Part 1 of 3
A trader on a prediction market loses nine bets in ten and finishes up $29,128. The cheap contracts he buys lose money for everyone who buys them without him. This series is the search for what he does differently.
Polymarket is a prediction market. A contract pays one dollar if an event resolves yes and nothing if it resolves no. Prices are quoted in cents, and one cent is one percentage point of probability. A contract trading at six cents is the market saying six in a hundred.
The markets in this story are five-minute Bitcoin Up or Down. Each one asks a single question: will Bitcoin be higher at the end of this five-minute window than it was at the start? Settlement uses a 60-second average of the price, which is why the venue calls it TWAP60. A new window opens every five minutes, roughly 288 times a day, and every trade is public.
That makes it an unusually clean laboratory. The outcome is binary. The payoff is fixed. The experiment repeats all day. Anyone with a laptop can read the tape.
We read it. One wallet traded these markets and almost nothing else.
Over the recent stretch of its life, from mid-July to late October, it bought 4,950 times across 4,707 positions. It won 398 of them.
That is 8.46%. It finished up $29,128.
Most people reading that number would assume a mistake. A wallet that wins less than one bet in ten should be bleeding, and the usual explanation is that the winning bets must be large while the losing ones are small. That is not what is happening. The median entry price is six cents. The median bet is about five dollars of premium. It buys cheap lottery tickets, thousands of them, and comes out ahead.
Here is why that is not automatically luck. If you pay roughly one cent for a contract that pays a dollar, you need to win slightly more than 1% of the time to break even. Across the cheapest band it trades, the wallet paid an average of 0.95 cents and won 2.82% of the time.
It won almost three times as often as the price implied. The 95% interval around that win rate runs from 2.03% to 3.92%, which sits entirely above the 1.0x break-even line. That interval matters: it is the range that accounts for how few winning settlements there are. With 34 wins, a point estimate alone would overstate the precision. The range still clears break-even.
| Entry price | Positions | Win rate | Implied | Ratio |
|---|---|---|---|---|
| under 2 cents | 1,205 | 2.82% | 0.95% | 2.95x |
| 2 to 5 cents | 811 | 3.45% | 3.24% | 1.07x |
| 5 to 10 cents | 1,132 | 7.33% | 7.15% | 1.03x |
| 10 to 20 cents | 781 | 13.57% | 11.84% | 1.15x |
| 20 to 50 cents | 403 | 23.82% | 24.50% | 0.97x |
Three months of one wallet, grouped by the average price it paid per position. The table is the rate, and ratio is realised win rate divided by the implied probability, so 1.0x is fairly priced. The chart is the money on the same five bands: one bar is the whole window, and the other four cancel out.
The whole profit for that stretch is $23,871, and it does not come from five bands in proportion.
Which raises the obvious question.
If cheap contracts are systematically underpriced, we should be able to buy them ourselves. So we did, mechanically. Buy every contract under two cents and hold it to settlement. That is the whole rule.
The rule lost $23,579 across 2,407 contracts.
So the band is not the edge. Something inside the band is, and it is not the price it pays.
We looked for that something wherever a trading edge usually hides.
| We asked | A yes would have meant | What came back |
|---|---|---|
| Who was selling into him? | A block seller to absorb | A small buying crowd, and no seller |
| Was he a maker, working the spread? | Quiet execution rather than prediction | A taker, paying the posted price |
| Did the book predict the next move? | A microstructure edge | Nothing that beat the midpoint |
| Could a flexible model beat the price? | A machine that sees more | None did, and the Bayesian one dropped every extra feature |
Four questions we could put to the tape, and what a yes to each would have meant. None of them paid, which is why the search turned to the wallet itself.
Every model we built said the market was fair. The wallet kept taking money out of it.
So the question became narrower. Not what the wallet knows, but what it does in the last few minutes of a window that we are not measuring.
There are only two ways to be that good. You see the next five minutes better than the crowd does, or you see the same five minutes sooner than the quote does.
Both can be checked against the public tape he was reading. Both were.
That is Part 2.
Part 1 of 3. Windows are stated explicitly throughout: the $29,128 headline covers 4,707 positions from 17 July to 28 October, while the band table covers 4,332 contracts from 23 July to 29 October. The two totals differ because the windows differ, and both are reported as measured.