From Imitation to Collusion: Long-run Learning in a Low-Information Environment

We study long-run learning in an experimental Cournot game with no explicit information about the payo function. Subjects see only the quantities and payos of each oligopolist after every period. In line with theoretical predictions and previous experimental ndings, duopolies and triopolies both reach highly competitive levels, with price approaching marginal cost within 50 periods. Using the new ConG software, we extend the horizon to 1,200 periods, far beyond that previously investigated. Already after 100 periods we observe a qualitative change in behavior, and quantity choices start to drop. Without pausing at the Cournot-Nash level quantities continue to drop, eventually reaching almost fully collusive levels in duopolies and often reaching deep into collusive territory for triopolies. Fitted models of individual adjustment suggest that subjects switch from imitation of the most protable rival to other behavior that, intentionally or otherwise, facilitates collusion via eective punishment and forgiveness. Remarkably, subjects never learn the best-reply correspondence of the one-shot game. Our results suggest a new explanation for the emergence of cooperation.