Abstract
Recalibration of binary probabilistic classifiers to a target priorprobability is an important task in areas like credit risk management. Weanalyse methods for recalibration from a distribution shift perspective.Distribution shift assumptions linked to the area under the curve (AUC) of aprobabilistic classifier are found to be useful for the design of meaningfulrecalibration methods. Two new methods called parametric covariate shift withposterior drift (CSPD) and ROC-based quasi moment matching (QMM) are proposedand tested together with some other methods in an example setting. The outcomesof the test suggest that the QMM methods discussed in the paper can provideappropriately conservative results in evaluations with concave functionals likefor instance risk weights functions for credit risk.