What a 70% Win Probability Really Means

A probability is a promise about frequency. If a model gives 100 teams a 70% chance, about 70 of them should win and about 30 should lose. When the 30 losses arrive, the model was not wrong; it would be wrong if 45 of them lost, or if only 10 did.

From a projected margin to a probability

NBA final margins scatter widely around any projection: the typical miss is about 10-11 points even for a good model. We treat the final margin as normally distributed around the projection and fit its spread (about 12 points) on past seasons by maximum likelihood. A team projected to win by 6 then comes out at roughly 69%; by 3, about 60%; by 10, about 80%.

Checking the promise

Calibration is checked by grouping games by the favourite's probability and comparing with how often the favourite actually won. In our four-season backtest the buckets line up within a few percentage points across the range, which is why we show probabilities rather than labels like "strong" or "lock".

Why not just show picks?

• A 52% game and an 85% game are not the same kind of call. Collapsing both into "pick: home team" hides most of the information.
• Scores such as the Brier score reward honest uncertainty and punish overconfidence, so they are the fairest yardstick for comparing models.
• Probabilities make disagreement measurable: over a season you can see whether the model's 60% calls won 60% of the time.

The calibration table from the backtest is on the backtest page, and the live season's record accumulates on the track record.

📚 More Articles