Invest1 publisher3 min readPublished
Coin Metrics rewrites Ethereum's exchange-flow history back to genesis with wallets it found later
Coin Metrics has recomputed all daily and hourly Ethereum exchange-flow metrics from the first block, using the exchange wallets it knows today. Outflow backtests run on a fresh download can now count wallets nobody had identified on the trade date.
The Investor · Invest desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Coin Metrics outlined the rebuild on Sept. 28, expected Ethereum to finish on Sept. 30, and posted its completion notice on Oct. 1 at 17:04 UTC.
- Its separate Point-in-Time series counts an exchange address only from the date it was discovered, so later finds leave earlier intervals untouched.
- The notice did not include revision amounts or any comparison of an ETH strategy run on old and new data, according to CryptoSlate.
- CryptoQuant's documentation warns that its ETH exchange-flow endpoint does not support point-in-time accuracy as clustering updates add exchange wallets.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- cost Any analyst who refetched ETH flow history has lost the vintage the old test ran on, and a saved query date cannot restore values pulled again from a mutable endpoint.
- decision Research desks have to pick the series by question: PIT for any claim about what a trader could have acted on, Standard for supply studies that want today's address coverage.
- constraint Changing data vendors does not remove the problem, because CryptoQuant's history also moves and each provider's revisions have to be measured against its own update records.
Coin Metrics' Standard flow metrics count every address the company currently knows belongs to an exchange, and each address contributes from its first nonzero balance [4]. A wallet identified this year that was funded years ago gets its whole history booked as exchange flow, back to that first deposit [4]. Past values change each time the label list grows [4]. Coin Metrics said the recomputation was meant to keep that Standard logic separate from its PIT series [17].
Measuring the effect on any ETH signal takes two copies of the same product. According to CryptoSlate, the comparison that measures this revision is Standard history saved before the rebuild against Standard history after it, while Standard against PIT tests a different rule about what was known [12]. Even the timing is loose. The Oct. 1 notice does not date when each affected value became available [7].
The rules most exposed are the ones that trade on small gaps between averages. Glassnode's March 13, 2026 hypothetical used Binance's bitcoin exchange balance [10]. It entered when a five-day moving average fell below a 14-day average, exited when the shorter rose back above, and ran from Jan. 1, 2024 to March 9, 2026 on $1,000 [10]. That window is about 798 days of possible crossovers [1]. If the averages are simple ones, revising a single day's reading moves the five-day average by a fifth of the revision and the 14-day by a fourteenth, so the gap the rule trades on shifts by about 13% of it [3]. That can move the day the averages cross, and the price of the trade taken on it. CryptoSlate uses the bitcoin test to isolate data-vintage risk; it does not size the ETH revision [16].
CryptoQuant has the same problem on a fixed schedule. It runs automatic updates every Tuesday at 00:00 UTC and says values can change slightly [9]. That is 52 scheduled revision windows a year [2]. The newest observations, the ones a live signal acts on, are the likeliest to change [9].
The ETH revisions could turn out small enough that no crossover date moves, leaving the rebuild relevant to supply charts and little else. They could cluster in the chain's early years, flattering long-dated ETH outflow backtests most. Or the hourly series could move more than the daily one, and intraday rules would take the hit first.
I think any ETH outflow result built on a Standard download made after Oct. 1 is unverified until it is rerun on PIT data or on a saved pre-rebuild copy. CryptoSlate says that test needs paired data vintages, fixed rules and publication timing, and that no change in returns has been established [11]. A positive result would also need more than flow data: a withdrawal measures movement relative to wallets attributed to exchanges, and calling it buying takes other evidence [13]. For a given rule, this view is wrong if the two Standard vintages produce the same entry and exit dates.
What to watch
- A published diff of pre- and post-rebuild Coin Metrics Standard ETH flows that gives actual revision amounts by year and frequency.
- A rerun of an ETH outflow crossover rule on PIT versus current Standard data, with fixed rules and publication timing, showing whether entry and exit dates move.
- CryptoQuant's Tuesday update records, to see how far its recent ETH exchange-flow values change from week to week.