Polymarket built a casino. Almost all of the players have left.

At ten to four in the morning in New York, nearly 800 people were awake and betting on bitcoin. Not on where it would be next year, or next month. On whether it would be higher 5 minutes later.

They were on Polymarket, a site where people bet on how real events will turn out. Most of it is what you would expect: elections, football, whether a central bank will move rates. You take a side, you put money down, and if you are right the people who took the other side pay you.

The market these 800 were in had opened 17 minutes earlier and would be settled before most of them finished watching. The typical stake was under $3. When the clock ran out bitcoin was higher, so the people who had said it would rise collected and the people who had said it would fall lost what they put in. Then a fresh market opened for the next 5 minutes, and the crowd bet again.

These are Polymarket up and down markets. The exchange runs thousands of them a day now, on bitcoin and a handful of other coins, right through the night.

Bitquery indexes every transaction on Polygon, the blockchain Polymarket runs on. We took the complete record of these markets from the point our data begins, in September of last year, through the middle of August: every market created, every bet matched against another bet, every result, and every fee charged. It comes to close to 700 million matched trades. What it shows is a product that grew about three times faster than the exchange around it, and then lost almost everyone who had turned up for it. Somewhere in the middle, Polymarket started charging people to play.

What you are actually betting on

An up and down market is a bet with 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 the exchange holds the money until the question resolves.

Getting in is cheap and easy. A few dollars buys a position and the answer arrives within minutes. Nothing is borrowed, so the stake is the most that can be lost, which is what separates this from betting the same price move with borrowed money.

What resolves it does not live on the blockchain. Polymarket reads an averaged bitcoin price published by Chainlink, an outside data service, and compares the value at the start of the window with the value at the end. Higher means Up. The chain records the verdict and never the number that produced it.

Over the whole record the two answers have won almost exactly as often as one another, with Down a shade ahead of Up across the hundreds of thousands of markets that have resolved. Whatever is emptying these players' accounts, it is not a rigged coin.

The five-minute market

For the first months of our record these were mostly 15-minute markets, with a slower 4-hour version and a handful of daily and weekly questions around them. Polymarket also tried the same format on shares, stock indexes, gold and oil. Those experiments are effectively over.

In February the exchange began running crypto markets on a 5-minute clock. That single change is visible in the record as a wall.

A shorter clock means more markets, more rounds, and more chances to bet in an evening. Volume followed it almost exactly.

March, and the machine that never sleeps

March was the top of our record. More money went through these markets that month than in any other we can see, and more wallets bought into them. Almost all of that money sits on one coin. By August, bitcoin on the 5-minute clock alone took close to two thirds of everything traded, and bitcoin across all three window lengths took three quarters.

Polymarket's weight is not confined to its own markets. It has become the single largest source of activity on Polygon itself, which we measured separately.

The other thing the record shows about this market is that it has no closing time. Betting dips in the small hours of the New York morning and recovers by breakfast, but it never stops, and the quietest hour of the night still turns over thousands of dollars a minute, drawing tens of thousands of separate wallets across the month.

The house starts charging

For the first four months of our record, trading these markets cost nothing beyond the price of the shares. Polymarket fees arrived in January, and the record shows them starting precisely: not one trade before that month carried a fee, and tens of millions of trades after it did.

The fee falls on the taker, meaning whoever accepts an offer that is already sitting on the order book. The trader who posted that offer and waited pays nothing, and is paid a rebate out of what the taker was charged. By Polymarket's published schedule the charge is heaviest on markets priced near an even chance and tapers away as the odds move towards certainty, which is where the money in a 5-minute market mostly sits.

Measured against the volume it was charged on, the rate at the peak was a little over four cents in every dollar. That is a heavy toll for a market where the average bet is settled inside 5 minutes.

The money is collected by the exchange contract itself, which is named as the recipient on every charged trade in the record. Across the eight months from the first fee to the middle of August, the charge on these markets alone comes to about $150 million, before the rebates paid back to the traders posting offers. That is the gross take on one product, over eight months, from a crowd whose typical bet is worth less than a cup of coffee.

Where the money went

When you bet against other people rather than against a bookmaker, the money only moves sideways. Every dollar won is a dollar somebody else lost, so the crowd as a whole can never be up. A fee changes that, and we can measure by exactly how much.

We took one recent week, every up and down market on every coin that opened, traded and settled inside it, and rebuilt each wallet's position from its own trades: what it paid, what it was charged, and what it was paid out when the markets resolved. Before fees, the crowd finished level with itself to within a tenth of one per cent, which is the arithmetic working as it should. After fees, the same crowd was down by nearly the whole amount charged.

The players are not losing to a smarter opponent. They are losing the toll, and they are losing all of it.

Level is an average, not an experience. Fewer than a third of the wallets finished the week ahead, and the typical one was already a couple of dollars behind before any fee was charged, because whoever crosses the spread pays it to whoever was waiting there. The median wallet ended the week a few dollars down, while the extremes ran to tens of thousands in both directions.

The more you play, the worse it gets

Sorting those wallets by how much they played makes the pattern plain. People who wandered in for a single market lost the least, about a dollar at the median, and roughly every step up in play multiplied the typical loss, out to about $100 for the wallets that played more than 500 markets in the week. A fee charged per trade compounds with every extra trade.

The losses concentrate where the play does. Wallets that played more than 100 markets were fewer than one in five of the crowd, and between them they carried about three fifths of everything the whole crowd lost. A fee charged per trade is a tax on frequency, and the people trading most frequently pay most of it.

Who is on either side

The extremes of that week are not ordinary players. The best-performing wallet, 0xb55f…64d4, played more than 6,000 separate markets in 8 days, paid more in fees than any other player, and still finished tens of thousands of dollars ahead. Its account has never been given a display name.

The worst-performing wallet lost more than the best one made. It trades under the name uninformedfloww, on an account opened in June, and it played more than 15,000 markets that week while paying almost nothing in fees: it rests its orders and waits to be hit, and being hit is what cost it. A third wallet has no Polymarket profile at all, which means it has almost certainly never touched the website. It traded through the API at a rate no person could sustain, paid more than any other wallet in fees, and still edged out a small profit over them.

Not everyone at the extremes is a machine. One wallet lost a five-figure sum that week across just 11 markets, a handful of large, unlucky bets rather than a strategy. Another five-figure loss belongs to an account opened 6 days before the week began.

Almost nobody stayed

The people are the part of this record that has not recovered. February was the month the crowd arrived, drawn by the new 5-minute clock. We took every wallet that bought an up or down share that month and looked for it again in August.

Almost none of them were there. Fewer than four in every hundred of February's players were still buying six months later. The rest stopped, and the volume chart is largely the story of the exchange replacing them with new arrivals, more slowly each month.

The obvious guess is that the losers gave up. What the record shows is stranger. Taking one week of February, sorting every wallet by what it made or lost, and looking for those wallets again in August, the ones still betting cluster at both ends: the biggest winners and the biggest losers came back at almost exactly the same rate, about twice the rate of everyone whose week ended near level.

Some of that is simply that heavy traders stay heavy, since the largest swings belong to the busiest wallets. But win or lose, the direction of the result did not decide who stayed. The players this market keeps are the ones to whom something happened.

What is left is the same machine, smaller and narrower than it was. Almost all of it is bitcoin on a 5-minute clock, still running through the night. It is still the busiest thing on Polymarket by a wide margin, accounting for about three in every five trades matched anywhere on the exchange in August. In money it is roughly a sixth of the business. Hundreds of millions of small, quick bets on a coin flip, taxed on every one.

What is on the chain

Every figure here comes from Bitquery's Polymarket historical data: the decoded trade, resolution and fee records of the contracts below, read from Polygon. The wallets are linked to the public block explorer. The names beside them are the handles their owners chose on Polymarket's own public profiles.