Learning by Matching ∗

We study how a continuum of agents learn about disseminated information in a dynamic beauty contest model when they do not observe aggregate variables, such as prices or quantities, but randomly observe each other’s actions. We solve for the market equilibrium and find that the average learning curve is S-shaped: learning is slow initially, intensifies rapidly and finally converges slowly to the truth. We show that increasing public information always slows down learning in the long run. It also reduces welfare if agents are sufficiently patient, even when there is no coordination motive. Lastly, optimal diffusion of information requires that agents “strive to be different”: agents need to be rewarded for choosing actions away from the population average. We’d like to thank, for fruitful discussions and suggestions, George-Marios Angeletos, Raquel Fernandez, Douglas Gale, Christian Hellwig, Boyan Jovanovic, Enrichetta Ravina, Stijn Van Nieuwerburgh, Laura Veldkamp, and Ivan Werning. Stanford University, e-mail: [email protected] Department of Economics, University of California, Los Angeles, e-mail: [email protected] 1