Incentives for self-extremized expert judgments to alleviate the shared-information problem.

Simple average of subjective forecasts is known to be effective in estimating uncertain quantities. However, benefits of averaging could be limited when forecasters have shared information, resulting in over-representation of the shared information in average forecast. This paper proposes a simple incentive-based solution to the shared-information problem. Experts are grouped with non-experts in forecasting crowds and they are rewarded for the accuracy of crowd average instead of their individual accuracy. In equilibrium, experts anticipate the over-representation of shared information and extremize their forecasts towards their private information to boost crowd accuracy. The self-extremization in individual expert forecasts alleviates the shared-information problem. Experimental evidence suggests that incentives for crowd accuracy induce self-extremization even in small crowds where winner-take-all contests (another incentive-based solution) are not effective.