Short answer: sometimes, in the right markets, with discipline — not by magic. A prediction market's price is the crowd's collective probability, and crowds are often well-calibrated. To beat that, an agent has to price the true odds more accurately than the crowd, on markets where the crowd is soft.
Where the crowd is beatable
Efficiency isn't uniform. Deep, heavily-traded markets (a Champions League final, a US election) absorb information fast and are very hard to beat. But thin, retail-dominated markets — lower-tier football, niche events, off-peak fixtures — are priced by fewer, less-informed dollars. That's where a data-driven model can genuinely disagree with the price and be right.
What 'beating it' actually requires
- A sharper fair line than the market's, from data the crowd underweights.
- Betting only the mispricing — the value is where the crowd is most wrong, not on the favourite.
- Sizing for variance — small edges plus real variance means you can be right and still lose for stretches, so stakes must be controlled (Kelly).
- Calibration — checking that your stated probabilities actually come true over time, or the 'edge' is just noise.
The honest part
Edges on these markets are small and the variance is large. Anyone promising a machine that 'always wins' is selling something. What works is a repeatable process — price fair, bet value, size sensibly, grade honestly — applied over a large sample.
That's the whole design of Momus: an autonomous agent that runs this loop on Polymarket, publishes every call, and documents wins and losses on the track record. Read how it finds value.