I Tried to Copy Every Winner on the Leaderboard

Here is the most repeated piece of advice on the internet about prediction markets: find the people who keep winning, and copy them. On Polymarket you can actually do it — so I tried. I ended up copying no one, because after taking the entire leaderboard apart, wallet by wallet, I could not find one person worth copying.

Polymarket is a prediction market — the same idea as the last story, you buy Yes or No on some future event — with one property that makes it a forensic accountant's dream. Every trade settles on a blockchain, which just means a public ledger that anyone can read and no one can quietly edit. Every bet every account has ever placed is sitting out in the open, permanently, attached to a wallet. Nobody can hide a loss or invent a win. If a trader is genuinely good, the proof is right there. If they're not, that's right there too. This is the rare corner of finance where you can fully audit the person you're thinking of copying — down to the penny, across their entire history.

And Polymarket helpfully publishes a leaderboard: the accounts with the most profit, ranked. The plan writes itself. Read the leaderboard, find the accounts that win and keep winning, mirror their trades, collect a slice of their genius. I built the tooling to reconstruct every one of those accounts from the chain, trade by trade. Then, exactly like the scanner, I looked. And the closer I looked, the more the geniuses evaporated.

The first sign: everyone was losing money

My very first pass added up, for each account, the value of the bets it was currently holding. Almost every account on the leaderboard came out roughly 35% in the red. The most profitable traders on the entire platform were, apparently, all losing money — which is a bit like walking into a casino and finding that the house, the dealers, and the whale at the high-limit table are all somehow broke. When your data says everyone loses, the broken thing is your data.

The mistake was quietly beautiful. On Polymarket, when a bet wins, the payout is collected and the position vanishes from the account — it's been cashed in, so it's gone. When a bet loses, the worthless position just sits there forever. So if you total up only what an account is holding right now, you are adding up its losers and skipping every one of its winners. Every account looks like a catastrophe. Once I rebuilt each account's full history from the chain — winners included — the numbers finally made sense. That one fix moved the entire picture by about thirty-five percentage points, which is a good measure of how confidently a small wrong assumption will lie to your face.

The winners don't stay winners

With honest numbers at last, I asked the only question that matters for copying anyone: do this period's winners keep winning in the next one? If turning a profit here is a skill, the same names should keep floating to the top. So I lined up each account's rank in one stretch of time against its rank in the next.

The relationship was essentially zero — in fact, very slightly negative. Sitting at the top of the leaderboard told you, if anything, a hair less than nothing about whether you'd be there next time. That is not what skill looks like. Skill persists: good is good again next time. What I was looking at was the signature of a fresh coin flip every time.

Ten heads in a row

Think about what a leaderboard actually is. Take a big enough crowd, have everyone flip coins, and someone will flip ten heads in a row. They did nothing — but they're now on top of the coin-flipping leaderboard, and from the outside they look untouchable. A prediction-market leaderboard is exactly that, at scale: with hundreds of thousands of accounts, a handful will win five, eight, ten bets in a row on pure luck, and those lucky few are precisely the ones that rise high enough for you to notice and admire.

And here's what makes a leaderboard measured in dollars even crueler than a coin-flipping one: the luckiest players don't have to be good — they just have to be big. A whale with a deep bankroll needs no edge whatsoever. Open a fresh account, bet large on a couple of near-coin-flips, win one or two of them, and you've booked a five-figure "profit" and vaulted straight onto the board — on nothing but variance and a fat wallet. Bet enough, and one lucky night is indistinguishable from genius.

And that is exactly what the biggest names turned out to be. To tell luck from skill I lean on a standard piece of statistics — a z-score. It counts how many standard deviations an account's win total lands above what the prices it paid should have produced: a textbook measure of how hard a record is to write off as luck. (It's a mouthful, so I'll keep calling it the luck-o-meter — snappier, exact same thing.) The scale is the familiar bell curve. A score around 1 is nothing — a no-skill coin-flipper clears it about one time in six by pure chance. Around 3 — a fluke that shows up less than once in seven hundred tries — is where it starts to look like genuine skill. Below 1 is noise. Run it on the multi-million-dollar whales and most of them land at zero or below: no real edge at all. They bet favorites, they bet volume, they bet size, and they finished ahead on variance and one good tournament. The smaller "geniuses" dissolved the same way — a striking share of the accounts near the top were brand-new, thin on history, with almost all their profit from a single hot run that evaporated the moment the tournament ended. That is not what a room full of skilled traders looks like. It's what you get when you take hundreds of thousands of accounts, let variance do its work, and publish a ranking of whoever happens to be hottest — and richest — at this exact moment.

The consistent ones aren't predicting anything

A few accounts, though, really were persistently, boringly profitable — the luck-o-meter adored them, period after period. These, obviously, were the ones I most wanted to copy. So I looked at what they were doing, and discovered you cannot copy them, because they are not betting on anything.

They're market-makers. Instead of predicting who wins, a market-maker offers to buy at, say, 49 cents and to sell at 50 cents at the very same time, and simply pockets the penny in between — thousands of times a day. It's the "spread" from the last story, that little gap between the buy price and the sell price, except here the gap isn't an obstacle, it's the entire business model. And on top of that penny, Polymarket pays rewards to accounts that keep those offers standing, so the profit isn't a forecast at all — it's rent for standing at the counter. One of these bots ran something like $99M of volume, around the clock, never sleeping, never having an opinion about anything. What makes an account like that money isn't being right — it's being fast, and being paid to stand there. You can't copy it even in principle: half of a market-maker's trades are just the opposite side of its own other trades. (How that whole business works — and why the leaderboard's "records" are the wrong thing to admire — is a story of its own, and it gets one.)

The real ones had an expiry date

That left a small, stubborn handful of accounts that looked like the genuine article — real directional bets, really up, not obviously vending machines. If a copyable genius lived anywhere, it lived here. And these you really could copy: each looked like a distinct, skilled individual with a method of their own, and nothing stops you mirroring a single wallet's trades. The only question that mattered was whether the edge underneath was the kind that keeps paying — so I put the whole pile under a microscope.

A handful of behavioral tells did most of the sorting. The first is the clock: plot every trade by the hour it was placed, and a real person leaves a hole in the chart where they sleep — five or six dead hours, night after night — while a bot trades straight through, 3 a.m. included, forever. The second is pace — how many bets a day: a human lays down a handful, where some of these accounts fired off dozens, and a few of them hundreds. The third I think of as turns per market: how many times an account trades in and out of a single market. A genuine bettor buys once and waits for the game to be played — one turn. A market-making bot buys and sells the very same contract again and again — thirty times, a hundred, sometimes four hundred — because it's skimming the sliver of spread on each round-trip, not backing anyone to win. And the last is timing: the real sharps were diligent in a very particular way, consistently placing their bets in the final hour before kickoff, once the lineups and the late news and the injuries were finally in — they were reacting to information, not to a calendar. The rest just sprayed and prayed — scattering small bets across everything in sight on a fixed schedule, with no late read and no conviction, hoping a couple would land. Every wallet came out the other side classified: human or bot, sharp or schedule-filler, edge or noise. Not one escaped the sorting.

The best of the survivors was a single wallet — for a while it looked like the best trader on the entire platform: a gorgeous record, up around 30%, luck-o-meter comfortably clear of noise. Then I split its profit by event, and 93% of it came from exactly one tournament — the 2026 World Cup. And a World Cup is only a few weeks long. A clean, dominant record over a few weeks is a genuinely slippery thing: it might be a real edge, or it might be a very lucky run — someone flipping heads ten times in the one stretch everyone happened to be watching, and riding it to the top of the board. Over a window that short you cannot fully tell the two apart. And it barely matters which it was, because the Cup ends: copy it while the tournament was on and you'd have done beautifully; copy it the week after and there was nothing there — the markets gone, the edge (or the luck) gone with them. It was still the best of the bunch: every other real-looking name told a version of the same story once measured honestly, a narrow record tied to one hot run, already closing. A wonderful thing to have found, and a terrible thing to bet your savings on continuing.

The most convincing one of all

The single most convincing account on the whole board wasn't that one — it was another, and it fooled me longer than any of them. No one-tournament trick here: a long, patient record spread across a full calendar of games, with the bets placed carefully in the last hour before each one, right as the lineups and the late injuries landed. That is textbook sharp behavior — money moving on fresh information, not on a schedule — and it looked, for all the world, like a disciplined professional quietly grinding out an edge.

Two things took it apart. The first was the shape of the record: it opened with a streak — a run of early wins that launched the account up the rankings, the same way ten heads in a row launches a lucky flipper to the top of a coin-flipping contest. The second was the sizing. Early on, while the bankroll was still small, some of the bets ran a fifth to a third of everything the account had — which is not how you back an edge you trust, it's how you press a hot hand. The early wins compounded, the stack swelled, and from the outside it all looked like conviction being rewarded.

So I ran the same two-step I ran on everyone: the luck-o-meter, and then the luck-o-meter again with the single best streak deleted. It came in around 1.4 — the kind of record a lucky nobody backs into about one time in twelve, already a long way from proof of anything. Strip out that one hot run and it dropped to 0.6: roughly one in four, the sort of thing pure chance coughs up all the time — not a trader at all, just a person flipping coins with real enthusiasm. Take away the lucky opening, and there was no edge left underneath it. The most diligent-looking trader on the entire platform turned out to be a good first streak, pressed hard and dressed in the costume of discipline.

One footnote, because it is too good to leave out. When I traced where these "independent" stars got their money, the trails all ran back to the same place: a single fund quietly staking around 45 wallets, eight of them up on the board. The leaderboard's "diverse field of self-made talent" was, underneath, one desk wearing eight name tags. It changes nothing for a copycat — you can still mirror any one of these wallets you like — but it's a lovely piece of trivia, and it gets its own story further down the shelf.

So here is the audit, top to bottom, of the visible peak of the largest prediction market on earth — every trade public, nothing hidden: coin-flippers on hot streaks, vending-machine bots collecting rent a penny at a time, and a few genuine sharps whose edges were real, narrow, and already closing. The number of accounts you could copy and expect to keep winning with was zero.

And it had to come out this way, for the same reason a screenshot of "free money" on Kalshi was never free. A public leaderboard is the single most look-up-able signal that could possibly exist — it is literally a ranking, published and updated live for everyone. The instant a signal is that easy to reference, it has already been referenced by everyone, so what's left sitting on it is either noise wearing a genius costume, or a service you'd have to become a bot to provide. Easy to look up; therefore impossible to profit from just by looking it up. That idea turned out to be the one law sitting under every dead end on this whole site, so I gave it a name — the Referenceability Law — and it held everywhere I pointed it: the easier an edge is to look up or compute, the more certain it is that everyone already has, and the less there is left to take.

Two threads to pull from here: how you actually win a prediction market (it is not win rate), and the full teardown of the syndicate — one fund, forty-five wallets — coming to the shelf.