Can a Polymarket AI bot really win 168 trades in a row? A 67-year-old Vermont retiree reportedly made $290,000

A 67-year-old retiree in Vermont is said to have made $290,000 in one year with a trading bot on Polymarket. Posts describing him say he placed 168 trades, won all of them and never had a losing stretch.

The posts say the bot scans 188 markets a day. About 12 clear a confidence filter, and roughly one gets a trade. Anything the model rates below 85% is dropped.

The probability comes from a vote. 31 models each pick an outcome. If 28 agree, that’s treated as about 90%. Fewer than 26 is rejected, which works out to about 84%, just under the cutoff.


Bet size follows the Kelly formula, a standard way to size bets by edge, cut to 15% of full size. Take a market priced at 87 cents where the model says 90%. The formula gives a bet of about 3.5% of the bankroll.

What a win at 90 cents is worth

A contract bought at 90 cents pays 10 cents if it settles at $1. That’s an 11% return on the stake. A loss costs the whole 90 cents.

One miss wipes out roughly nine wins of the same size.

That’s why 168 straight wins would matter so much. It also hints at the bankroll. Using the post’s own formula and the 87-cent example, each trade adds about half a percent to the account.

Compounded over 168 trades, that’s a bit more than doubling. A $290,000 gain would then imply a starting stake near $200,000.

Why the Brier score doesn’t fit

The Brier score measures forecast error. Zero is perfect, and lower is better. The post cites 0.041 across all 168 predictions.

If every forecast was right, each one scores the square of the gap between its stated probability and 100%. At 90% that’s 0.01. At 85% it’s 0.0225. To average 0.041 with every bet winning, the model’s typical confidence would have to be near 80%. That’s below the stated 85% floor.

Either some bets lost, or the confidence figures aren’t what the posts say.

Does the weather comparison hold up?

Weather services do run ensembles, which are many versions of a model started from slightly different data. Agreement among them is a useful signal.

But ensemble members share assumptions and inputs, so they tend to be wrong together. 28 out of 31 agreeing isn’t the same as a 90% chance. Whether the vote is well calibrated can only be shown by testing it against outcomes, and the posts offer no such test.

What the wider data shows

Bloomberg analyzed every Polymarket wallet active since the start of 2025. More than 100,000 accounts lost at least $1,000, almost twice the number that made that much.

Bots do profit. The typical bot averaged 89 trades per active day, against 2.2 for non-bots, and high-volume accounts together earned $131 million, mostly concentrated among 823 users who each netted more than $100,000. A bot that trades about once every two days would be unusually slow by that standard.

Not every bot works. One published post-mortem described a bot that ran for months, placed 104 trades, won 32.69% of them and lost money every month. And on claims of high win rates for Claude-based bots, CryptoBriefing found weak verification and advised treating community figures as preliminary.

Polymarket’s ledger is public. A wallet address would let anyone check each entry price and settlement.