Calibration

The promise

When we say a player has a 70% chance of something, it should happen about 70% of the time. That is a different thing from being close on average, and it is the one we care about — you are usually choosing between players whose middles are nearly identical, so what separates them is the shape.

What we will publish

Reliability

Bucket every prediction by the probability we stated, then show what actually happened. The 70% bucket should land near 70%.

Where the truth fell in our range

Across thousands of games this should come out flat. A U-shape means our ranges are too narrow; a hump in the middle means too wide.

Score on the whole distribution

Not just the midpoint, so a confidently wrong range is punished.

Nothing published yet

We have no measured results on this page because none have been produced and locked. Putting a number here before then would be exactly the thing this page exists to prevent. When the backtest runs against a held-out season, its results appear here — good or bad.

Why we don't look at other sites' projections

Market projections are deliberately kept away from the model while it is being built. A model tuned toward someone else's numbers cannot beat them, and the damage leaves no trace in the code. After a version passes its own test they become a benchmark, not an input.