I have been testing whether trade-size distributions on major crypto exchanges behave like real market data. One of the tools for doing that is a leading-digit test, better known as Benford's Law. Most venues looked roughly as expected. Binance did not.
On the full UTC day I sampled, 57% of all the individual BTC/USDT prints on one of the highest-volume and most systemically important spot exchanges in crypto were for exactly $5.60. Not approximately. Exactly. They were also all for the same precise quantity of Bitcoin: 0.00007 BTC. The world's largest crypto exchange runs, in trade-count terms, on a flood of $5.60 executions, and there is a perfectly reasonable explanation for this, and the explanation has the unfortunate side effect of breaking one of the standard statistical tests for whether an exchange is faking its volume.
When I first ran the numbers I assumed I'd made a mistake. 58.79% of executions starting with a single digit is not a value a normal leading-digit distribution produces, and the obvious explanation was that my pipeline had double-counted or filtered something wrong somewhere upstream. I spent the better part of an afternoon checking the pipeline before I started taking the result seriously. The thing that broke the puzzle was checking which specific trade sizes were in the digit-7 bucket. I had expected noise across many magnitudes — different algorithms, different sizes, all happening to share a leading digit. The trades were not across many magnitudes. They were at one magnitude. Almost all of them were exactly 0.00007 BTC.
That's when the explanation got mechanical.
Binance, sensibly, does not want people to clog its matching engine with one-cent trades. So it enforces a minimum trade size: every BTC/USDT execution must clear five dollars of notional value. (Five dollars, not five dollars and sixty cents; we will get there.) Binance also requires trade quantities to be in increments of 0.00001 BTC, because matching engines like round numbers in their preferred units even if those units look weird to humans. Combine these two rules and you get the smallest eligible trade on Binance: take five dollars, divide by the current Bitcoin price, round up to the next 0.00001 BTC. When BTC was trading at $80,000 on the day I sampled, that smallest-eligible-trade worked out to 0.00007 BTC, which at $80,000 happens to be $5.60. So if you wanted to make as many trades as possible while spending as little money as possible — which is exactly what some algorithms may want to do, for reasons such as latency probing, queue interaction, fee optimisation, or other microstructure incentives — you would generate executions for exactly 0.00007 BTC and pay exactly $5.60 for each one. This is what 57% of Binance's BTC/USDT executions did on the day I looked at. (Same-day sample: a full UTC day of executed trades from Binance's public trade tape. I have not yet checked every recent day, but spot checks on adjacent days show the same pattern.)
This mechanical quirk doesn't just dominate the order book; it breaks one of the most famous statistical tools for detecting financial fraud.
There is a thing called Benford's Law, which is a statistical regularity in real-world numerical data: in many naturally occurring datasets, more numbers start with the digit 1 than start with 2, more start with 2 than start with 3, and so on down to 9. Roughly 30% of numbers in a Benford-distributed dataset start with 1; only about 5% start with 9. Trade sizes on real exchanges generally obey this. When they don't — when leading digits cluster in odd patterns — this is sometimes a sign that the trades are not real, that they are being generated by a program whose author did not know to make the numbers look natural. So "does this exchange's trade size distribution match Benford" has become one of the standard statistical tests for whether an exchange is faking its volume. It is not the only such test, but it is a popular one and a reasonable one. It is also the test I started with, which is part of why I noticed Binance in the first place.
If you ran the Benford test on Binance for the day I sampled, here is what you would find: 58.79% of executions had a leading digit of 7. The Benford-expected value is 5.80%. Binance fails the Benford test by an order of magnitude more than any other major exchange. Binance is also, by a large margin, the highest-volume real spot exchange in the world. You can see where this is going.
The reason 58.79% of Binance's executions start with the digit 7 is that 57% of them are exactly 0.00007 BTC, which starts with a 7. The cluster of executions at the minimum-notional floor doesn't just dominate the trade size distribution; it dominates the leading digit of the trade size distribution. The Benford test, naively applied to trade sizes on Binance, looks at this and says: this exchange has anomalously many trades starting with 7, which is suspicious, which suggests fake volume.
The Benford test, naively applied, is wrong. The executions are fully consistent with real flow clustering at Binance's minimum-order constraint. They are very small, but they appear to be genuine — generated by algorithms (market makers, latency arbitrageurs, the various species of microstructure participants who have reasons to want a high count of cheap fills) that have collectively figured out that the cheapest eligible trade on Binance is 0.00007 BTC and have collectively decided to make a lot of them. The math is mechanical. The pattern is structural. This particular anomaly is not good evidence that Binance is faking BTC/USDT volume.
It also has a testable prediction, which is the part that made me feel better about the finding. If Bitcoin's price moves enough, the minimum-notional floor shifts. At $90,000 BTC, the smallest eligible trade becomes 0.00006 BTC, and the leading-digit cluster should migrate from 7 to 6. At $70,000 it becomes 0.00008. At $50,000, 0.00010, which is digit 1. If you look at Binance's leading-digit distribution across different price regimes, you should see the dominant cluster walking around the digit axis in lockstep with Bitcoin's price. The transition points are predictable: digit-7 dominance holds for any BTC price in roughly $71,429 to $83,333 (that is, $5 / 0.00007 to $5 / 0.00006), and then jumps to digit-6 or digit-8.
I have not yet tested this prediction across enough days to verify it. I will. If it turns out to be wrong I will say so.
None of this means Benford-style tests are useless. It means they need to be applied with venue context. A Benford test that first filters out trades at the minimum-notional floor — or stratifies the distribution by trade-size band — would recover useful signal on Binance. The lesson is not that Benford is broken. The lesson is that exchange-quality metrics need to understand venue rules, tick sizes, lot sizes, minimum notionals, and market microstructure before converting statistical anomalies into accusations. The methodologies that don't do this work will mislabel Binance specifically, and mislabel it badly, because Binance happens to be the venue where the minimum-notional cluster is largest in absolute terms.
I want to be honest that I have spent several weeks on this project getting precisely this kind of result, where a metric I expected to behave one way turned out to be measuring something subtly different. The Binance finding is one of the cleaner cases. There are messier ones — the Coinbase result is different in shape and not yet fully understood, and there is a separate problem with the way cross-venue price deviation is computed on volatile days that I'm still working through. I'll write those up as I get to them.
That is the work: not finding anomalies, but finding out what they actually mean. The methodology I'm building, called Assay, is mostly the sum of taking these results seriously instead of explaining them away.