What is model calibration?

A model is calibrated when the probabilities it gives match how often things actually happen: of the matches where it says 60%, about 60% come true. Picking the favorite isn't enough; the number has to be right.

To measure it, you group predictions into bands (40% to 50%, 50% to 60%…) and compare each band's average probability with what actually happened. On a chart, a calibrated model sits close to the diagonal.

xG PRO publishes its own calibration, bad months and bad rules included.

See also

• Brier score and log loss
• Expected value (EV) and value bets
• Poisson distribution
• How xG PRO calculates every match