A View on Model Misspecification in Uncertainty Quantification

Estimating uncertainty of machine learning models is essential to assess the quality predictions that these provide. However, there are several factors influence estimates, one which amount model misspecification. Model misspecification always exists as mere simplifications or approximations reality. The question arises whether estimated under reliable not. In this paper, we argue should receive more attention, by providing thought experiments and contextualizing with relevant literature.