Creating Wise Crowds: Diversity Maintenance Through Incentives

Regents Of The University Of Michigan - Ann Arbor, Ann Arbor MI

Investigators

Abstract

The success of democracies, markets and organizations often depends upon accurate forecasts by collections of individuals. Collective forecasts can involve a handful of people on a board of directors forecasting business trends, hundreds of people in congress making a policy forecast, or millions of people determining prices through a stock market. In each case, more accurate forecasts lead to better outcomes and forecasts that are way off the mark can lead to crises. In this project, the research team will develop a new analytic framework for understanding collective forecasting that combines models from cognitive psychology and mathematics. In this framework, individuals construct models that they use to make forecasts. These models rely on attributes -- components of reality that the individual sees as most relevant in the unfolding future. For example, an individual might see a company's spending on research and development as portending better long run growth. In this approach, different people choose non-overlapping sets of attributes creating a diversity of forecasts. Collective accuracy hinges on a combination of diversity -- seeing different parts of the problem and connections between them -- and sophistication. This approach extends the standard statistical framework in which collective accuracy arises through the cancellation of errors. By unpacking the contributions of individual cognitive depth and collective diversity on forecast accuracy, the framework can be used to examine the knotty problem of how to create incentives to produce collective accuracy. Accurate groups, teams, or societies need individual level sophistication and collective diversity. How much of each will depend on the forecasting task. An effective incentive structure must therefore take into account the complexity of the task. To test incentive structures, the team will develop a suite of forecasting problems with real world features. Ideally, those structures that perform best can be tested in real applications.