Impact of decision theoretic models in information elicitation

点击翻译按钮获取中文摘要

Prediction markets and scoring rules are used daily to elicit personal probability estimates on the outcome of future events. For example, “Will project B be complete by June 1, 2018?” is one security that may be traded in a prediction market, or have agents report their beliefs in a scoring rule. Prediction markets and scoring rules are used for questions and instances in which there is no historical data, and the only way to forecast the outcome of a future event is to ask experts, workers, or anyone having an opinion, thereby leveraging the “wisdom of the crowd.”When initially proposed, both prediction markets and scoring rules were shown to work for agents that are risk-neutral (are indifferent between getting $50 or $100 with probability 0.5, and nothing otherwise), myopic (do not take future payoffs into account), and rational (are maximizing their total reward). Risk-neutrality, rationality, and myopic behavior are known not to hold in practice, and more realistic models of human decision-making have been proposed and verified in various laboratory and field studies. In the proposed research program, we will characterize the efficacy of prediction markets and scoring rules when humans, agents, are modeled using these more realistic and modern decision models. In particular, we will characterize how well prediction markets and scoring rules aggregate information when participants are cumulative prospect theory agents, when external incentives exist, when participants are ambiguity averse agents, and when agents have hyperbolic discounting. All of these models have been shown to exist in humans in practice, but they have not been considered in prediction markets, and some have not been considered in scoring rules. The results of this research program will be new prediction market and scoring rule mechanisms that improve the accuracy of these probability elicitation methods when faced with agents modeled using the decision models discussed above. The outcomes of the proposed research program will be more accurate forecasts from prediction markets and scoring rules in settings that occur more often in practice than the settings considered when these mechanisms were proposed. This research will further highlight the limitations and benefits of prediction markets and scoring rules. In addition, new information elicitation mechanisms may be developed for the settings considered. Canada stands to benefit from the proposed research as we will work to make the developed methods available in software that may be deployed in Canadian organizations, thereby providing them with a distinct competitive advantage of having more accurate estimates on the likelihood of a future event, leading to faster and more accurate decisions. HQP trained under the developed program may deploy the developed mechanisms or use their analysis and technical skills in industry, education, or public sector throughout Canada.