There's a picture everyone has of a Polymarket user. A guy who reads three newspapers before breakfast, has strong opinions about swing states, and puts fifty dollars on an election because he's sure he knows something the polls don't.
He exists. He's just not the one moving the money.
Which is Daniel's question, more or less. He's been looking at the volume numbers on Polymarket, the millions on individual markets, and he wants to know who's actually on the other side of all of it. Is this thousands of small bets adding up, or is something much more professional going on.
And he asked it in three layers, which is what makes it a good question. Who's actually trading and where does the money flow. What liquidity and volume on a prediction market even mean when you look at the figure on the screen. And then the harder one, whether these markets can be manipulated, not by moving the price on the screen but by going after the real-world event the contract depends on.
His phrasing was that he wants to know what happens when the thing being traded is not a stock or a commodity but a prediction about reality itself.
Right. And that last part is where it stops being a market-structure story and becomes something closer to a governance story. Which is a good place to be, because the answer to all three layers is unusually concrete right now. There's a wave of on-chain work from this year, and there's the first actual CFTC insider-trading case involving event contracts.
So where do you want to start.
Start with who trades. The data on this is striking, and there's more than one study now pointing the same direction.
The short version is that Polymarket's user base is overwhelmingly casual, and its volume is overwhelmingly not. Those are two different populations. The people are retail. The money is a few thousand wallets running algorithms.
Define few thousand.
One analysis from March classified users into personas and found the high-frequency, high-capital group is about two percent of users and generates somewhere around eighty-eight to ninety percent of platform volume. They put it as closer to a ninety-two split than an eighty-twenty split.
So the 80/20 rule was too generous to the crowd.
Considerably. And it's not one study being clever with its own taxonomy. A separate analysis in August looked at wallets making more than a thousand trades a month and found they're three to six percent of active wallets but generate eighty-six to ninety percent of all fills. Every month. Not a spike, not one event, the baseline.
What does a thousand trades a month look like as a life.
It doesn't look like a life. It looks like a process running somewhere.
Then there's the third one, which I think is the cleanest version of it.
There's a classification of just over a million wallets, and it splits them into retail, professional, and whale. Retail is eighty-three point seven percent of wallets and fourteen percent of volume. Whales are nine-tenths of one percent of wallets and seventy percent of volume. And the retail bucket is down about a hundred and one million dollars in aggregate.
Fourteen percent. That's the number I'd put on the wall. Everyone you know is in that bucket and it's carrying a seventh of the market.
And the upshot of all three is the same sentence, which the authors of a London Business School and Yale working paper actually wrote out. They said Polymarket's accuracy reflects the wisdom of an informed minority, not the wisdom of the crowd.
That's the whole episode in one line, isn't it. The oracle isn't the crowd. It's three percent of the crowd and everyone else is the weather.
I'd keep that one.
But here's what I want to press on. If the pros are that dominant, and the retail bucket is losing, then the platform isn't really aggregating beliefs. It's harvesting a subsidy.
You're not wrong, and the numbers back you. The Galaxy analysis from the start of this month covered two point nine million retail accounts and found sixty-nine percent of them finished below break-even. The population was down three hundred and thirty-nine million dollars in aggregate.
And the median account?
Down about three dollars.
So the typical retail account isn't ruined. It's just slowly paying a subscription it didn't know it signed up for.
That's exactly the shape of it. The aggregate number is huge because there are millions of people, not because any one of them got cleaned out. And the counterparty on the other side of that subscription is the professional core.
Who's actually profitable, presumably.
Here it gets interesting, because the pros aren't uniformly winning either. One report covering seven hundred and eighty-seven million trades in the first half of this year found bots were the only group net profitable overall, up about a hundred and eight million. But only twenty-seven percent of bot wallets were profitable. The median bot wallet still loses money.
The median bot is a hobbyist with a script.
A hobbyist with a script who is funding the minority of bots that work. And the professional human traders had the highest hit rate of any group, about forty-two percent profitable, and still collectively lost about twenty-six million before fees.
So the humans in the smart bucket lost money too.
Everybody lost money except one narrow slice of the bot population. And the starkest single figure from that report: out of roughly one point seven five million wallets, a hundred and forty-six maintained positive cash flow for six consecutive months.
Out of one point seven five million.
A hundred and forty-six.
I want to sit with the fact that the number of consistently profitable traders on the biggest prediction market in the world is roughly a mid-sized wedding.
And the London Business School paper found essentially the same thing from the other direction. Three point one four percent of accounts qualify as skilled winners, and together with market makers they capture more than thirty percent of all gains while being under three and a half percent of accounts. The sixty-seven percent classified as unskilled losers absorb the platform's entire pool of aggregate losses.
So the platform's total losses and the unskilled bucket's total losses are the same number. That means the pros aren't losing. They're just transferring.
Net of fees, yes. That's what the accounting says.
And yet the product is marketed as a truth machine.
Galaxy's own conclusion, which I think is the most honest line in the whole report, is that nothing in it undermines the case for prediction markets as truth machines. Their framing is that the noise trading is what pays for the forecast.
So retail is the product.
They're the counterparty that makes the professional core's edge convertible into money. Without the noise traders, the informed minority has nobody to be right against.
Then the LeBron James thing makes a certain kind of sense.
Polymarket reportedly pays around fifteen million a year for that partnership. Galaxy's point is that they're spending heavily to attract what its own analysis identifies as the least profitable cohort on the platform.
They're buying inventory.
That's a fair way to describe it. And it's not a scandal, it's just a business. The pros need somebody on the other side.
Alright, so we've got the who. I want the mechanics, because Daniel's second question is what liquidity and volume on that screen actually mean, and I suspect the answer is less obvious than the number suggests.
It's very much less obvious. Polymarket is a hybrid. Orders are matched off-chain through a central limit order book, and settlement happens on-chain on Polygon through smart contracts. The YES and NO shares are conditional tokens, each collateralized one-to-one by a stablecoin. If you hold one YES and one NO, you can redeem the pair for a dollar, and arbitrage keeps that relationship tight.
So the price of a YES share is a probability in cents.
Structurally, yes. A YES at sixty-two cents is the market saying sixty-two percent.
And the liquidity number on the page is what, the depth of the book?
It's roughly how much size is resting in the order book at any moment. And the first thing to know about it is who put it there. Almost all wallets on the platform have taken liquidity, meaning they crossed the spread and hit somebody's resting order. Fewer than half have ever placed a limit order. So the supply side of the book is a very small club.
And the top of that club?
The top thousand makers provide most of the static liquidity. On the taker side it's even more concentrated, the top three-tenths of one percent of takers contribute about sixty-three percent of volume.
So when you place a bet and get filled instantly, you were filled by the same few dozen entities as everybody else.
Almost certainly.
And how do the pros actually make money, because I assume the answer isn't just being smarter than the public.
Partly the spread, partly fees. On fees, there's an explicit formula, and the important part isn't the formula, it's the shape of it. The fee scales with the product of the probability and its complement, so it's highest at fifty-fifty and falls to nothing at the extremes.
So a coin flip is the expensive trade.
That's where the fee's maximum is, and it's deliberately lopsided by category. Crypto pays the most, sports next, politics and finance less, and geopolitics is literally zero fee.
Geopolitics is free.
Deliberately. It's the traffic driver.
And the fee revenue tells you where the volume actually is. Crypto generated just over ninety-six million in gross fees in the first half of this year, which passed sports at about eighty-three million, and that's with crypto having roughly half the volume. Sports and crypto together are ninety-eight percent of all fees.
Which is a bit of a tell. The platform people describe as a politics forecasting tool makes almost none of its money on politics.
Almost all of it is crypto and sports.
Now the volume figures, because Daniel asked what those numbers represent, and I have a suspicion they're not what a reasonable person would assume.
They are not. The commonly cited figures get inflated two to three times by double-counting.
Double-counting how.
When you buy a YES share from someone, two wallets were involved in one transaction. If you count both sides, you've counted one trade twice. And the headline number for a market is often the cumulative matched total rather than the amount actually at risk. There's a set of widely quoted July figures, forty-one billion, seven billion, five billion, and on a one-sided taker basis they're closer to thirteen, three, and two.
So a market advertising a billion dollars in volume has maybe a third of that that anyone actually put on.
Roughly. It's not fraud, it's a convention, and the platform is being consistent about it. But if you're reading the number on a page and inferring that a billion dollars of conviction went into a question, you've inferred about three times too much.
I think that's the single most useful thing for a normal person to take from this. The screen is describing activity, not commitment.
That's well put. Liquidity is depth. Volume is matched flow. Neither one is a measure of how much money the market has collectively got behind an answer.
Which lets us get to the part Daniel was most interested in. He asked specifically about the manipulation cases where somebody takes a position and then tries to change the thing the contract is about, rather than just moving the price on the screen.
Right. And when you look at what's actually been documented, the answer is stranger than the question. The most alarming case isn't insider trading and it isn't somebody bribing an athlete. It's the resolution layer.
Walk me through the Ukraine mineral deal market.
Seven million dollars at stake on a market asking whether Ukraine would agree to a Trump mineral deal before April. In late March, 2025, over about a day, the price on that contract went from nine percent to a hundred percent YES, and the market resolved YES.
And Ukraine had not agreed to a deal.
No deal was signed. Polymarket called it an unprecedented situation and apologized on Discord. They refused refunds, on the basis that it wasn't a market failure.
Say the mechanism, because this is the part most people won't have heard.
Polymarket doesn't resolve ambiguous outcomes itself. It hands disputed resolutions to a decentralized oracle, UMA, where token holders vote on what actually happened. You stake tokens to vote, and if you vote against the majority you lose the stake.
So the incentive is to vote with what you think everyone else will say.
That's the critique. It turns resolution into a coordination problem rather than a truth problem, because the payoff is on being with the herd, not on being right. An equity analyst looking at this described Polymarket's silence on it as deafening and perhaps even inculpatory.
And what did the vote actually look like.
UMA token holders voted with ninety-five percent or greater majority consensus in twenty-six of the last thirty disputed markets. So when a dispute gets escalated, it doesn't produce a contested finding. It produces a landslide.
A landslide is not what a truth-finding process looks like. That's a process where everybody has already agreed what the answer needs to be.
And in the mineral deal case, the wallet that proposed the YES resolution had put two hundred thousand dollars on YES beforehand.
So the person who put the answer forward had a two-hundred-thousand-dollar reason to want that answer.
One of the named bettors denied coordination. Another admitted proposing the YES outcome but said he doesn't win because he cheats. UMA's CEO acknowledged that individuals holding large Polymarket positions in ambiguous markets are generally treated with suspicion, and that the comment sections have a rising spam problem.
Which is a very diplomatic way of saying the evidence threads get flooded.
There's a documented pattern of it. People with positions in ambiguous markets show up in the Discord, mass-post arguments, and build what looks like a consensus before the vote.
Now, insider trading, because that's the case Daniel will have seen and it's a first.
In April, the CFTC charged an active-duty Army master sergeant with insider trading on Polymarket using classified information about an operation to capture Maduro. He bought more than four hundred and thirty-six thousand YES shares on the Maduro question between December thirtieth and January second, and made over four hundred and four thousand dollars.
And the enforcement director's line was that it's the first time the agency has charged insider trading involving event contracts, and the first time it's used the so-called Eddie Murphy Rule.
The rule about misusing government information, from the movie. That's a real genre of enforcement action now, and it's the first one aimed at a prediction market.
So insider trading on event contracts is now a live legal category with a named precedent.
There's also a behavioral signature the researchers flagged, which I find more troubling than the single case. Nearly two thousand accounts that opened shortly before a single event and went dormant immediately after. Those accounts moved prices seven to twelve times more per dollar than skilled traders. And three accounts collectively cleared more than six hundred and thirty thousand dollars betting on Maduro's ouster before the operation was disclosed.
Somebody knew, and the market's own price didn't tell you that somebody knew. The market looked normal right up until it wasn't.
That's the sharpest thing about it. Insider trading doesn't break the price mechanism in a visible way. The price was wrong and it looked exactly like the price being right.
And that gets us to the taxonomy, because I think there's a distinction here that gets run together and shouldn't be.
There's a formal treatment of this from May. It separates two things. Market-price manipulation is moving the contract price without touching the event. Real-world outcome manipulation is affecting the event itself. Throwing the match, bribing the voter, a debate moderator behaving strategically.
For a normal derivative on a stock or a barrel of oil, the second category basically doesn't exist as a venue-level payoff, because the underlying is continuous and there's no single moment where one person decides the outcome.
Event contracts are different precisely because the payoff depends on a real-world event whose resolution is subject to influence.
So the platform is not just a market. It has an exposure that no ordinary exchange has. It's got a soft spot at the point where the future becomes the present.
The paper's argument is that the central integrity risk for prediction markets isn't informational asymmetry in the abstract, it's interference with the event, the benchmark, or the resolution process that determines the contract's value.
Which is a very different threat model than the one regulators are set up for.
It notes that leverage would shift the cost-benefit threshold on outcome manipulation, and that sports and sub-national political markets are where it matters most, because the manipulation cost is low. Large-electorate and macroeconomic markets are bounded below by the cost of actually moving them.
So a governor's race in a small state is a target. A national election is not.
That's the shape of it.
What safeguards exist for that third category.
None specifically. No jurisdiction has rules aimed at real-world outcome manipulation. The taxonomy paper treats it as largely theoretical, because nobody has caught anyone doing it.
Nobody has caught anyone doing it.
Which is not the same as nobody doing it.
And I'm going to be careful about the difference between the two things we've been describing. Insider trading is about information. The UMA problem is about governance of the resolution. And outcome manipulation is about touching the event. Those are three different failure modes and the platform's defenders argue the first one is a legal question, the second is a design question, and the third is basically a category error.
I don't think the third is a category error. The paper's point is precisely that it's a venue-level liability that event markets carry and derivatives don't.
What's the regulatory situation, then, because there's a lot happening at once.
It's a patchwork. The CFTC sent two linked rule proposals to the Office of Management and Budget at the end of September to formally classify event contracts as federal swaps, which would bring prediction markets under the derivatives regime, and that's on a track toward a Supreme Court fight because the operators say these aren't swaps.
And the states?
Missouri's attorney general sent cease-and-desist letters to six operators, including Polymarket, on the theory that this is unlicensed sports wagering. The UK regulator has held entry talks with Kalshi and Polymarket while retail binary options remain banned. And Polymarket was in court in The Hague arguing its contracts are financial products rather than gambling, against a Dutch ban.
The same product is simultaneously a swap, a security, and a lottery ticket, depending on who's asking.
That's where it sits. The legal identity is unresolved, and the operators are clearly trying to get it classified as a financial product because that's a better regulatory neighborhood than gambling.
But here's what nags at me. Even if the CFTC wins and event contracts are swaps, and the whole derivatives rulebook gets pointed at Polymarket, that rulebook has nothing to say about a UMA vote going ninety-five percent one way because the voters expect each other to vote that way. It doesn't have a category for interfering with the real-world event either.
Correct. Swap classification resolves who has jurisdiction. It doesn't resolve what to do about the oracle or about outcome manipulation.
The thing that actually scares me about these markets is going to remain outside the perimeter even after the regulators finish.
That's my read. You'd need a new regulatory concept, something like event integrity, that doesn't currently exist anywhere.
Let me push on one thing, because Daniel asked whether the identity of the liquidity suppliers changes how we should read the predictions. I think it plainly does, and I want to say why in a way that's not just a complaint about rich people.
Go ahead.
A prediction market quote is presented to the public as a belief. Sixty-two cents means the world thinks there's a sixty-two percent chance. But if you trace who's actually holding the other side of the trade, you discover it's a few dozen entities running models, and the price is a model output, not a survey.
It's a market-clearing price. Which is what a price is.
And I'm not saying it's wrong. I'm saying it's a different kind of object than people think they're reading. When a poll says sixty-two percent, you know it's a sample of humans answering a question. When Polymarket says sixty-two cents, that's a machine's estimate of what the future distribution looks like, and the machines are the ones setting it.
That's why the accuracy is real. If the price is set by the people with the best information and the fastest execution, of course it's going to beat a poll of uninformed adults.
The accuracy is genuine and it's because it's not a crowd.
The informed minority sets the price and the crowd subsidizes the venue.
Which leads me to a slightly uncomfortable conclusion about the truth machine framing. If the forecast is as good as it is only because retail noise pays for the pros to play, then the forecast comes with a hidden cost that isn't in the quote.
It's not hidden. It's in the aggregate loss number we already cited.
It's hidden from anyone reading a probability off a website. They see a number and don't see the subsidy.
Fair.
The truth machine line is also doing something else, which is that it launders the question. If the output is truth, then the people losing money behind the output aren't gamblers, they're participants in a public good. That's a very load-bearing piece of marketing.
It's also, partly true. Prediction markets do aggregate information in a way that polls can't. It's just that the aggregation is being done by fewer people than anyone imagines.
Which is the difference between a crowd and a crowd with a professional core. The first one is a poll. The second one is a market, and markets have counterparties.
The counterparties are the crowd, which is a strange sentence but accurate.
I'll take it. Herman, let me ask you the one I can't work out. If someone actually wanted to game this, not by insider trading and not by oracle gaming, but by affecting the underlying event, how would you even do it and would anyone find out?
The honest answer is that we don't have a confirmed case, which is either reassuring or the opposite.
Say more.
The taxonomy paper suggests the markets where it's cheapest to do are sports and small political races. If you want a college basketball outcome to go the wrong way, the cheapest method is talking to a player. If you want a small city council vote to go the wrong way, the cheapest method is talking to a few voters.
On a market with millions of dollars in it, how much does talking to the right person cost?
There's no published number, and I'd be wary of guessing. But the paper's implicit argument is that it can be less than the market's notional value, which is what makes it a rational crime.
If you did it, would it be detectable?
Only if somebody went looking, and only if they could establish the link between the position and the interference. Which is exactly the problem with the whole domain. The surveillance pattern is looking for price anomalies, and a real-world event going slightly the wrong way doesn't create a price anomaly, it creates a resolution. It looks like a market being right.
The snag we don't have a category for is also the one that would be hardest to detect.
That's the sum of it. And the CFTC, the FCA, the Missouri AG, the Dutch court, none of them are set up to look for that.
One more thing before we close this part out, because Daniel asked the classification question directly. Is Polymarket a forecasting tool, a financial market, a gambling venue, or all three?
It's all three and the answer isn't evasion. It's a forecasting tool whose accuracy comes from a professional core. It's a financial marketplace where retail is the product. And it's a gambling venue that regulators can't decide how to classify, which is why operators are fighting to be classified as finance rather than betting.
The classification matters, by the way, not just for the operators. If it's a swap, the derivatives rulebook applies. If it's gambling, sports-betting rules apply. Both are different, neither has anything to say about event integrity.
That's the gap.
The trading data and the oracle case together argue that the gap is where the risk actually lives. Insider trading is real and now has a precedent. Resolution gaming is documented. The thing we'd need to catch is the thing we've got no framework for.
Which is a good place for us to be sitting, because this is the part of the show where we get the other side of it.
