Information Characteristics And Errors In Expectations: Experimental Evidence
University of Warwick – Warwick Business School
Georgia State University – J. Mack Robinson College of Business
Durham Business School
University of Warwick – Warwick Business School
We design an experiment to test the hypothesis that, in violation of Bayes Rule, some people respond more forcefully to the strength of information than to its weight. We provide incentives to motivate effort, use naturally occurring information, and control for risk attitude. We find that the strength-weight bias affects expectations, but that its magnitude is significantly lower than originally reported. Controls for non-linear utility further reduce the bias. Our results suggest that incentive compatibility and controls for risk attitude considerably affect inferences on errors in expectations.
Information Characteristics And Errors In Expectations: Experimental Evidence – Introduction
Behavioral finance explains market anomalies by drawing on evidence from psychology that some people respond to information in a systematically biased manner. However, several studies show that behavioral biases are not always robust when tested in tasks that reward subjects for being accurate. We design an experiment to test a psychological hypothesis related to errors in expectations, and widely cited in finance, first proposed by Griffin and Tversky (1992) (GT).
According to the GT hypothesis information can be broadly characterized along two dimensions: strength and weight. Strength is how saliently the information supports a specific outcome, and weight refers to its predictive validity. GT suggest that, in violation of Bayes Rule, some decision makers pay too much attention to strength and too little attention to weight, thus overreact to high strength, low weight signals, and underreact to low strength and high weight ones. The magnitude of the bias reported by GT is significant, as in some cases probabilities that should be equal under Bayes Rule diverged by 28%.
Because the reported strength-weight can parsimoniously explain both underreaction and overreaction, it received several applications in finance. Barberis, Shleifer, and Vishny (1998) use the GT findings as a basis of a theory that explains several asset pricing anomalies. Liang (2003) and Sorescu and Subrahmanyam (2006) similarly use the GT findings to explain the pricing of earnings surprises and analyst recommendations, respectively. Other finance studies which cite GT to behaviorally explain their findings include Daniel and Titman (2006), Hackbarth (2009), De Dreu and Bikker (2012), Puetz and Ruenzi (2011) and Gupta-Mukherjee (2013).
However, there is tension in the literature whether such behavioral biases are as significant as initially reported in tasks with an incentive compatible reward system. For example, Grether (1980) and Charness, Karni, and Levin (2008) report that violations of Bayes Rule reduce substantially among financially motivated subjects.
We test the strength/weight hypothesis using an incentive compatible design to encourage effort in the experimental tasks. In addition, to avoid confusion that may arise from subjects being asked to imagine signals from a hypothetical process, as GT asked their subjects to, we generate all the relevant information in front of our subjects during the experiment using physical urns and dice. Finally, in our experiment we elicit subjective beliefs using revealed preference, as opposed to the stated preference methods used by GT, which avoids the need for introspection.
Our elicitation methods are based on the principles of subjective probability elicitation initially outlined by Ramsey (1931) and Savage (1954, 1971). Our respondents observed information signals generated by random draws from urns, and chose between bets that varied the payoff they offered if different states of the world were true. From these bets we inferred the underlying subjective probabilities for the different states of nature, and examined whether they are influenced by the strength-weight heuristic.
Because subjects’ choices will depend on both subjective beliefs and preferences, in our estimations we use data from a separate experimental task to control for the distorting effect of the utility function on inferences about subjective beliefs, estimating the relevant parameters using a structural model. We start our analysis assuming risk neutrality, moving on to a Subjective Expected Utility (SEU) specification that allows for non-linear utility. This approach allows us to examine whether inferences on decision heuristics are affected when one relaxes the assumption of risk neutrality, commonly employed in experiments (e.g., Grether (1980)).
See full PDF below.