I Said Polymarket Was Efficient. Then I Tried Two More Ways to Break It.

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

I published the verdict a few weeks ago. Ten ways to beat Polymarket, ten failures, one efficient market. A clean no.

Then the comments did what comments do. But did you try X. Most of the X’s I had already tried. Two I hadn’t, and they were good enough to keep me up at night. So I went back in.

I’ll give you the ending now, same as last time, because the suspense was never the point. Still efficient. But one of the two got further into my head than anything in the original ten. It had me, for an afternoon, spending money I had not made.

(Still a paper project. No real money, no real trades, none of this is advice. You already know.)

The first idea: let a sharper market do my thinking

There is a forecaster more efficient than Polymarket, and it has been around far longer: a sharp sportsbook. Pinnacle takes bets from professionals and prices them so well that its closing line is, by reputation, the most accurate number in betting. The plan was simple. Use that line as a signal. When the bookie and the prediction market disagree about who wins a match, trust the bookie, bet the prediction market.

It died on a fact, not a calculation.

Polymarket does not sell the bet I needed. I pulled every single-match market resolving in the next ten days. There were 331 of them. Three were “will this team win.” Three. The other 328 were over/under on total goals, both-teams-to-score, draw-at-halftime, a whole confetti of side bets. The World Cup, the biggest thing on the board, sells only as “who lifts the trophy,” never as “who wins Tuesday’s game.”

You cannot run a horse race when the track only sells hot dogs. The idea didn’t lose. It never got to play. It died on the loading dock while the statistics were still lacing their boots.

I had been here before, in miniature. The calendar-arbitrage “edge” from the first round was the same shape, two things that looked like the same bet and were not. This time the mismatch was total. Sometimes the market structure ends the idea before the math is even invited.

The second idea: stop betting, become the house

This is the one that got me.

Stop predicting. Start providing. A market maker does not forecast the dice, it rents the table. Quote both sides, sell to the eager, buy from the panicked, keep the spread. On a venue full of people guessing, be the one who does not have to guess. Just collect rent.

Polymarket publishes every trade on a public chain, so I could rebuild what a patient maker would have earned: filled at the prices the takers actually paid, then carried to the finish line and settled.

The number came back at plus 4.6 percent. Per dollar provided, held to resolution, with a confidence interval sitting comfortably above zero. I felt it land in my chest. Spread capture is real, I thought, the house always wins, and for one warm afternoon I was the house.

You recognize this feeling, because I confessed to it once already. Sixteen wallets. A correlation of 0.43. The dream is alive.

So I did the only honest thing you can do with a number you have fallen for. I tried to kill it.

Two cuts. The first was toxicity. A real maker does not win every fill, and the fills you do win are the poisoned ones, the trades where someone who knows something is running you over at the exact moment the price moves against you. So I weighted toward the large trades, the ones a maker at the back of the queue actually catches. The second cut was simply more markets, because the 4.6 had come from a small handful, and small handfuls lie.

It flipped. Minus 5.5 percent, the whole confidence interval now below zero. On the biggest, most informed trades it got worse, down past minus 7. In the early markets, negative. In the late markets, negative. And the short-horizon number, the thirty-minute markout, had been sitting at zero the entire time, telling me the truth I did not want to hear: the spread you capture is exactly the spread the smart money takes back off you.

The plus 4.6 was the sixteen wallets again. Same ghost, new costume. A small sample wearing a genius’s coat.

Two more doors, same lock

I owe you a caveat, and I would rather say it than have you catch me. These two did not get the full treatment the original ten got. No pre-registration, no Bonferroni gauntlet, no paid skeptics tearing at the method. They were probes. Fast and cheap.

They did not need more. One died on a fact you can read off the order book in an afternoon. The other died on a sample size, which is the most ordinary cause of death there is for a trading idea, and the one almost nobody writes down, because admitting it is not flattering.

Twelve angles now. The same answer from every one. The price is sharper than the sportsbook I tried to borrow, sharper than the spread I tried to farm, sharper than me on my best day. I keep walking up to this market with a new key, and it keeps being the same door.

Go prove me wrong

Everything is still open. The code, the full method, and a collector that rebuilds the data from Polymarket’s free public APIs, so you are never trusting my spreadsheet, only your own rerun. If there is an edge in here that I strangled by accident, it is sitting in the method, waiting for someone with steadier hands than mine.

One thread is still live. I left a cross-venue test running, Kalshi against Polymarket, pre-registered and resolving at the end of July. I genuinely do not know that answer yet. Subscribe if you want to watch me find out in public, whichever way it falls.

Twelve doors. I am out of keys for now. I still would not bet on the thirteenth.

Keep going: ‹ Part 3  ·  Series home ›

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.