The Exchange Balance Is Just the Flows Added Up

When bitcoin rose 33.9% between its 30 June low and 21 August, one explanation travelled further than the rest. Large holders had been quietly buying through the fall. They moved their coins off exchanges into private storage. That left fewer coins available to sell, which put a floor under the price, and the rise followed.

It is a good story, and it has the shape of something checkable. Exchanges are watched. Coins moving in and out of them are counted daily. If the story is right, then the counting should show it, and it should show it early.

That last word is what makes this worth a day of work. Something worth watching has to move before the thing it is meant to warn about. This site publishes a score built from thirteen such measures. If money leaving exchanges reliably came before a rise, it would belong on that list as a fourteenth, and anyone reading it would have had days of notice ahead of a 33.9% move. If it does not, then the story is a description of the rise, written after the fact, in the same words either way.

What the Repo Can and Cannot Check

The version of the claim that gets quoted most often puts a figure on it: the share of all bitcoin held by large wallets went from about 27% to about 29.5%. That figure cannot be checked here, and it is worth saying why rather than pretending otherwise.

Splitting supply by the size of the wallet holding it is a paid product. The free Coin Metrics tier this site draws on carries 31 daily measures for bitcoin and not one of them is a supply-by-wallet-size series. A direct request for one is refused outright. So the number circulates precisely because checking it costs money, and most people repeating it have not checked it either. What follows tests the claim, not that figure.

What the free tier does carry is the pair this article needs. The first is the count of coins moving onto exchanges each day and the count moving off. The second is the total sitting on exchanges. Both come with a warning that belongs in every sentence about them: the balance is an estimate built from addresses the vendor believes belong to exchanges. It is a labelled guess, not a fact about custody, and one of this article's findings is about exactly how much weight that guess can carry.

The Coins Did Not Leave

Start with the simplest reading. On 1 June, exchanges held 2,639,145 coins by this measure, and bitcoin traded at $71,329. The price then fell, to $58,525 on 30 June. The exchange balance did not fall with it. It rose, and kept rising, for ten weeks, to a peak of 2,732,484 coins on 15 August.

That is 93,338 more coins on exchanges, a rise of 3.54%, across the whole drawdown. The story needs float to leave. It arrived.

Say only that. Coins arriving on an exchange are not proof of anyone selling. There is a plain reason for caution this month. A defensive migration has been under way since 30 July, after a flaw was disclosed in one wallet maker's firmware, and a coin moved somewhere safer looks identical in this data to a coin moved somewhere to sell. Why the coins came is not in the file. That they did not leave is.

Now the ten days everybody quoted. Net flow below is coins in minus coins out, so a positive number means coins arriving on exchanges.

Coins came in on all five days to 15 August, 10,836 of them, while the price sat between $62,925 and $63,547. Over the six days after, the net ran the other way by 35,982 coins, after the low and on the way up. The biggest outflow day of the run, 20 August, is also one of the biggest up days. That is the wrong order for a cause.

The last row is the one to sit with. On 21 August the price rose 7.24%, the largest single day of the whole window. The net flow that day was plus 58 coins: in, not out, and near enough to nothing on a day when 44,593 coins arrived and 44,535 left. Whatever produced the biggest move here, it was not visible in this number.

Ten days prove nothing on their own. They are the illustration. The test comes next.

The Balance Is Just the Flows Added Up

Before testing whether flow leads price, it is worth asking what these two series are. There is an arithmetic rule tying them together. If the exchange balance changed only because coins moved, then each day's change in the balance must equal coins in minus coins out. Nothing else can move it.

Across 5,596 days of paired readings, the two match almost perfectly: the correlation is 0.9966, and the typical gap between them is zero. Not small. Zero. On 5,441 of the 5,596 days the two agree to the last digit the file carries, which is one hundred-millionth of a coin.

That is worth stating plainly, because it changes what the balance is for. For most of bitcoin's traded history the exchange balance was not an independent measurement standing beside the flows. It was the flows, added up. Charting it next to them and treating the two as separate evidence was always double counting.

Then it stopped. Sort the days where the two disagree by more than one coin and they are not scattered across fifteen years. Before 10 April 2026 there are seven such days out of 5,464. From 10 April 2026 to the end of the file there are 132 out of 132: every single day, without a break, with a typical gap of 485 coins.

Two dated facts sit beside that, and this article states both and joins neither. Coin Metrics announced on 21 April 2026 that it had recomputed the full history of these exchange measures. The per-date stamps on the recomputed rows cluster across early April. What caused the break is not established here, and a date that fits is not a cause.

One check does rule out the obvious alternative, which is that this site's own plumbing broke. The rows before 24 May and the rows after it reached this repo through two different routes, an outside mirror and then the vendor's own service after a fix landed on 22 August. The daily disagreement runs straight across that seam, unchanged. It is in the data, not in the pipe.

The Day the Balance Moved More Than the Coins

The rule above also sets a hard ceiling. On any day, the balance can rise by at most the number of coins that arrived, which happens if every one of them stays. It can fall by at most the number of coins that left, which happens if none arrived and stayed. A day outside that range cannot be coins moving, whatever else it is.

In 5,596 days there is exactly one.

On 13 August 2026 the exchange balance rose by 46,954 coins. That day 21,360 coins arrived and 18,168 left, a net of 3,192. Even if every arriving coin had stayed and none had left, the balance could have risen by 21,360 and no more. So 25,594 coins of that day's rise are not coins moving anywhere. They are addresses being relabelled as exchange addresses, or the vendor's estimate being corrected, or both. The full gap between the balance change and the flows that day is 43,763 coins.

This is the same class of error an earlier piece here found in two of the dashboard's own indicators: a number moving because the accounting moved, not because the economics did. The lesson repeats. Build the test on the flows, which are counts of coins, and treat the balance as a cross-check at best. One day in 5,596 is rare. It landed eight days before the window everyone was reading, and the part of it that cannot be coins is larger than any single day's net flow in the table above.

Does the Flow Lead the Price?

Now the test the article exists for. Take the daily net flow and the daily price change, and measure how closely they track each other. Do it not once but sixty-one times: with the flow read thirty days early, twenty-nine days early, and so on through to thirty days late. If coins leaving exchanges lead the price, the strongest link should sit on the early side. The whole set is reported below rather than the best one, because picking the best of sixty-one is how a coin flip gets published as a finding.

There is nothing there. Every one of the sixty-one readings is under 0.06, on a scale where 1.0 is a perfect match and 0 is none. The strongest is same-day, at -0.0569, which is statistically firm on 5,596 days and explains 0.32% of daily price movement. The best of the early readings is eight days ahead at -0.0228, which on this much data is indistinguishable from chance. Ranking the days instead of using their raw sizes moves nothing: the largest reading anywhere in the grid falls to 0.0454.

So the flow and the price move together on the day, faintly, and the flow knows nothing about tomorrow. That is the answer to the question, and the rest of this section is about the ways it could have been faked.

The first is the oldest trap in this business. Compare the level of the exchange balance with the level of the price and the reading is 0.82 in logs, which looks enormous: two thirds of the variation shared. It is worthless. Both numbers rose for fifteen years, so any two of them would agree. Compare the daily changes instead, which is what an indicator would see, and 0.82 becomes -0.0568. This site published exactly the same collapse for stablecoin supply last month, and the pattern is the same one both times.

The second is to give the flow more room. Instead of tomorrow, ask about the next thirty days, with each day's flow measured against its own trailing year so that 2012 and 2026 are on comparable terms. Pooled over 5,204 days the answer is +0.0141, the wrong sign for the story and near zero anyway. Measured inside each calendar year and averaged, it is -0.0152, the right sign and as small. Of the fifteen years, nine lean the way the story needs and six lean against, which is what a fair coin looks like.

The third is to sort the days into five groups by flow and look at what followed. If the story held, the days with the heaviest outflows should be followed by the strongest months, in order. They are not. The heaviest-outflow group is followed by +5.57% over thirty days on average, the heaviest-inflow group by +6.25%, and the gap runs 0.68 points the wrong way. The middle groups do not line up either.

The Version of This Test That Passed

Two of the checks above came back looking like support before they were finished. Both are worth showing, because the article would have been wrong in an interesting way if either had been left alone.

The first was a fix for a real problem. Thirty-day windows starting on consecutive days overlap almost completely, so counting them as separate evidence counts the same month thirty times. The standard repair is to take every thirtieth day instead, giving 173 or 174 windows that share no days at all. Done that way, the result flips: the correlation comes out at -0.1884, the right sign for the story, and the heaviest-outflow windows beat the heaviest-inflow windows by 9.09 points. That is the strongest result in this entire article, and it points the other way.

It is an artefact of where the counting starts. There are thirty ways to slice the history into non-overlapping windows, depending on which day the count begins, and every one is as valid as the others. Run all thirty. The correlation lands anywhere from -0.1884 to +0.1811, and it is negative in exactly 15 of the 30. The gap between the outflow and inflow groups runs from -9.29 points to +9.26 points, and is positive in exactly 15 of the 30. The average across all of them is -0.53 points, which is nothing.

So the one test that supported the story was a coin flip, and the run reported above happened to start on day zero. A single non-overlapping slice is one arbitrary draw dressed as a fix for double counting.

The second near miss was smaller and stranger. Scaling each day's flow by the size of the exchange balance seems obviously sensible. Doing it produces the only reading above 0.10 anywhere in the sixty-one: -0.1449 with the flow six days early, on a test statistic that would pass any ordinary bar. Three days in mid-2011 produce it. Exchanges then held 2.16 coins, 302 coins and 775 coins, so dividing by the balance turns one ordinary day into a number 9,684 times its own denominator. Rank the days rather than using those sizes and the reading falls to +0.0032. Start the sample in 2012 and it falls to -0.0037.

Both of these are the same failure with different clothes: a handful of observations wearing the authority of five thousand.

The Verdict, and What It Does Not Cover

Exchange netflow was assessed as a fourteenth indicator for the dashboard and it is rejected, for two separate reasons. It carries no lead worth acting on at any horizon tested from thirty days early to thirty days late. And the series it is usually charted beside is not independent evidence. The exchange balance is just the flows added up. On the one day in fifteen years when it was not, it moved by more coins than the network sent.

Four things this article does not say. It does not audit the 27% to 29.5% figure, which is behind a paid tier and was not checked. It does not say why coins arrived on exchanges during the fall, because the file records movement and not motive. It does not explain the rise from $58,525 to $78,359, and other mechanical accounts of it were reported elsewhere and are not tested here. And it does not treat the exchange balance as ground truth, because it is a vendor's estimate of which addresses belong to exchanges, which is the whole point of the middle of this piece.

What it does say is narrow and holds. Float did not leave exchanges during the drawdown. The daily flow did not lead the price, at any lag, in fifteen years of data. And the number most often shown alongside it has been telling the same thing twice.

Every figure here comes from this site's own copy of the Coin Metrics daily file, which ends on 21 August 2026. One command produced them: python3 ml/research/exchange_netflow_test.py.