Every day the model produces raw win probabilities. Those numbers are often too extreme early in the season because the sample of games played is still small.
What calibration actually does
We fit a single shrinkage factor (currently around 0.5–0.7) on all previously graded picks. The calibrated probability is:
calibrated = 0.5 + shrink × (raw − 0.5)
This pulls extreme forecasts (95% or 8%) closer to a coin flip while leaving moderate forecasts almost unchanged.
Why we publish the Brier score
The Brier score measures how close our probability forecasts were to the actual outcomes (lower is better). We show both the raw Brier and the calibrated Brier on the record page.
When the calibrated number is lower than the raw number, the shrinkage is helping. When it is higher, we are over-shrinking and may need to relax the factor.
How often we re-fit
We re-run the calibration whenever the graded sample grows by another 20–30 picks. The homepage shows the current fit size ("Calibrated vs 368 graded results") so you always know how much data is behind the shrink value you see on each pick.
Transparency on calibration is part of the same philosophy as publishing every graded result: if the model is getting better (or worse), you should be able to see it in the numbers.