Thompson Sampling : A Behavioral Model of Expectation Formation for Economics

Despite much criticism, economic analyses often use Bayesian belief formation. Yet, widely applicable, empirically founded, non-Bayesian approaches are scarce. I investigate whether Thompson Sampling is a suitable unifying model of belief formation by testing its predictions in a pre-registered laboratory experiment. Thompson Samplers update their beliefs using Bayes' rule, randomly draw from the posterior, and then optimize given their random draws. Contrary to standard Bayesian learning and rational inattention, Thompson Sampling predicts that cross-sectional dispersion increases with objective uncertainty. My data deliver strong causal evidence for this feature. Thompson Sampling also stands out regarding statistical fit and predictive accuracy