How much of Polymarket is bots?

Somebody sat down and chose the name swisstony. It has been on a Polymarket account since July of last year. Behind the name is a record no person could have typed: close to 6 million trades across nearly 190,000 separate markets, placed in almost every hour of every single day for 11 straight months, typically a second or two apart. They cost more than $6 million in fees.

Polymarket is a site where people bet on how real events will turn out. Elections, football, whether a central bank will move rates, whether bitcoin will be higher than it is right now. You pick a side and put money down, and if you are right the people who took the other side pay you. It settles on a public blockchain, so every bet anyone has ever placed is written down where anyone can read it.

Most people place those bets by opening the website and clicking. The Polymarket API does the same job for software: a program can place exactly the same bets without anyone touching a screen, and it can do it far faster than a hand can. Whole guides exist explaining how to build one.

Bitquery indexes every transaction on Polygon, the blockchain Polymarket settles on. We took the complete record from the point our data begins, in September of last year, to the middle of August: every trade, and every wallet that placed one. Then we asked Polymarket's own public directory, one wallet at a time, whether each of them had ever registered an account, and we measured how each of them behaved: what hours it kept, and how long it waited between one trade and the next. The machines here have been written up one at a time, usually by the people running them. This is a count of all of them, and it shows that most of the trading is done by a very small number of wallets, almost all of which hold an ordinary Polymarket account.

What you are actually betting on

A Polymarket bet has two sides and a deadline. You buy shares in one answer, and each share pays a dollar if that answer proves right and nothing if it does not. The price is the crowd's estimate of the odds: a share at 60 cents means the market gives that answer about a 60% chance, and somebody is taking the other side at 40. The two prices always add to a dollar, so nobody is betting against the house. They are betting against each other, and Polymarket holds the money until the question resolves.

Getting in is cheap, the stake is the most that can be lost, and on the fastest markets the answer arrives within minutes. That combination is what draws people. It is also what makes the place worth automating: the questions are simple, the settlement is mechanical, and a new one opens every few minutes, right through the night.

Polymarket takes its cut from whoever crosses the spread rather than waits at it. That charge is the only reliable income the venue makes from these markets, and it turns out to be one of the sharpest ways to see who is doing the work. Over the record it came to about $346 million, against $34 billion of matched trading. We measured how those fees were split in a companion piece on the 5-minute markets.

How you tell a machine from a person

Two things about a wallet are observable from outside, and they answer different questions.

The first is whether it ever used the website. Polymarket keeps a public directory of every account that has signed up, and it will tell anyone who asks whether a given wallet has one. A wallet that trades heavily and has no entry in that directory has never opened the site. It reached Polymarket through the API instead.

The second is whether it ever stops. We measured, for each wallet in each month, whether it placed orders in all 24 hours of at least one single day, and how long it typically waited between one order and the next. A wallet that acts right around the clock, and usually within a minute of its last move, is not being operated by hand. Orders matter here, not just trades: a big resting order can keep filling while its owner sleeps, so we timed what each wallet did, not what happened to it.

It matters that neither test is the same as counting robots. A person can write a program and let it trade while they sleep, and that wallet will hold an ordinary account with a name on it. Someone else may prefer to place their bets through the API by hand and never register, which would leave a human wallet with no directory entry. Both tests are deliberately strict, so both undercount. Between them they mark out a floor, and the floor turns out to be high enough to answer the question.

Fewer than 1 wallet in 100

Over the record, 2,347,464 wallets placed a trade. About 1 in every 200 of them met the test for a machine. That sliver placed more than 6 of every 10 trades on Polymarket, moved more than a quarter of the money, and paid more than a third of the fees.

No single operator explains this. The biggest machine accounts for a twentieth of what the machines traded, and the largest 10 together for about a sixth, so what the record shows is a crowd rather than a firm. A typical one of them ran for 4 months and traded 2,925 separate markets.

They also cluster at the fast end of the wire. The wallets that usually wait less than 5 seconds between one order and the next are a small crowd on their own, and between them they place well over half of every trade on the venue.

The machines signed up

The obvious guess is that automated wallets skip the website, because a program has no use for one. The record says the opposite, and it says it clearly.

Nearly 8 in every 10 machine-run wallets carry a username somebody picked. Another 13.3% registered and never bothered to name themselves. Fewer than 1 in 10 have no account at all, which is a smaller share than the crowd at large manages. Whatever the operators are, they signed up, and some of them told the room what they were doing when they chose a handle.

Read across all of Polymarket, having no account is a much weaker signal than it looks. Wallets with no directory entry are 14.6% of everyone who traded but only 4.4% of the money and 7.1% of the fees. They are numerous and small, which is the opposite of what a hidden professional fleet would look like.

So the account test is worth running and worth reporting, and it is a floor rather than an answer. What it establishes is that at least 1 wallet in 7 here has never seen the website. What the behaviour test establishes is that the trading itself is mostly automated, whoever signed the forms.

The faster the market, the more mechanical it is

Polymarket sells questions on every timescale, from an election months away to a bitcoin market that opens and closes inside 5 minutes. We sorted every trade by how much time was left before its market settled. The pattern runs one way and it does not wobble.

On markets settling within a quarter of an hour, machine-run wallets are on more than half the money. On markets more than a week out, about a sixth of it. That is what you would expect if the advantage automation buys is reaction rather than judgement.

The same gradient shows in the account test, faintly. Wallets with no profile take 7.7% of the money in the fastest markets against 4.2% in markets more than a week out, which points the same way at a fraction of the strength.

Who actually wins

The received story about bots on prediction markets, told in headlines about millions in profits and in screenshots posted by the operators themselves, is that the machines are cleaning up while everyone else falls behind. Individual machines certainly do make money. Across the whole population the picture is different, and the difference is the fee.

Over those 9 days, everybody lost. The machines went in slightly behind before any fee was charged, so on the trading itself the people were a touch ahead of them. Then the fees arrived, and both sides of the room finished down, the people by rather more in total. What the machines are is not the winners but the extremes: they are about 1 in 7 of the wallets that made over $10,000, about 1 in 6 of those that lost over $1,000, and far scarcer in the crowd that finished near level. Automation does not buy a better result here. It buys a bigger one.

Every month, a bigger share

The machines' own share of the money has roughly doubled across the record, from about 16% in September to about 29% in July. A second measure says more. Every matched trade has two sides, so you can ask how much of the money had a machine on either one of them. A year ago three quarters of it was people trading against people. By this August it was only just over half.

Set beside the retention we measured separately, where fewer than 4 in every 100 of the wallets that turned up in February were still trading 6 months later, the direction is not hard to read. The people leave and the programs stay, because a program has no reason to get bored. The share of money coming from wallets with no Polymarket account has crept up too, from about 4% to about 7%, counting a wallet as unregistered only in the months before its account actually existed.

What is on the chain

The wallets below are the largest machine-run accounts by fees paid, and the exchange contracts the trades ran through. Everything above was read from the public record on Polygon, and anyone can read it again: the live markets are listed on DexRabbit, the trade feed behind them is the Polymarket API, and the queries used here are written out in the prediction market documentation. We measured how much of Polygon this one market now accounts for separately.