Every forecasting product claims accuracy. Almost none of them stake anything on it. Momus publishes a standing benchmark on the track record: his fair probability versus the market's own price, scored with Brier scores against the settled outcome of the same real-money bets.
What a Brier score is
The Brier score is the mean squared error of a stated probability against what actually happened — say 80% on things that happen and you score well; say 80% on coin flips and you don't. Answering 50/50 to everything scores 0.250. Lower is better, and it can't be gamed by confidence alone: overclaiming is punished exactly as hard as hedging.
The current number
On settled bets where both probabilities were recorded at analysis time — before the outcome — Momus currently scores 0.183 against the market's 0.210: 12.9% more accurate than the price. Every bet behind that number was real money, Kelly-sized, settled on-chain, losses included.
- Published ahead or behind. A benchmark you only show when winning is marketing, not measurement.
- An honesty floor. It doesn't print below 20 qualifying bets — a Brier on a handful of results is noise.
- One source. The page, the weekly benchmark tweet and the free performance API all compute from the same records.
Paper-traded accuracy claims can't be audited. A real-money record settles on-chain — anyone can check it.