Learning about the Future and Dynamic Efficiency

摘要

We have analyzed the conditions under which an efficient allocation of resources is implementable in a dynamic private values setting where the designer gradually learns about the distribution of agents' values. Learning by the designer generates indirect informational externalities which may destroy the incentives for truthful revelation. Efficient implementation is possible only if the designer's beliefs do not dramatically change after each new observation. Analogous phenomena occur in a model where the designer is completely informed about the distribution of the agents' values, but these are interdependent. In particular, we can use an insight from Jehiel and Moldovanu (2001) to conclude that dynamic efficient implementation is generically impossible in the model with learning, if values are private and if agents' signals are multidimensional.