Uncertainty Aversion in Experiential Learning

Abstract

Most theories and experimental studies of choice under uncertainty focus on choices between alternatives with known probability distributions. In many real-life contexts, however, the distribution of payoffs is not known, but have to be estimated from observations of outcomes. March (1996) illustrated how such experiential learning could produce seemingly risk-averse behavior even among risk-neutral individuals. This paper formally demonstrates that a large class of learning procedures produces seemingly risk-averse behavior in situations where information about an uncertain alternative can only be obtained by choosing it. Moreover, I show that such risk sensitivity is actually an adaptive response to the trade-off in sequential choice under uncertainty between exploration and exploitation. In particular, uncertainty aversion follows even if a risk neutral decision maker would follow an optimal policy of experimentation in the socalled one-armed bandit problem. These results illustrate the implications of the costs of learning about unknown and uncertain alternatives and suggest a different interpretation of observations of a preference for a certain alternative over an uncertain alternative. They also have implications for the rationality of positive hypothesis testing strategies. 1