The loss of predictive validity that results when moving from the calibration data to tests on holdout data. Do not confuse this with shrinking.