Some do, narrowly, for a while. That is the honest answer after two months of live testing across four exchanges, and it is more useful than either "yes" or "no".
The base rates
- On-chain analysis of over a million Polymarket addresses found about 70 percent net losers, with well under one percent of wallets taking most of the profit.
- Every public "98 percent win rate" bot checked was either unverifiable or a win rate on favourites, which is not an edge.
- Of twelve strategy families tested live or on tick data, three cleared their fee: the late-window favourite, one calibrated model on the 5-minute markets, and a maker seat on the 4-hour markets. The rest lost or were noise.
Where the money actually sits
Three places, and none of them is a smarter forecast. The book priced its own settlement better than any model built against it, on crypto, on weather and on sport.
- Fees. The taker fee is the only sure profit in these markets. A strategy has to be measured against it at the price paid, every trade.
- Latency. The one lever that moved a live result was moving the server. The market reacts to spot in about 16 milliseconds; you will not beat it, but you can stop losing to it.
- Subsidies. Liquidity rewards on long windows pay more per share of forced flow than the adverse selection costs. That is the maker seat, and it is capped by capacity, not capital.
Are they allowed?
Yes, through the official APIs, on every exchange tested. Geography is the real limit: Kalshi is US-only, and Polymarket, Predict.fun and Limitless block the US and, in Polymarket's case, several other countries. The location warning on the API guides reads your country from the CDN and says which exchanges will refuse you.
So should you build one?
Build the logger first. If after a week of your own data the rule you had in mind still clears the fee at the price paid, with the official outcomes, you have something most people asking this question never get: a number.