Probabilities with receipts
A forecast is not a result.
Every NFL probability is recorded before kickoff, linked into an immutable hash chain, and left in the public record after the game. Good weeks and bad weeks stay together.
The measured model
7,276walk-forward test games
10.1%Brier skill vs. base rate
1.1%expected calibration error
272pregame forecasts recorded
These are two different records. The 7,276-game walk-forward test measures the method on history. The public record begins with the first 16 Week 1 forecasts and grows from there. The historical test is evidence, not a promise about the new record.
What 10.1% skill means
It is a Brier skill score: the calibrated probabilities reduced squared probability error by 10.1% compared with a running home-win base rate in the historical walk-forward test. It is not a win rate, a profit figure or a claim that the model is right 10.1% more often.
What 1.1% calibration error means
Across probability bands in the same test, the average gap between forecast probability and observed frequency was about 1.1 percentage points. A low historical gap is useful evidence. It does not prevent a 70% team from losing one game.