Evaluation
Calibration
Calibration measures whether predicted probabilities match observed frequencies.
How it works
Among cases assigned probability 0.8, about 80 percent should occur in a calibrated system. A model can rank well while giving poorly calibrated probabilities.
A concrete example
Illustrative: of 100 cases scored near 0.7, 68 are positive.
Do not confuse it with Accuracy
calibration evaluates probability reliability; accuracy evaluates final decisions.