Polymarket's Best Traders Made Millions. I Checked If You Could Copy Them.

The Polymarket Series  |  Start · Part 1 · Part 2 · Part 3 · Part 4

By now I’d lost to a single number twice.

The AI swarm lost to the market price. The six classic edges I bolted on afterward lost to it too. So I did the thing you do when every system fails. I stopped believing in systems and started believing in people.

Some traders are just good. You know the type. They smell the resolution before it happens. And the dangerous, beautiful thing about Polymarket is that you can watch them do it. Every trade settles on a public blockchain. Every wallet, every size, every price, timestamped and permanent. So I pulled it. Crypto’s transparency, finally useful for something: you don’t have to guess who the smart money is. You read the receipts.

The dream writes itself. Find the wallets that win. Copy them. Let someone else have the talent.

(Paper project, by the way. No real money rode on any of this. Nothing here is financial advice, and you’d be unwell to treat it as such.)

The trap I almost walked straight into

Here’s how you fool yourself. I nearly did it on day one.

You sort every wallet by how much money it has right now. Take the top hundred. Look at their history and, my god, they’re geniuses. Up and to the right, all of them. Clearly the smart money. Clearly you copy it.

Except you picked them because they got rich. You selected for the outcome, then acted amazed by the outcome. Flip a thousand coins enough times and someone lands ten heads in a row, and if you go find that person afterward they look blessed. Spotless past. Tells you nothing about the next flip.

Survivorship bias. Oldest trap in the book, and it wears a very convincing costume when the costume is an on-chain track record.

The only honest test throws profit out as a selection criterion entirely. You pick wallets by activity instead, whoever’s trading a lot, winners and losers alike. Then you cut each wallet’s history in half by time. First half, second half. And you ask the one question that matters: does being good in the first half predict being good in the second?

Real skill persists. Shows up early, shows up late. Luck doesn’t make appointments.

105,885 wallets walk into a bar

The pull was not small. I harvested 600 markets from the training period and watched 105,885 distinct wallets move through them. From that crowd I took the 1,200 most active, chosen for showing up and not for winning, and pulled each one’s complete trade history. Both halves.

After filtering to wallets that were genuinely active and sizeable in both periods, 157 were left. Clean sample. Big enough to believe.

The correlation between a trader’s early return and their later return came back at 0.035.

That’s not a weak signal. That’s the absence of one. Zero in a lab coat. Being good in the first half told you essentially nothing about the second. The people who looked like geniuses in spring were a coin toss by summer.

And here’s the part I have to confess, because it’s the most human moment in the whole project. An early pilot, sixteen wallets, had shown a correlation of 0.43. I saw that number and something in my chest actually moved. Skill is real, I thought. The dream is alive. I started mentally spending the money.

Sixteen wallets. At a real sample size the 0.43 didn’t shrink. It vanished. The gap between 0.43 and 0.035 is the whole reason you don’t trust a result until it’s tired of being tested.

Then it got mean

A correlation near zero just means copying winners won’t help. I assumed, reasonably, that it also wouldn’t hurt. You’d track the market, eat the vig, shrug.

So I took the fifteen best traders from the training period, the actual top of the board, last season’s MVPs, and followed them out into fresh markets they hadn’t touched. If skill were even slightly real, these were the people who had it.

They lost 13.6 percent.

Across $349,158 of bets, the champions dropped $47,590 in money that would have been real. Their out-of-sample return landed at -0.229, with a confidence interval running from -0.47 to -0.01. The whole range sits below zero. Not “we can’t tell.” Below zero, top to bottom. Copying last season’s winners didn’t fail to win. It reliably, measurably lost.

Everyone else, on average, lost about 2.7 percent. Sit with that number a second. The house edge on Polymarket, the vig, the cost of being allowed to play, is almost exactly 2.7 percent. The average trader’s edge over the market is precisely nothing, and they pay the toll to discover it. An efficient market, written in one number, in money.

Seven more doors, all locked

A reasonable person stops here. I am not always a reasonable person.

If skill doesn’t live in the traders, maybe it lives in some corner of the market nobody’s watching. So I sliced the data seven more ways, hunting a pocket. Opening prices against settled prices. Short markets against long ones. Category by category. Liquidity tier by liquidity tier. Every time I found a candidate edge, I handed it to independent skeptics whose only job was to kill it.

They killed all seven.

The closest thing to a finding, and I want to be fair to it because it’s genuinely real, is that opening prices are about twice as noisy as later prices. Makes sense. The market hasn’t woken up, nobody’s priced anything in, it’s a cold start. A true inefficiency lives there.

It’s under two cents wide. And it costs about 2.7 percent to trade. The edge is real and it’s smaller than the door you have to walk through to reach it. You can see the twenty-dollar bill on the sidewalk. Bending down costs twenty-five.

Two more, because I couldn’t help it

Two ideas were still on the bench when I last wrote, and they’re off it now. Both came back the same color as everything else.

The first was a hunch about round numbers. Crypto markets ask whether Bitcoin will be above some price on some date, and a price like $100,000 feels different from $97,300. People should anchor on the round one, lean on it, misprice it. I built a proper volatility model to call the fair value and went looking for the bias. There wasn’t one. Round strikes and ragged strikes were mispriced by the same flat nothing, a quarter of a cent apart. The strategy looked profitable for about an afternoon, until I noticed why: my sample landed in the middle of a crypto bull run, so anything that said “yes, it’ll go up” made money. Betting “yes” on everything, no model required, beat my clever model. And the only genuinely contrarian calls it made, the few times it said the market was too high and bet against the crowd, lost 8 percent. The one place the idea had to stand on its own, it fell over.

The second was patience. When a market is all but decided, the near-certain winner doesn’t quite trade at a dollar. It sits a hair under, because the oracle that settles Polymarket takes its time, and your money is locked up until it does. I wondered if that little discount was free money for anyone willing to wait. It is real. The winner trades under a dollar for about four days, sometimes longer, a genuine settlement shelf you can measure. But the discount is smaller than the 2.7 percent it costs to get in and out. So it isn’t a mispricing. It’s a parking fee. The market is paying you, correctly, for the inconvenience of waiting, and not one cent more. Fair pay, not free money.

What the whole thing actually found

Step back and the series tells one story, told over and over with the props swapped out.

An army of arguing AI personas lost to the price. Six textbook edges lost to the price. The on-chain elite, the literal best traders measured honestly across seasons, didn’t just lose to the price, they lost money, significantly, out of sample. And four more exotic swings, a riskless multi-leg arbitrage, seven desperate re-slices, the round-number bias, the settlement-lag discount, each turned up either nothing or one real inefficiency too small to be worth the toll.

Ten independent angles. Same answer every time, from every one. Polymarket is efficient. The price isn’t merely smart. It’s smarter than the smartest individual humans I could find, tracked across seasons. They couldn’t beat it. Neither could I.

That’s the result. It’s also exactly the result nobody posts, because “I tested the dream rigorously and it isn’t real” doesn’t sell a course or a Discord seat. A clean, honest no has no upsell. It just sits there being true.

Every line of code and the full methodology are open. The data isn’t a file I hand you — a one-command collector rebuilds it from Polymarket’s free public APIs, so you reproduce the whole thing from source instead of trusting my CSV. If there’s an edge in there I strangled by accident, it’s all there in the method, waiting for someone sharper than me. Go prove me wrong. I’d genuinely love it.

I just wouldn’t bet on it.

Keep going: ‹ Part 2  ·  Part 4 ›

Want every number, the methodology, and the code to rebuild this yourself from the public APIs? Get the reproducible code bundle → Or subscribe for the next experiment.