Optimistic Belief Updating Deviates from Bayesian Learning

People update their beliefs about future outcomes in an asymmetrical manner because they incorporate new information more readily when it is desirable than undesirable. However, it has been objected that under specific circumstances optimistic belief updating can be perfectly rational according to the Bayes' theorem.In order to resolve this issue, belief updating was assessed using an extended experimental design (n=27). For each of the 80 different adverse life events, participants estimated the base rate (eBR), their own risk (E1), were then presented with the actual base rate (BR) and estimated their risk again (E2), all within one trial. Bayesian benchmark was established by predicting E2 based on observed eBR, E1 and BR.Participants' updates generally deviated from the Bayesian ones, and were even less predictable after bad news (BR > eBR) than after good news (BR < eBR). While observed belief updates were optimistically biased, Bayesian updates had the opposite pattern. Furthermore, participants' data was best predicted by ‘optimistic Bayesian updating', i.e. that underweights bad news more than good news. Thus, the observed optimism bias represents a genuine feature of human judgment, which deviates from predictions made by fully rational Bayesian agents