Six Ways to Beat Polymarket That All Don't Work

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

Last time, I built an eight-headed AI swarm to trade a prediction market. It lost to a single number: the price already sitting on the screen. Embarrassing. The lesson I took away was that I’d brought imagination to a problem that wanted arithmetic.

So I put the personas in a drawer and did the boring thing. I went and got the data. Six hundred and forty-eight resolved Polymarket markets, every one with at least ten thousand dollars of real volume behind it, from September 2025 through July 2026. Liquid enough that the price meant something. Resolved, so I knew the answer. Then I lined up the six oldest, most famous edges in betting and pointed them at the wall one at a time.

This is a paper project. No real money rode on any of it. Nothing here is financial advice, and if you treat it as such, that’s between you and your wallet.

What follows is six autopsies.

First, the body on the table

Before any clever strategy, one number rearranged how I thought about everything else. Of those 648 markets, only 16.82 percent resolved YES. Roughly one in six. The confidence interval runs from about 13.9 to 19.8 percent, so it isn’t a quirk of my sample.

That’s not random noise. Polymarket questions are overwhelmingly shaped like “will this dramatic thing happen by this date.” Will the ceasefire collapse. Will the coin hit the number. Will the resignation come. Dramatic thresholds mostly don’t get crossed, because that’s what makes them dramatic in the first place. So the house favorite, the thing you’d win by betting against the headline, is NO. Loud, even. Remember that. It comes back.

Autopsy one: the longshot that wasn’t overpriced

The most famous edge in all of gambling is the favorite-longshot bias. At the track, people overpay for the 50-to-1 nag because a tiny shot at a big payout feels good, and they underpay for the favorite because winning two bucks on a one-dollar bet feels like nothing. Longshots get systematically overpriced. People have been writing this down for ninety years.

So I found the longshots — everything priced at 0.20 or below — and checked. The market said those events should win about 0.73 percent of the time. They actually won 1.15 percent of the time.

Read that again. The longshots didn’t win less than their price implied. They won slightly more. If anything they were a hair underpriced, the exact opposite of the track. I tried betting NO on them anyway, sweeping the threshold across 0.05, 0.10, 0.15, 0.20. The profit-and-loss confidence interval sat at minus 1.5 to plus 0.4 percent. Straddling zero like a fence. No edge to find, because the edge wasn’t there.

Autopsy two: the prices are just honest

This one died of good health. I bucketed every market by price and asked the rude question: when Polymarket says 90 percent, does the thing happen 90 percent of the time?

In the 0.90-to-1.00 bucket, eighty-five markets, the average price was 0.982. The actual win rate was 0.9765. That gap is nothing. The crowd quoting those odds was telling the truth to within half a percentage point. You can’t make money off a forecaster who’s already right.

Autopsy three: momentum, and the mirage that fooled me for an afternoon

This is the one that got my heart rate up.

The idea is ancient: things in motion stay in motion. If a price is drifting toward YES, ride it. I measured price drift against the final outcome and the number came back 0.8386. Eighty-four percent. Momentum picked the winner five times out of six. I sat there a good while feeling like I’d found the door in the wall.

Then I looked at when I was measuring it. I’d taken the drift near the end, right before resolution. But that’s exactly when a prediction market does the one thing it’s built to do: converge. As the answer becomes obvious, the price walks to 0 or 1 mechanically, dragged there by reality, not by any signal you could trade. Of course late drift “predicts” the outcome. The outcome was already happening.

So I rebuilt it honestly. Take an early price move, before anyone knows anything, and use it to predict the end. That number was 0.541. A coin flip with a slight limp. The beautiful 84 percent was real and completely untradeable. You can’t bet on a convergence that only exists once the result is already in.

I keep this one around as a reminder. The most exciting result is usually the one where you forgot to check your own clock.

Autopsy four: heavy favorites, already spoken for

If longshots were a bust, maybe the money was at the other end, the near-locks at 0.90 and up. Surely a 95-cent favorite that pays out is free money.

It is not. After you pay the price, the expected value on heavy favorites came to minus 0.6 percent. You’re buying a dollar of probability for a dollar and change. The crowd already did this math. Everything that looks safe is priced like it’s safe.

Autopsy five: the thin markets weren’t sloppy either

Here’s where I was sure I had them. Big liquid markets are efficient because everyone’s watching. But the thin ones, the dusty corner with five hundred to ten thousand dollars of volume, nobody serious is in there, so the prices should be lazy and wrong.

I pulled 290 of those thin markets and scored them. Brier score 0.022. The liquid markets scored 0.024. The illiquid ones were, if anything, a microscopic hair sharper. There was one twitch that looked like blood in the water: among thin heavy-favorite markets, twenty-eight out of twenty-eight resolved the favored way. A perfect streak. But it’s twenty-eight bets in a market thin enough that one motivated trader explains the whole thing. That’s not an edge. That’s a small number wearing a costume.

Autopsy six: the calendar arbitrage that was my own bug

This one I have to own.

“Trump does X by March” and “Trump does X by June” are logically chained. The March market can never be worth more than the June one, because anything that happens by March also happened by June. When that relationship breaks, you lock a spread risk-free. Pure arbitrage. I ran it across roughly 2,100 live markets, hungry.

Every hit was a false alarm. My matching code had cheerfully paired markets that shared a few keywords but not the actual thing: different thresholds, different opponents, different events entirely. I wasn’t finding arbitrage. I was finding my own sloppy string comparison. Once I tightened the matching, the “opportunities” evaporated. The market wasn’t broken. My script was.

A seventh thing, briefly, that wasn’t even a strategy

While I was in this mood I audited a public GitHub repo advertising Polymarket “trading bot strategies.” Ten of them. Zero were net-profitable once you did the accounting honestly. The impressive returns in the README came from forgetting to subtract costs. The orders the bot generated weren’t even cryptographically signed, which means Polymarket would have rejected them outright, and the fallback path was broken too. I’m critiquing the code, not the author. The verdict is just plain: don’t run it.

Why every door was locked

Run forty-some statistical tests and chance alone hands you a few exciting false positives. So I held everything to a stricter bar than my excitement wanted, the kind of correction you apply when you know you’ve taken that many swings. Nothing cleared it. Not one of the six.

And the pattern underneath is the same every time. Every one of these edges is famous because it genuinely worked somewhere else: the horse track, the sportsbook, the old inefficient corners of finance. They’re famous because they worked, which is exactly why they don’t anymore. Famous edges get crowded, and a crowded edge is just the new fair price. Polymarket already ate all six. The crowd had been there before me and left nothing on the plate.

Which finally forced the obvious move. If the prices are this honest, stop staring at prices. Start staring at people. Specifically, the handful of wallets that had quietly turned this market into millions of dollars, not by finding a famous edge, but by being a different kind of animal entirely.

That’s next.

Keep going: ‹ Part 1  ·  Part 3 ›

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.