Title: A framework for probabilistic inferences from imperfect models

Abstract: The Bayesian paradigm provides a natural way to deal with uncertainty in model selection through assigning each model in a list of models under consideration a posterior probability, with these probabilities providing a basis for inferences or used as weights in model-averaged predictions. Unfortunately, this framework relies on the assumption that one of the models in the list is the true model. When this assumption is violated, the model that is closest in Kullback-Leibler divergence to the true model is often assigned probability converging to one asymptotically. However, when all the models are imperfect, interpretation of posterior model probabilities is unclear. We propose a new approach which relies on evaluating parametric Bayesian models relative to a nonparametric Bayesian reference using Kullback-Leibler divergence. This leads to a new notion of {\em absolute} posterior model probabilities, which can be used to assess the quality of imperfect models. Some properties of this framework are described. We consider an application to linear model selection against a Gaussian process reference, providing simple analytic forms for routine implementation. The framework is illustrated through simulations and applications.