We provide an axiomatic characterization of Bayesian updating, viewed as a mapping from prior beliefs and new information to posteriors, which is disentangled from any reference to preferences. Bayesian updating is characterized by Non-Innovativeness (events considered impossible in the prior remain impossible in the posterior), Dropping (events contradicted by new evidence are considered impossible in the posterior), and Proportionality (for other events, the posterior simply rescales the prior’s probabilities proportionally). The result clarifies the differences between the normative Bayesian benchmark, alternative models, and actual human behavior.
Insight
is our reward