Can You Actually Beat Polymarket? I Spent Weeks Trying Everything.

The Polymarket Series. A three-part investigation plus a follow-up, readable in any order:

1. Part 1
2. Part 2
3. Part 3
4. Part 4

The honest answer arrived at 23:43 one night, when a market resolved against a swarm of AI pundits I’d built and my phone buzzed to tell me they were wrong. They’d been so confident. That was the moment I stopped expecting to win and started expecting to learn something, which is a less fun way to live but a better one to write from.

Here’s the question I couldn’t put down. A prediction market is a price between 0 and 1, and that price is just a crowd betting real money on whether a thing will happen. Ceasefire by month’s end. This candidate wins. The price moves as the crowd changes its mind. So the question is brutally simple: can you, with a computer and some cleverness and ten months of public Polymarket history, see the future even slightly more clearly than that price?

If you can, even a little, it’s worth money. If you can’t, the price already knows everything you do.

I spent weeks finding out. Not with money — with a paper ledger, timestamped predictions I couldn’t quietly edit later, and the most boring statistical discipline I could force myself to apply. Ten independent ways at the problem. I’ll show you all of them.

This is the front door to four pieces. Read them in order or don’t; the ending doesn’t change.

Part 1 — I built an army of AI pundits to beat a prediction market. They lost to a single number. Eight little AI personas, each with a name and a backstory, arguing on a simulated Twitter for thirty rounds before spitting out a probability. It cost 27 cents a run and it had one beautiful night that made +81.74 € on paper. Then I did the thing those Medium articles never do: I checked what the rest of it added up to.

Part 2 — six ways to beat Polymarket that all don’t work. No AI this time. Just the oldest, most famous edges in betting, tested one by one against 648 liquid markets and ten months of history. Longshot bias. Momentum. The thin illiquid corners where mispricing is supposed to hide. Six classic strategies, every number shown, including the one that looked like an 84% signal until I figured out what it actually was.

Part 3 — following the smart money, and other ways to lose it. The last refuge of the hopeful: somewhere on-chain there must be a wallet that just knows. I harvested 157 of the most active ones, split their history down the middle, and asked whether yesterday’s winners were tomorrow’s. Then I went after Dutch-books, crypto rounding bias, and the settlement lag everyone swears is free money.

Part 4: I kept trying. Two more ways to beat it, two more failures. After the series went up, readers asked the obvious follow-ups, so I tried two more. Borrow a sharp sportsbook’s closing line as a signal, and flip sides to become the market maker and farm the spread. One died on a fact you can read straight off the order book: Polymarket barely sells the bet I needed. The other dangled a 4.6% edge in front of me, and I fell for it, right up until a bigger, meaner sample turned it into a 5.5% loss.

I’ll tell you the shape of the ending without spoiling the good bits. Ten approaches. One result. The market kept winning, and it didn’t lose to any of them. Not the swarm. Not the quant edges. Not the smart wallets. Not me. “Efficient” turned out to be the most expensive word I learned all season. You can bet on Polymarket. On average you cannot beat it. The price is already the best guess in the room.

That’s a negative result, and I’ve made my peace with it being one. It’s pre-registered in parts, run with bootstrap CIs and out-of-sample splits and a couple of skeptics paid to tear my own findings apart. The whole point was to be the kind of wrong you can trust.

If that sounds like your idea of a good time — careful work, flat opinions, dead-ends left in — the series is yours to read.

Update (June 2026): A reader asked the obvious follow-up — does this hold across venues? If the same contract trades on both Kalshi and Polymarket, is there a gap between them to arbitrage? The original study only looked inside Polymarket, so I don’t know yet — I’m testing it now, pre-registered, same bar as everything else (a real edge has to clear a confidence interval above zero and the cost to trade it), with results by the end of July. I’ll post what I find right here, whichever way it goes.

Get the next experiment by email. The original three parts are done, but I keep poking at it, and Part 4 is the latest proof of that. One loose end is still open: a cross-venue test, Kalshi against Polymarket, pre-registered and resolving at the end of July. You’ll get the result when I find out, whichever way it falls. New experiments land in the same inbox. No noise, just the work. Sign up here.

Want to check my math? The full reproducible bundle — every script, the analysis harness, and the methodology write-up — is here: the full code + methodology. It ships the code, not a dataset: a one-command collector rebuilds the market data straight from Polymarket’s free public APIs, so you’re not trusting my numbers — you’re regenerating them from source. Pull it, rerun it, and tell me where I’m wrong. That’s the only kind of argument I’ll find persuasive at this point.

Start with Part 1.

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.