Models, Inattention and Bayesian Updates

We formulate a theory of expectations updating that fits the dynamics of accuracy and disagreement in a new survey dataset where agents can update at any time while observing each other’s expectations. Agents use heterogeneous models and can be inattentive but, when updating, they follow Bayes’ rule and assign homogeneous weights to public information. Our empirical findings suggest that agents do not herd and, despite disagreement, they place high faith in their models, whereas during a crisis they lose this faith and undergo a paradigm shift. Bayesian updating fits the data well, but only in non-crisis years. Furthermore, we empirically evaluate this theory’s relative strengths and weaknesses in both crisis- and non-crisis years vis-a-vis several leading alternatives and find that it fits better on average and in non-crisis years.