Who Really Sets the Bitcoin Price? We Measured It to the Millisecond

Correlation of two prices at the same moment only shows that they move together, not which one moves first. So we shift one series in time and ask how well the past of A predicts the future of B. For a single shift this is the same as correlation at a lag: R² is its square. On top of that, the forecast is tested on days the model has not seen, in both directions, and we also check what an exchange adds beyond the common leaders. The timing of first moves alone is unreliable: it depends on how an exchange packs its feed. And that is exactly the question: one coin trades on every major exchange at once, the prices move almost together, and what matters is which exchange moves first and how long the others take to catch up.

So for us "A leads B" means three things at once:

• the past of A predicts the future of B on days the model has not seen;
• the past of B predicts the future of A clearly worse;
• this holds not on average over the days, but day after day.

The set: the main perpetual future of Binance, KuCoin, Huobi, Gate, Bitget, Bybit and OKX on Bitcoin, Ether and Solana, days from 12 to 30 July 2026. We compare every exchange with every other within a coin. Seven exchanges give 21 pairs on Bitcoin and 21 on Ether; our Solana set has no OKX, so there it is six exchanges and 15 pairs. In all, 21 + 21 + 15 = 57 pairs.

How we counted, in short: we take the last 100 ms of one exchange's price and use them to predict the next 100 ms of another. The model is trained on yesterday and tested on today, a day it has not seen. Then we swap the exchanges and see which way the forecast is better.

Each exchange's price is sampled every 10 ms as the mid of its quote. We take time two ways: the moment the message reached us in Tokyo, and the exchange's own timestamp. We call a pair leading only if the result holds on both clocks. The grid is the same as in our article on latency arbitrage, except that there it was one pair of exchanges and here it is every pair.

Binance by its book; in the wave, by trades. The Binance Futures book arrives in batches every 25–28 ms and reflects Binance's own trades on average 14–15 ms late, while trades arrive one by one. For the forecast we take the book, as for everyone else: the trade price is noisy, bouncing between bid and ask, nobody can predict that noise, and pairs with Binance taken by trades would look more lopsided than they are. Trades remain a check. Hours when the Binance book stood without updates are excluded.

The wave. The second check works on events. An event is a price move of 2 bp or more within 10 ms on any of the exchanges; for each event we note which exchange covered half the move first and how many milliseconds the others took to cover half of it, within a 2 s window, on the Tokyo clock. Here the moment of the first move matters, so Binance is taken by trades, and every "X moved this many ms after Binance" also contains the book delay of exchange X itself (it is an upper bound). This is timing, not a link between series.

Let $r_F(t,\,t+h)$ be the follower's return over the next $h$, and $r_L(t-h,\,t)$ the leader's return over the previous $h$:

With the coefficients fitted on yesterday and tested on today, that gives one out-of-sample $R^2$ per pair, direction, horizon, clock and day, 16–18 test days per pair: the share of the follower's next move that the leader's past move explains on an unseen day. We run the same regression with the exchanges swapped, and the measure of leading is the difference by day:

There are seventeen horizons, from 10 ms to 120 s; the headline numbers are at $h = 100$ ms. Significance across days comes from the Wilcoxon signed-rank test with the Holm correction for the number of tests, and a bootstrap interval for the median.

At 100 ms, KuCoin's past explains 23.3% of Bybit's next 100 ms, while Bybit's past explains 1.0% of KuCoin's future. This pair has no special feeds: KuCoin and Bybit on Bitcoin, both taken by the quote mid. The difference is 22.4 pp, the 95% interval for the median is 21.0–23.1, and KuCoin wins on all 16 days. On the exchanges' clocks the gap is smaller, but the direction is the same: 10.0% against 2.4%. On Ether, in Tokyo, 25.6% against 0.8%.

In the chart, both lines for the KuCoin – Binance pair sit above zero. So does each lead the other? Leading is not either-or. Each exchange hints a little at the other: each has its own buyers and sellers who move the price first. What matters is whose hint is stronger, and here, at short horizons, the KuCoin → Binance line is about twice as high as the Binance → KuCoin one. For Binance – Bybit the Bybit → Binance line is almost at zero: there one side leads. Still, we don't name who is first between KuCoin and Binance: Binance's book arrives in batches every 25–28 ms, and with Binance taken by trades the picture flips (see the chart in the section on the first pair).

One pair could be luck. So here are all fifty-seven.

Leading holds beyond one pair and one day: at 100 ms the difference in $R^2$ between the two directions is significant (Wilcoxon with the Holm correction) in 55 pairs of 57 on the exchanges' clocks – 20 of 21 on Bitcoin, 21 of 21 on Ether and 14 of 15 on Solana – and in 54 on the Tokyo clock, and in 47 and 44 of them one side wins on every day.

We call a pair leading by a stricter rule: the past of A improves the forecast of B beyond B's own past and beyond the common leaders (Binance and KuCoin, on Solana Huobi too) on at least 80% of days, significantly after the Holm correction, and on both clocks. One way, 13 pairs of 57 passed it (3 on Bitcoin, 9 on Ether, 1 on Solana), and 5 more only on the exchanges' clocks: in Tokyo, delivery eats them. The other 39 passed the rule both ways. That doesn't mean "at the same time": a day holds about 8.6 million grid points, and with that many even a weak reverse influence is significant. So for those pairs we look at which side is ahead: in 28 of the 39 one side is three times stronger or more. On Bitcoin, for example, KuCoin's past adds 3.9 pp to OKX's forecast, OKX's past adds 0.19 pp to KuCoin's.

The days differ in how strong the link is, but almost never in its direction.

A check against a fake signal. We take the leader's past not from today but from another day, at the same time of day. It has no connection to today's market, so if the method is honest, the forecast must drop to zero. It does: the median absolute $R^2$ is below 0.001% on every pair and in both directions. So the link above is not a property of the price series themselves: it appears only when one exchange's past and the other's future come from the same day.

The second objection: Bybit's forecast already knows Bybit's own past and Binance's past, and perhaps KuCoin has nothing left to add. It does: KuCoin's past adds 4.2 pp to such a forecast in Tokyo and 2.5 pp on the exchanges' clocks, while Bybit's past adds 0.1 and 0.5 pp to KuCoin's (a version of the Granger test).

The reverse gain is not zero: the two exchanges carry some information about each other. The direction is proven by the difference in size.

On all three coins and on both clocks, KuCoin and Binance take the first two places. The order by the sum of leads at 100 ms, on the exchanges' clocks: on Bitcoin KuCoin, Binance, Bitget, Bybit, Huobi, Gate, OKX; on Ether KuCoin, Binance, Bybit, Bitget, Huobi, Gate, OKX; on Solana KuCoin, Binance, Huobi, Bybit, Bitget, Gate (no OKX in our Solana set). In Tokyo Bybit drops to sixth place; more on that below.

Both measures are skewed, in opposite directions. Binance's book arrives in batches every 25–28 ms and lags, which favours KuCoin. Its trade price arrives at once but bounces between bid and ask, and nobody can predict that noise, which favours Binance. So by the book KuCoin is ahead, by trades Binance, and we don't know where the truth lies in between. On Solana, at 100 ms, KuCoin is ahead by both measures.

On a common target they are almost equal. Bybit's future is the mid of its quote for both. In Tokyo, Binance's past explains 25.1% of Bybit's next 100 ms on Bitcoin, KuCoin's past 23.3%; on the exchanges' clocks, 8.8% against 10.0%.

That Binance shows Bybit's future, though, is visible on both clocks. By the strict rule this pair is two-way, but Binance is the side ahead: the other way, Bybit's past explains 1.4% of Binance's future in Tokyo and 3.3% on the exchanges' clocks, and on the exchanges' clocks, beyond KuCoin's past, Binance adds 1.6 pp to Bybit's forecast and Bybit adds 0.6 to Binance's, and Binance wins the $R^2$ difference on all 16 days. On Ether it is ahead twice over, on Solana three times.

In Tokyo, Bybit is second to last on Bitcoin and Ether and last on Solana. On the exchanges' own clocks it rises to third or fourth: about 33–35 ms of extra road to Tokyo were hiding it. Bybit itself leads OKX on Bitcoin and Ether on both clocks, and Gate on Solana; on Bitcoin and Ether it is ahead of Gate only on the exchanges' clocks, seven and eighteen times over, while in Tokyo Gate is ahead in those two pairs.

Gate and OKX trail on every clock (in Tokyo only Bybit sits below Gate, hidden by delivery): almost any exchange adds to their forecast, and they add tenths of a pp to anyone else's. Huobi is a follower: KuCoin and Binance lead it on Bitcoin and Ether, and beyond them its past adds almost nothing, 0.00–0.12 pp on Bitcoin on the exchanges' clocks; on Solana Huobi is third.

On Solana KuCoin comes first on both clocks and Binance second. KuCoin is twelve times ahead of Binance: on the exchanges' clocks its past adds 2.0 pp to Binance's forecast, and 0.17 the other way. In Tokyo KuCoin's past explains 12.7% of Bybit's next 100 ms, 10.4% of Gate's and 10.3% of Bitget's, against 0.3–0.5% the other way. Beyond that it looks like Bitcoin: Binance's past explains 13.1% of Bybit's next 100 ms against 0.7% the other way, and on the exchanges' clocks it is three times ahead. On a common target, Bybit, KuCoin and Binance are again almost equal, but in the direct pair KuCoin is twelve times ahead. Only one pair on Solana passed the rule one way: Bybit leads Gate.

With Binance taken by trades, Solana looks different: Binance's past explains only 2.3% at Bybit instead of 13.1%. Solana's trade price on Binance jumps between bid and ask by about 0.7 bp, and a large market order walks through several levels in separate trades; nobody can predict that noise.

In the wave, where Binance is taken by trades, the same jumps become events in their own right: with Binance by trades, Solana has 18% more events than with Binance by its book. Binance is first in 39% of Solana events, but the other exchanges' books don't follow its jumps: when Binance was first, the other Solana exchanges fail to catch up within 2 s in 11–19% of events; on Bitcoin (except Huobi), in 3–7%. From KuCoin the Solana wave runs like this: Huobi after 40 ms, Gate and Bitget after 80, Bybit after 100, Binance after 130 ms.

When Binance moved first on Bitcoin, KuCoin covers half the move after 20 ms, Bitget after 60, Gate after 70, Huobi after 80, Bybit after 90 and OKX after 110 ms. These are medians on our clock in Tokyo and an upper bound: they include the road from the exchange and the book delay of the exchange itself. On the exchanges' clocks the steps are shorter, most of all for Bybit (see above). Our article on latency arbitrage gives the same Binance–Bybit pair 70 ms on our clock and 30 ms on the exchanges' clocks, but that is a different measurement: the shift at which the link is strongest, with Binance taken by its book.

There are 11,035 events on Bitcoin over 17 days, 25,776 on Ether and 44,194 on Solana. Binance is first in 49% of events on Bitcoin, 53% on Ether and 39% on Solana; KuCoin in 25, 23 and 26%. Here Binance is taken by trades, which arrive before its book, so these shares don't settle whether it or KuCoin is earlier. When Binance moved first, Huobi doesn't catch up within 2 s in 30% of events: its quote updates least often of all.

On Ether the steps are the same: KuCoin after 20 ms, Bitget after 50, Gate and Huobi after 70, Bybit after 80 and OKX after 110 ms.

Moves that start somewhere other than Binance spread less well. Almost half of Bitcoin events (45%) are started by an exchange other than Binance, most often KuCoin. Gate, Bitget, Bybit and OKX catch up with them in almost the same median time as with Binance moves, but the tail is longer and they fail more often: in 8–12% of events against 3–5%. Binance itself fails to catch up within 2 s in 21% of such events.

Because $R^2$ is a share of a predictable move, not profit: between "predicts" and "earns" stand execution latency, the spread and fees. And 90 ms is time on our clock in Tokyo, an upper bound, and about 33–35 ms of it is the road of Bybit's messages; your window will be different.

What this looks like in money we measured on the same pair, in a different period and with Binance taken by its book, in our article on latency arbitrage. A simple rule, entering Bybit right after a Binance jump, won 88% of its trades and all 30 days of 30 before fees, 2.8 bp a trade on average, and lost money at every public Bybit fee: it breaks even while the fee is no more than 1.4 bp a side, and the whole public table sits above that.

Here is the conclusion we stand behind: who leads among crypto exchanges, and how fast a move reaches the rest, is proven. What leading is worth to you depends on your fee, on where you stand and which feed you read, and you will have to measure it on your own clock.

On Bitcoin and Ether, watch KuCoin and Binance together; on Solana, KuCoin above all; and only on the scale from 10 to 700 ms. This answer rests not on who flipped first and not on one chart, but on a forecast tested both ways, on days the model had not seen, and on two clocks.