MODEL CALIBRATION
How to judge a 55% forecast
A calibrated 55% forecast still loses about 45% of the time across comparable observations. One outcome can grade the pick. Evaluating the probability takes a cohort.
Score the same games
The Brier score is the mean of squared probability errors. A 55% probability on a win contributes 0.2025. On a loss, it contributes 0.3025. Log loss penalizes confident errors more heavily.
Watch the denominator
Compare frozen predictions with a same-line market benchmark on the same outcomes. Keep evaluation dates after model fitting, show excluded and unmatched records, and use blocks when games on a slate may be related. Show uncertainty before interpreting a difference.