What accuracy and calibration actually mean

Any sports analytics platform can quote an accuracy percentage. On its own that number tells you very little. Two models with identical accuracy can be wildly different in how trustworthy they are, and the difference shows up in calibration.

Accuracy: how often the call was right

Accuracy is the share of graded games where the favored side actually won. It is easy to understand and easy to misread. A model that only ever predicts heavy favorites will post a high accuracy figure while telling you nothing you did not already know.

That is why sample size matters as much as the percentage. Twenty games is a fortnight of noise. Several hundred graded games starts to mean something.

Calibration: is the confidence honest?

Calibration asks a sharper question: of every game the model called at 70%, did roughly 70% of them land? If the answer is 85%, the model is underconfident. If it is 55%, the model is overconfident and its high-confidence calls are worth far less than they look.

A well-calibrated model lets you size your own conviction correctly. A poorly calibrated one misleads you most exactly when you lean on it hardest.

Score error: how close the numbers were

For projected scores, EdgeIQ tracks average error — the typical distance between projection and reality. It is the number that tells you how wide to draw your mental range around any projected score.

Where to check EdgeIQ's own record

EdgeIQ publishes two layers of results. The engine's per-league accuracy is measured on held-out games it never saw during training, and this season's games are graded live after being predicted before kickoff. Both are shown with sample sizes on the how-it-works page and in Model Lab, with accuracy by month, error on projected scores, a calibration view comparing stated confidence with realized results. Nothing is filtered to a flattering window.