Thompson Sampling: Predicting Behavior in Games and Markets
65 Pages Posted: 31 Oct 2017 Last revised: 29 Nov 2019
Date Written: November 14, 2019
This paper proposes Thompson Sampling as a tractable theory of expectation formation across very different situations. Thompson Sampling accounts for learning behavior and for the fact that dispersion in expectations varies over time and across environments. Thompson Sampling means that agents, having limited information about their environments, update their subjective belief distributions in a Bayesian way, and subsequently make a random draw from the posterior. Conditional on that random draw, agents optimize. Thompson Sampling provides a unifying explanation for several empirical puzzles: (1) human behavior often differs from the predictions of the Nash equilibrium; (2) market volatility varies both over time and across markets; (3) there is much dispersion across individual forecasters in survey data. Thompson Sampling predicts human behavior better across very different datasets than commonly used theories of decision-making in economics such as Nash equilibrium, Bayesian learning with exogenous shocks, and quantal response equilibrium (QRE).
Keywords: Learning, bounded rationality, behavioral game theory, expectations, stochastic choice
JEL Classification: C91, C92, D84, E37
Suggested Citation: Suggested Citation