Exchange Netflows: What the Numbers Show

Exchange netflow is the coins sent to labelled exchange wallets minus the coins withdrawn from them over a period, and a positive number means more went in than came out. The subtraction is trivial. The hard part is the labelling behind it, and in Coin Metrics' bitcoin data the two halves of one ledger now disagree: from January 1 to October 21, 2026 the flow series says 345,380 more BTC left exchanges than entered, while the same provider's exchange balance rose by 369 BTC.

That gap is not a quirk of one dataset. It is what the page below is about: how a netflow figure is built from a list of addresses that keeps changing, what happens to the number when the list changes, and which questions a netflow chart can honestly answer. Everything computed here comes from one public series and can be repeated. Nothing in it is a signal about price.

What is being subtracted

Coin Metrics defines the inputs in its metric catalog. Exchange deposits (FlowInExNtv) are "the sum number of native units sent to exchanges that interval, excluding exchange to exchange activity," and exchange withdrawals (FlowOutExNtv) are the equivalent for coins leaving. The balance (SplyExNtv) is "the sum of all native units held in hot or cold exchange wallets that interval." All three carry the same qualification: "Metric values reflect activity of addresses currently known to belong to the entity, from each address's first non-zero balance."

That last sentence carries most of the weight. The series is not a record of what was known on each day. It is today's list of known exchange addresses, applied backward to each address's first non-zero balance. When a provider learns that an address belongs to an exchange, history is recalculated as though it had always been known.

CryptoQuant's data guide describes the same process from its side. Netflow, inflow and outflow share one definition there: the "amount of coins deposited/withdrawn into/from the exchange wallets." It says "the heuristic running to label the exchange addresses continues to evolve around time" and that "addresses that are newly created or not labeled need some time to be corrected." And it gives the failure it worries about in the vendor's own words: a movement that "gives an impression of an outflow but instead it is just the same bitcoin staying in the same exchange."

Neither document gives an accuracy rate. That is typical of this kind of data, and it means an outsider has to test the series with the only tool available: whether its own parts add up.

A ledger that should close

If deposits and withdrawals are the only ways coins enter or leave the labelled set, then each day the change in the balance must equal deposits minus withdrawals. The test is a one-line subtraction: balance today minus balance yesterday, minus (deposits minus withdrawals). Call the result the gap. A closed ledger has a gap of zero.

Here is the gap on two stretches of three days in 2026, in whole bitcoin (so columns can differ by one from rounding):

On August 13 the balance rose by 46,954 BTC, 1.63 percent of the balance the day before, while the flow series showed a net inflow of only 3,192. The other 43,763 coins arrived in the balance without arriving as deposits. A flow-based reading of that day would say "modest accumulation"; a balance-based reading would say "the largest one-day rise of the year." Neither is wrong about its own data. The gap is the part the two cannot explain.

When the ledger closed, and when it stopped

The ledger was exact for almost all of the series. Then it stopped.

In the first block the gap is within 0.1 BTC on 2,648 of 2,655 days. Seven days break it. They are 36.7 BTC on October 17, 2023 (reversed by minus 36.7 on October 28), 16.9 on March 9, 2025, 389.7 on May 24, 2025, 10.3 on August 12, 2025, 1.7 on November 5, 2025 and 1,271.4 on February 9, 2026. A ledger that closes to the fraction of a coin for more than seven years is not one that was measured day by day. It is a balance computed as the running total of the flows.

The second block is different in kind. On 194 of its 195 days the balance rose by more than the flows implied, and the one exception was minus 2,502 BTC. The gaps add up to 346,979 BTC on the positive days. The flow series for those days says 274,465 BTC net left the labelled wallets; the balance says 70,012 BTC net arrived.

The data carry their own explanation of where the break falls. Every value in this pull has status "flash," one of the three status labels (flash, reviewed, revised) the API lets a caller filter by; the catalog marks these metrics as reviewable, and the documentation reached for this page does not define the labels. Each row also carries the time its status was set. For every row dated before April 10, 2026, that timestamp falls between April 2 and April 20, 2026. They were written in a single backfill, days to years after the dates they describe. For rows from April 10 on, the timestamp is 1.04 to 1.91 days after the date, with a median of 1.09 days: written the morning after, one day at a time.

So the pattern fits a simple account: when the provider rebuilt history in April it applied one address list to the whole past and the ledger closed by construction, and since then each day has been written with whatever list was current that morning. The measured gap is then the effect of the address list growing between rebuilds. That account matches the data, and it is an inference from timestamps and the catalog wording, not something either provider has stated. It makes a prediction that anyone can test: if the history is rebuilt again, the recent rows should be rewritten and the 2026 gap should shrink toward zero. Comparing this pull with an earlier pull of the same series made the same day, none of 2,851 balance values differed, so nothing was rewritten in between.

The practical consequence is for anyone using the netflow series for the last six months: the flow figures there are the ones the provider's own balance does not corroborate. The balance for December 31, 2025 was 2,674,747 BTC and for October 21, 2026 it was 2,675,115, almost the same, while the flows over that stretch imply a decline of 345,380.

Reading a headline with the two-number check

A weekly headline is the easiest place to apply the test. Take the seven trading days from October 15 to October 21, 2026. Deposits summed to 135,955 BTC and withdrawals to 141,167 BTC, so the netflow was minus 5,211 BTC: a modest net outflow. The balance was 2,675,717 BTC at the end of October 14 and 2,675,115 BTC at the end of October 21, a fall of 602 BTC. The direction agrees and the size does not, because the gap for the week is 4,610 BTC, almost as large as the flow itself. Over the 30 days from September 22 to October 21 the netflow was minus 86,562 BTC while the balance fell 44,379 BTC.

Neither version is the better one by default. A reader who quotes the flow figure should say that it comes from the flow series, and a reader who quotes the balance should say it comes from the balance series. What the check does is turn "coins left exchanges this week" into two numbers whose disagreement is itself information: it measures how much of the weekly figure depends on the provider's list changing.

Year by year, and the same test on Ether

Summed by calendar year, the two sides of the bitcoin ledger match to the coin until 2025 and part company in 2026. The sums are of daily changes, so 2019 starts on January 2.

Ether is the only second case that could be tested: the Ethereum versions of USDT and USDC return a "forbidden" error for these metrics on the community tier. It behaves differently, and the difference is informative. The Ether history was rebuilt later, with timestamps between September 5 and September 21, 2026, and rows from September 19 onward are written the morning after. In the rebuilt block, from January 2, 2019 to September 18, 2026, the ledger does not close exactly: the balance fell by more than the flows implied on 2,815 of 2,817 days, by a median of 114 ETH a day, and the median of the daily ratio of that shortfall to the day's gross flows is 0.03 percent. Over the full block the balance fell by 14,052,960 ETH while the flows implied a fall of 13,122,374, so about 930,586 ETH of decline is unexplained by transfers. Gas payments from exchange wallets and other balance changes that are not deposits or withdrawals are possible causes; this page did not test them. The live block, only 34 days from September 19 to October 22, is too short to compare with Bitcoin's 195.

Two conclusions follow. A rebuilt history is not the same thing as a closed ledger: bitcoin's closes to the coin and Ether's does not. And the rebuild date is specific to each asset, so a break in one asset's series says nothing about the other's.

Why a small mislabel can flip the sign

A netflow is a small difference between two large numbers, so a modest labelling error in either one can change its sign. Over the 2,851 days in the pull, the median day had 54,831 BTC of gross movement (deposits plus withdrawals) and a net of 2,701 BTC in absolute size, 5.5 percent of gross. On 76.5 percent of days the net was under 10 percent of gross. Deposits and withdrawals were almost balanced: the median deposit figure was 27,076 BTC and the median withdrawal figure 27,148 BTC, and withdrawals exceeded deposits on 49.5 percent of days.

Take the real daily values and overstate only the deposits by a fixed percentage, as would happen if internal moves between an exchange's own wallets were wrongly tagged as deposits:

A ten percent error in one input, roughly the size of a median day's net itself, reverses the headline direction on more than one day in five. The calculation is a what-if on real inputs, not a measurement of how wrong the labels are. It shows how much accuracy is needed before the sign is meaningful, and neither vendor publishes one.

Where the errors come from

Each of these is a mechanism, written from the vendors' own descriptions of how they work, and none is observed in this data except where noted.

The second row is the CryptoQuant example in a different form. The sixth is the reason gross flows are better understood as churn than as unique coins, as the realized price guide notes when it uses the same series for the turnover of exchange wallets.

Custodians deserve a note because of the spot funds. The trusts described in the spot ETF custody guide hold their coins with a custodian, and the iShares Bitcoin Trust's June 30, 2026 10-Q names Coinbase Custody Trust Company in that role. Whether any provider counts such custody wallets as "exchange" balances depends on its labels, and the two docs read for this page do not say. A reader who wants to attribute a large flow to ETF creations should ask that question of the provider first.

The largest days in the record

The biggest daily netflows in the 2019 to 2026 data show a pattern that is hard to read as demand.

Seven of the 15 largest absolute daily netflows in the series fall between November 2 and November 23, 2022, the weeks of the FTX collapse, which is the obvious context and is covered in the page on exchange bankruptcy and customer funds. Flows alone cannot say which addresses moved.

Two consecutive rows show what a reconciliation looks like. On November 17, 2022 the series records a net outflow of 52,318 BTC, and on November 18 a net inflow of 88,036 BTC. Added together the two days are a net inflow of 35,718 BTC. If the first day's coins went to a wallet that was not labelled at the time and the second day's inflow included their return, the two numbers would overstate each other, and the two-day sum would be the more reliable one. That is a hypothesis about how to read the pair, not a finding about those coins: the data do not show where the bitcoin went.

What netflow did next to price

The folk rule printed in the CryptoQuant guide is that an increase of inflow is a "bearish sign" and an increase of outflow a "bullish sign." Looking at what this series actually did, with the caveat that it is today's address list applied to the past:

The same-day correlation between netflow in coins and bitcoin's return is minus 0.05, and the correlation between one day's netflow and the next day's return is 0.01. Days of net inflow did not produce lower next-day returns; the mean was higher, 0.22 percent against 0.10 percent, on sets of nearly equal size. Gross movement is another matter. Its correlation with the size of the day's price move (the absolute daily percentage change), up or down, is 0.40, and the table shows the same thing in bins.

Medians are taken within each bin. Exchange activity rises with how far the price moves, which fits the plain reading that more people trade when the market is moving, and it says nothing about direction. It also shifts over the years: the median day's gross movement was 70,528 BTC in 2020, 51,297 BTC in 2025 and 43,747 BTC in 2026 to October 21. This is a description of one series on one address list. The next-day columns are not a trading rule, and an effect that disappears with a different list or a different year is not one.

What would change this analysis

Three things would. A re-pull of the same series in some weeks that shows the April 10 to October 21 rows rewritten and the gap closed would confirm the timestamp reading. A vendor statement of how newly labelled addresses are treated would replace the inference with a rule. And a second provider whose balance and flows agree over the same months would show the break is specific to one pipeline and not a feature of the measurement. The sensible use of netflow figures meanwhile is as a rough measure of activity, with the balance cross-check run before trusting any figure's sign. The exchange reserve risk checklist covers the separate question of what an exchange discloses about its own holdings.

Frequently Asked Questions

What does exchange netflow mean?

It is deposits to labelled exchange wallets minus withdrawals from them over a period. Positive means more coins entered than left. CryptoQuant defines the underlying flows as the "amount of coins deposited/withdrawn into/from the exchange wallets," and Coin Metrics excludes exchange-to-exchange activity.

Does a large exchange inflow mean people are about to sell?

This page's data do not show that. Across 2,850 days of Coin Metrics bitcoin flows the correlation between a day's netflow and the same day's return was minus 0.05, and between a day's netflow and the next day's return it was 0.01. Coins deposited to an exchange can be sold, lent, used as collateral or moved internally, and the series cannot tell which.

Why do different sites show different netflow numbers?

Each provider keeps its own list of exchange addresses and its own rules for exchange-to-exchange and internal transfers, and the list keeps changing. Coin Metrics states that its values reflect "addresses currently known to belong to the entity." Two providers with different lists will report different flows for the same day.

Why did Coin Metrics' flows and balance disagree in 2026?

The measured fact is that from April 10 to October 21, 2026 the balance rose by 70,012 BTC while the flows implied a fall of 274,465, a gap of 344,477 BTC. The cause is not documented in what this page read. The row timestamps suggest earlier history was rebuilt in one batch in April and later rows were written daily, but that is an inference.

Are exchange outflows to self-custody a good sign?

Whether coins leave an exchange for a person's own wallet, for another venue's unlabelled wallet or for a fund's custodian is not visible in a netflow total. This page does not rate outflows as good or bad. For how to move coins to your own wallet, see the guide to withdrawing from an exchange to self-custody.

Does a gap between flows and balance mean coins are missing?

Not on this evidence. The gap is the difference between two series from the same provider, and the likeliest sources are changes to the list of addresses counted as exchange wallets. Coin Metrics' own definition says values reflect addresses "currently known to belong to the entity," which means the list is not fixed.

Is the balance series more reliable than the flow series?

Neither can be called more reliable from outside. Through April 9, 2026 the balance behaved as the running total of the flows, so it carried no independent information. Since then the two disagree by hundreds of coins on a median day, and only the provider can say which one reflects a change in labelling.

Can I check the numbers myself?

Yes. The Coin Metrics community API serves FlowInExNtv, FlowOutExNtv and SplyExNtv for bitcoin at daily frequency. Subtract each day's balance from the previous day's and compare with deposits minus withdrawals. Where the two differ, the series is not internally consistent.

Sources

CryptosEyes publishes general educational research, not investment, legal, or tax advice. This article describes how a public data series is built and how it behaved. Nothing here predicts price, and exchange flow figures do not show anyone's intent or any exchange's solvency.