Abstract
A statistical model is said to be calibrated if the resulting mean estimatesperfectly match the true means of the underlying responses. Aiming forcalibration is often not achievable in practice as one has to deal with finitesamples of noisy observations. A weaker notion of calibration isauto-calibration. An auto-calibrated model satisfies that the expected value ofthe responses being given the same mean estimate matches this estimate. Testingfor auto-calibration has only been considered recently in the literature and wepropose a new approach based on calibration bands. Calibration bands denote aset of lower and upper bounds such that the probability that the true means liesimultaneously inside those bounds exceeds some given confidence level. Suchbands were constructed by Yang-Barber (2019) for sub-Gaussian distributions.Dimitriadis et al. (2023) then introduced narrower bands for the Bernoullidistribution and we use the same idea in order to extend the construction tothe entire exponential dispersion family that contains for example thebinomial, Poisson, negative binomial, gamma and normal distributions. Moreover,we show that the obtained calibration bands allow us to construct various testsfor calibration and auto-calibration, respectively.