Who Actually Wins on Polymarket

Polymarket is a prediction market: you buy YES or NO contracts on real-world events (will Trump win, will the Fed cut, will it rain in Manhattan tomorrow), prices settle on-chain in USDC (a US-dollar-pegged stablecoin), and contracts pay out $1 if your side resolves true. For about four years its core pitch had one clean sentence in it: trading is free. No taker fee (paid by the trader hitting the order book), no maker fee (paid by the trader resting an order on it), no rake. Just prices and the wisdom of the crowd. In January 2026 they started charging takers.^[1] By March 30 the fee schedule was tiered: peak effective rates of 0.75% on sports, 1.00% on politics and finance, 1.80% on crypto.^[2] Geopolitics stayed at zero, which is about the only place the original pitch survived.

The fees fund what Polymarket calls the Maker Rebates Program, which pays USDC daily to the people posting resting orders on both sides of a market.^[1]The fee curve on the 15-minute crypto markets was specifically designed to slow down latency arbitrage. Those are bots that watch BTC, ETH, and SOL prices on Binance and Coinbase and buy or sell Polymarket's short-dated crypto contracts a fraction of a second before Polymarket's prices catch up.^[3] Plain English: Polymarket figured out that one specific kind of trader was extracting value from another specific kind of trader, and built a tax and a redistribution scheme on top of it.

That is the tell. If the wisdom of the crowd were doing the work here, the fee schedule would be unnecessary. Charging the takers and rebating the makers is what an exchange does when it figures out that one specific kind of trader is extracting value from another specific kind, and decides to take its cut on the way through. The way I've come to think about Polymarket is this: the prices are good, the architecture is genuinely cool, the platform has changed forecasting in real ways, and the “wisdom of the crowd” story it tells about itself is the inverse of what's actually happening on the order book. A small minority does the price discovery. The rest of the user base is the liquidity that minority trades against. The ICE deal in late 2025 was an institutional bet on exactly that arrangement. If you've thought about putting money on the platform, you should know which side of the table you're sitting on.

The 3-vs-97 split

Two studies dropped in April 2026 with the same conclusion from different angles.

The first is a working paper from Roberto Gomez-Cram, Yunhan Guo, and Howard Kung at London Business School, with Theis Ingerslev Jensen at the Yale School of Management.^[4] It analyzes every Polymarket wallet from 2023 through 2025: 1.72 million accounts, 210,322 markets, $13.76 billion in trading volume. They find that roughly 3.14% of accounts generate the bulk of price discovery, the process of pushing market prices toward the actual probability of an event. The abstract is unusually direct for an academic paper, and worth reading in full.^[4]

The second is a separate on-chain analysis by the blockchain analytics handle “defioasis,” covering 86 million Polymarket transactions.^[5] About 70% of trader addresses showed realized losses. The top 0.04% of accounts captured roughly 70% of all realized profits, which works out to around $3.7 billion. About 140 addresses lost more than $1 million each.

Yes, retail loses everywhere. That part isn't news. The Commodity Futures Trading Commission's required broker disclosures show retail forex traders lose at roughly the same rate.^[6] A widely-cited study of Brazilian day traders by Chague, De-Losso, and Giovannetti found that 97% of people who day-traded futures for more than 300 days ended up net losers.^[22] The 70% figure on its own is just what speculation looks like.

What's Polymarket-specific isn't the loss rate. It's how few accounts actually make the prices, combined with a marketing claim that pretends otherwise. The platform doesn't describe itself as speculation against sharps. It describes itself as wisdom of the crowd. The Gomez-Cram paper is the clean math saying those two stories aren't the same.

What edge actually looks like

If 3% of traders are doing the price discovery and the other 97% are funding it, the next obvious question is what the 3% are actually doing. Four edges show up in the on-chain data: combinatorial arbitrage (bots catching cross-market mispricings), market-making for the maker rebate (posting two-sided liquidity all day), vertical specialization (domain experts in a niche), and UMA oracle vote capture (gaming the dispute-resolution layer). Each one has a mirror loser pattern on the other side of the table.

The arbitrage piece is the cleanest case. A team at IMDEA and Northeastern (Saguillo, Ghafouri, Kiffer, and Suarez-Tangil) published a paper in August 2025 showing that arbitrage traders extracted around $40 million from Polymarket between April 2024 and April 2025.^[7]Most of that wasn't Polymarket-vs-Kalshi cross-venue arbitrage, as most coverage assumes. It was intra-Polymarket combinatorialarbitrage, where the same event is split across multiple markets (“will X win?”, “what month will Y happen in?”) and the prices on related markets drift apart. Whoever notices first, inside the median 2.7-second window, pockets the spread. The median speed isn't a human game. It's a bot game, and the people on the losing side of those bots are mostly retail.

Liquidity provision is the second largest. A trader operating as defiance_cr documented scaling from $200 a day on $10,000 of capital to $700–800 a day at peak by running an automated market-making bot that places orders on both sides of low-volatility markets.^[8]The Polymarket rebate formula favors two-sided posting at roughly 3x the one-sided rate, which means the structurally profitable move is to be the person sitting on resting orders, not the person taking liquidity. The losing mirror is exactly what you'd guess. People crossing the spread by reflex are paying the bid-ask the market makers are collecting.

The third edge is vertical specialization, which is what people usually mean when they talk about “edge” on prediction markets. A trader posting under the name Axios runs a 96% win rate on “mention markets,” which resolve on whether a specific word or phrase appears in a public statement.^[9]abeautifulmind specializes in sports. HaileyWelsh in crypto. They're domain experts running their domain expertise. The losing mirror is tourists who have a strong feeling about a topic and assume that feeling is an edge. (My favorite category of this on Polymarket: people trading the popularity of a politician based on how much they personally like the politician.)

The fourth one is the one most coverage skips: UMA oracle vote capture. Polymarket resolves disputed markets through UMA's optimistic oracle. The “optimistic” part means the system assumes a proposed answer is correct unless someone disputes it inside a challenge window; disputed markets get settled by holders of the UMA token, voting weighted by how much UMA they hold. In March 2025, a market on whether Trump and Zelensky would sign a Ukraine rare-earth-minerals deal by March 31 resolved YES against the obvious facts after a single whale voted approximately 5 million UMA tokens across three accounts, about 25% of the vote.^[10]Over $7 million in bets resolved against the honestly-priced answer. Polymarket refused refunds, called it an “unprecedented situation,” and committed to building “systems, monitoring, and more” to prevent recurrence.^[10] A separate ~$160 million Zelensky-suit market saw UMA holders rule No against the apparent media consensus.^[11]Losing mirror: pricing the correct answer and finding out the resolution mechanism doesn't agree with you, with no recourse.

Théo and Domer

Théo, the French whale, is the case study every Polymarket piece leads with. He bet around $80 million on Trump in 2024 across roughly eleven accounts (Fredi9999, PrincessCaro, Theo4, Michie, and others), and walked out with profits reported at $80 to $85 million depending on the source.^[12] His edge was specific: he noticed alternative pollsters were asking “who do your neighbors plan to vote for?” rather than “who are you voting for?” and Trump was overperforming on the neighbor-poll version. He commissioned his own private poll for under $100,000 to confirm. Then he sized the position at the limit of what the market could absorb. Sherwood News reported he held about 25% of the Electoral College Trump contracts and 40% of the popular-vote contracts at peak, large enough that he was “effectively stuck, unable to exit his wagers without crashing the market.”^[13]He told the Free Press, “I had conviction, so I bet a lot of money.”^[14]

Théo gets held up as evidence that you can win on Polymarket. He's better evidence of why you probably can't. His information edge was a private commissioned poll. His position size required eleven accounts to evade per-account limits. Neither of those is replicable. Théo proves the structure. He doesn't prove the alternative to it.

The other named trader worth knowing is Domer. He's based in Las Vegas, played poker professionally in the mid-2000s, started on Intrade in 2007, and is the all-time #1 trader on Polymarket by both volume and profit. More than 5,000 markets traded, roughly $300 million in cumulative volume, and around $700,000 in net profits over three years of play. Last year alone he made nearly $3 million.^[15]

He spends roughly a hundred hours a week on the platform, which I find charming and slightly worrying in equal measure. He talks to a small trusted group about every position before placing it. He admits to losing more bets than he wins but making more money on the ones he wins. And the line that should be on the wall in every serious trader's office:

He admits he's losing slightly more bets than he's winning, and making more money on the ones he wins. He says it's all manual, despite people not believing him. And he frames the core skill as the inverse of being smart: the truly hard part is avoiding losing money on stupid stuff.^[15]

Domer is the rule of the rule. A former poker pro spending a hundred hours a week researching, talking through every position with a small trusted group, sizing carefully, mostly trading by hand, betting on outcomes he understands better than the rest of the market. If you're not Domer, and you're not Théo, the question worth asking before you fund a Polymarket account is what your business model actually is. It's the same shape as the GameStop rally: retail believed it had coordinated power, and the order-book data told a different story about who was actually setting prices over time.

ICE bought the data feed, not the gambling product

The cleanest way to describe the ICE deal is this: ICE didn't buy a gambling site. It bought a sensor network and started selling the readings to its existing institutional clients. The 70% of users who lose money are the cost of producing the signal. The other 3% are the sensors. ICE is the layer on top, monetizing the data that comes off the order book and pushing it into the same feeds Wall Street already pays for.

The Intercontinental Exchange, owner of the New York Stock Exchange, committed up to $2 billion to Polymarket in October 2025 at a roughly $9 billion post-money valuation.^[16] Jeffrey Sprecher, ICE's chair and CEO, framed the deal in language that would have been considered eccentric on a Wall Street earnings call as recently as 2023. “Our investment blends ICE, the owner of the New York Stock Exchange, which was founded in 1792, with a forward-thinking, revolutionary company pioneering change within the Decentralized Finance space.”^[16]Sprecher later told Markets Media what specifically interested ICE was the on-chain architecture: “onchain banking allows for instantaneous margin calls and trade liquidations 24/7.”^[17] The deal completed with a final $600 million tranche on March 27, 2026.^[18]

On February 11, 2026, ICE launched the Polymarket Signals and Sentiment Tool. It plugs Polymarket-derived sentiment data into the ICE Consolidated Feed, the normalized market-data product that hedge funds and asset managers already pay for. Same plumbing, Polymarket overlay. Chris Edmonds, who runs ICE's Fixed Income and Data Services business, described it like this: “we're able to normalize the data to help our customers convert the wisdom of those markets into signals that they can use to develop alpha generation strategies and manage risk.”^[19] The consumer Polymarket markets itself as the wisdom of the crowd. The institutional pitch is “the wisdom of those markets,” which, given what the on-chain data shows about who actually makes those prices, is the more honest version of the same idea.

Bloomberg's terminal isn't priced on whether the FX traders providing the underlying tape make money. The same logic applies here. The 70% loss rate is sort of orthogonal to the $9 billion valuation, because ICE is monetizing the price tape itself, not the trading flow that produces it. Coplan put the institutional version about as cleanly as you can: “These markets reflect collective expectations on market-moving events in near real time and have quickly emerged as a credible input alongside traditional data sources.”^[19]That's a much narrower claim than the consumer-facing one, and the more defensible one. The regulatory thaw that brought Polymarket back to US usersin 2025 is what made this institutional layer possible at all. Without it, ICE doesn't write the check.

The strongest counter, and where it fails

The honest pushback on this whole argument runs as follows. Prediction markets are good. They beat the polls in 2024. The fact that 3% of traders generate the price discovery is a feature, not a bug. Information aggregation doesn't require breadth, just marginal informed trading. Wolfers and Zitzewitz already showed this in 2004.

That's largely correct. Polymarket's Trump-vs-Harris pricing in October 2024 sat closer to 60/40 for weeks while polling aggregators were stuck near 50/50. Polymarket genuinely beat the public-polling aggregates on Brier score (a standard measure of probabilistic-forecast accuracy). Robin Hanson's structural argument, that you don't need many informed traders so long as the marginal trade is informed, is essentially the right model. I don't want to undersell that. The 2024 election was a real win for prediction markets as a category.

The pushback fails in two specific places.

The first is the marketing language itself. “Wisdom of the crowd” implies thousands of citizens contributing distributed knowledge. The Gomez-Cram paper says it's a small minority running domain expertise and bots. Those are not the same shape. The second one is closer in spirit to an uncoordinated Goldman Sachs trading desk than to a town hall. Selling option two as option one is the part that's misleading, and the part that the fee schedule and the ICE deal both quietly admit.

The second is the failures, which the bull case glosses over. In May 2025 Polymarket gave Cardinal Robert Prevost a 1–2% probability of becoming pope. Prevost became Pope Leo XIV.^[20]The market completely missed the call, because the small group of cardinals who actually had information weren't on Polymarket. The Ukraine rare-earth and Zelensky-suit markets show what happens when the resolution layer can be bought. Chaos Labs estimated that roughly one-third of Polymarket's 2024 presidential election volume was wash trading, possibly to game the company's airdrop, possibly for other reasons.^[21] None of that destroys the case for prediction markets as a category. Each one chips at the wisdom-of-the-crowd frame specifically.

The honest version of the bull case is what ICE actually said in its institutional pitch: prediction markets are a credible new input alongside traditional data sources. That's a much narrower, much more defensible claim than “the crowd is smart.” And it's the version that the people writing the checks at the institutional layer believe, which tells you something.

What to do with all of this

Three things follow if you're a regular person reading this and considering whether to put real money on Polymarket.

If you're trading recreationally, on small positions, with budget you've already mentally written off, that's fine. It's the same as poker night with friends. Just don't kid yourself about which side of the poker table you're sitting on.

If you're trading because you think you have an edge, audit the edge specifically. Are you faster than the bots running combinatorial arbitrage? Are you running a market-making bot of your own? Do you have domain expertise that beats Axios on mention markets, abeautifulmind on sports, or Domer on whatever Domer's focused on this week? Did you commission a $100,000 private poll? If the answer to all four is no, you are providing liquidity to whoever did. That isn't a moral judgment. It's just an accurate description of what the on-chain data shows is happening.

If you're using Polymarket prices as a forecast input, for a bet, an investment thesis, or a story you're writing, you can. The prices are good. (The same question of where retail edge actually exists comes up in equities, with similar answers.) Just remember what the prices are. They are a sentiment feed produced by a few hundred sophisticated wallets, mostly bots and specialists, monetized by ICE, calibrated by the people taking the other side. They are not the wisdom of the crowd, in the sense the marketing suggests.

The wisdom isn't in the crowd. It's in a few hundred wallets, and you probably aren't one of them.