Learning Under Ambiguity

This paper considers learning when the distinction between risk and ambiguity (Knightian uncertainty) matters. Working within the framework of recursive multiple-priors utility, the paper formulates a counterpart of the Bayesian model of learning about an uncertain parameter from conditionally i.i.d. signals. Ambiguous signals capture responses to information that cannot be captured by noisy signals. They induce nonmonotonic changes in agent confidence and prevent ambiguity from vanishing in the limit. In a dynamic portfolio choice model, learning about ambiguous returns leads to endogenous stock market participation costs that depend on past market performance. Hedging of ambiguity provides a new reason why the investment horizon matters for portfolio choice.