Bayesian learning with variable prior

Abstract

How much can be learned from a noisy signal about the state of the world not only depends on the accuracy of the signal, but also on the distribution of the prior. Therefore, we define a general information system as a tuple consisting of both a signal technology and a prior. In this paper we develop a learning order for general information systems and characterize the order in two different ways: first, in terms of the dispersion of posterior beliefs about state quantiles and, second, in terms of the value of learning for two different classes of decision makers. The first class includes all agents with quasi-linear quantile preferences, and the second class contains all agents with supermodular quantile preferences.