revised or backfilled. Building the premium from both venues' raw prices means you know what is in it. Taking a finished series means you do not.
2. Align the clocks. Both prices must be pulled to the same timestamp. Computing a 12:00:05 Coinbase price against a 12:00:00 Binance price measures latency and data delivery as much as demand.
3. Multi-day averages, not ticks. At short horizons the premium is extremely noisy, because arbitrage keeps collapsing it. Even the vendors suggest reading it smoothed. But the smoothing window is then a free parameter you have to test, or you have quietly fitted it to the past.
4. Control for the USDT price. The decisive cross-check. Compute the premium additionally against a genuine dollar pair on a third venue, or adjust the Binance price by the USDT rate. If your result disappears, you were measuring the stablecoin.
5. Charge real costs. The premium normally lives in basis points. Fees and the spread between bid and ask consume it at the same order of magnitude. A backtest without realistic costs is not optimistic here, it is void.
6. Rule out look-ahead mechanically. The prefix test: compute the metric series once on all the data and once on a truncated leading segment. If old values change, future information is leaking in. That is not a judgment call, it is a diff of two files.
And the two gates that decide it
Sample size. Under 30 trades a result is an anecdote, not an edge. A filter that only engages in five market phases has five observations, however many candles sit in between.
Benchmark in the same window. The question is never whether the filtered strategy made money, but whether it made more than simply holding over exactly the same period. If the fair comparison does not support the filter, the filter goes — not the benchmark.
One regime warning
The Coinbase premium built its reputation in 2020 and 2021, when American institutions and high-net-worth individuals were the decisive marginal buyer and their route into the market ran through the Coinbase order book.
Since spot bitcoin ETFs launched, a substantial share of American demand routes through the funds' authorised participants and through over-the-counter desks. How far that has shifted the transmission path is an empirical question we do not answer here. But it is precisely the question to ask before carrying a metric from one market structure into another.
A metric is tied to a mechanism. Change the mechanism and the old test does not carry over — it has to be run again.
What holds up
The Coinbase premium is a serviceable context variable and a poor signal.
What it reliably describes: that buying pressure is unevenly distributed across two order books, and that the imbalance can persist for days or weeks. What it does not describe: how much money moved, who moved it, whether the dollar reference even holds, and where price goes next.
The distance between "describes where trading happens" and "says where price goes" is not a matter of interpretation. It is one backtest with a benchmark.
This piece describes a market metric and how it is calculated. It is not investment advice, a recommendation, or a forecast.
Sources: Igor Makarov and Antoinette Schoar, "Trading and Arbitrage in Cryptocurrency Markets", Journal of Financial Economics 135(2), February 2020, pp. 293–319 · CryptoQuant, user documentation for the Coinbase Premium metric and the "Bitcoin: Coinbase Premium Index" chart page, accessed 22 July 2026 · Coinbase, "Hello Advanced Trade, goodbye Coinbase Pro", company blog, announced June 2022, migration completed 20 November 2023 · Kaiko commentary on the June 2023 USDT deviation (pool imbalance) · University of Calgary, research on the Korea premium, April 2019 (peak of 54.48% in January 2018, average of 4.73% January 2016 to February 2018) · Dong-a Ilbo, 5 August 2019 report on the disappearance of the Korea premium, as relayed by CoinDesk · Ki-Young Ju (CryptoQuant) on Korea's share of global trading volume, CoinDesk, April 2021.