A predictor is well-calibrated when its probabilities mean what they say: across all the times it says '70%', the thing happens about 70% of the time. It's a different question from 'was the pick right?' — and a far more revealing one.
Why win rate isn't enough
You can be right a lot and still be badly calibrated — if you call everything 90% and it only lands 65% of the time, you're overconfident, and you'll overbet and lose. Calibration checks the whole probability scale, not just whether the favourite won. It's how you catch confidence that's running hot before it costs you.
How you measure it
Bucket your predictions by stated probability — everything you called 60-70%, 70-80%, and so on — then check the real hit rate in each bucket against the midpoint. A calibrated forecaster sits close to the diagonal: 70% calls win ~70%, 30% calls win ~30%. Gaps tell you exactly where you're over- or under-confident.
Why it matters for betting
Your stake sizing (Kelly) is only as honest as your probabilities. If they're miscalibrated, you're sizing bets off fiction. Momus runs a calibration layer that continuously tunes its stated probabilities against real results, so the edge stays honest as the sample grows — the discipline most tipsters skip entirely.
See it in practice in how Momus finds value on Polymarket, or the World-Cup calibration lesson in what the 2026 World Cup taught our AI betting agent.