Modeling the change of paradigm: Non-bayesian reactions to unexpected news

Bayes' rule has two well-known limitations: 1) it does not model the reaction to zero-probability events; 2) a sizable empirical evidence documents systematic violations of it. We characterize axiomatically an alternative updating rule, the Hypothesis Testing model. According to it, the agent follows Bayes' rule if she receives information to which she assigned a probability above a threshold. Otherwise, she looks at a prior over priors, updates it using Bayes' rule for secondorder priors, and chooses the prior to which the updated prior over priors assigns the highest likelihood. We also present an application to equilibrium refinement in game theory.