Definition
Calibration means your predicted probabilities match what happens in real results. If you label 100 bets as 60% likely to win, about 60 of them should win. The basic check is:
Observed hit rate = wins / total bets
A model is well calibrated when its predicted probability and observed hit rate are close across many similar picks.
Worked Example
A sportsbook posts a team at -110. The break-even probability is:
110 / (110 + 100) = 52.38%
If your model says the team has a 57% chance to win, that is a 4.62 percentage-point edge before accounting for limits, line movement, or execution. To test calibration, group past bets where your model predicted about 57%. If that bucket has 200 bets and 114 wins:
114 / 200 = 57%
That bucket is calibrated. If it won only 100 times:
100 / 200 = 50%
The model overstated the team’s true chance in that range.
Why It Matters
Calibration helps a bettor trust probability estimates before comparing them to market odds. It is most useful when sizing bets, testing a model, or deciding whether an apparent edge is real enough to act on.
