Eric Stone’s research while affiliated with University of Massachusetts Boston and other places

We report the results of the first large-scale, long-term, experimental test between two crowdsourcing methods: prediction markets and prediction polls. More than 2,400 participants made forecasts on 261 events over two seasons of a geopolitical prediction tournament. Forecasters were randomly assigned to either prediction markets (continuous double auction markets) in which they were ranked based on earnings, or prediction polls in which they submitted probability judgments, independently or in teams, and were ranked based on Brier scores. In both seasons of the tournament, prices from the prediction market were more accurate than the simple mean of forecasts from prediction polls. However, team prediction polls outperformed prediction markets when forecasts were statistically aggregated using temporal decay, differential weighting based on past performance, and recalibration. The biggest advantage of prediction polls was at the beginning of long-duration questions. Results suggest that prediction polls with proper scoring feedback, collaboration features, and statistical aggregation are an attractive alternative to prediction markets for distilling the wisdom of crowds. This paper was accepted by Uri Gneezy, behavioral economics.

We report the results of the first large-scale, long-term, experimental test between two crowdsourcing methods: prediction markets and prediction polls. More than 2,400 participants made forecasts on 261 events over two seasons of a geopolitical prediction tournament. Forecasters were randomly assigned to either prediction markets (continuous double auction markets) in which they were ranked based on earnings, or prediction polls in which they submitted probability judgments, independently or in teams, and were ranked based on Brier scores. In both seasons of the tournament, prices from the prediction market were more accurate than the simple mean of forecasts from prediction polls. However, team prediction polls outperformed prediction markets when forecasts were statistically aggregated using temporal decay, differential weighting based on past performance, and recalibration. The biggest advantage of prediction polls was atthe beginning of long-duration questions. Results suggest that prediction polls with proper scoring feedback, collaboration features, and statistical aggregation are an attractive alternative to prediction markets for distilling the wisdom of crowds.

Across a wide range of tasks, research has shown that people make poor probabilistic predictions of future events. Recently, the U.S. Intelligence Community sponsored a series of forecasting tournaments designed to explore the best strategies for generating accurate subjective probability estimates of geopolitical events. In this article, we describe the winning strategy: culling off top performers each year and assigning them into elite teams of superforecasters. Defying expectations of regression toward the mean 2 years in a row, superforecasters maintained high accuracy across hundreds of questions and a wide array of topics. We find support for four mutually reinforcing explanations of superforecaster performance: (a) cognitive abilities and styles, (b) task-specific skills, (c) motivation and commitment, and (d) enriched environments. These findings suggest that superforecasters are partly discovered and partly created-and that the high-performance incentives of tournaments highlight aspects of human judgment that would not come to light in laboratory paradigms focused on typical performance. © The Author(s) 2015.

This article extends psychological methods and concepts into a domain that is as profoundly consequential as it is poorly understood: intelligence analysis. We report findings from a geopolitical forecasting tournament that assessed the accuracy of more than 150,000 forecasts of 743 participants on 199 events occurring over 2 years. Participants were above average in intelligence and political knowledge relative to the general population. Individual differences in performance emerged, and forecasting skills were surprisingly consistent over time. Key predictors were (a) dispositional variables of cognitive ability, political knowledge, and open-mindedness; (b) situational variables of training in probabilistic reasoning and participation in collaborative teams that shared information and discussed rationales (Mellers, Ungar, et al., 2014); and (c) behavioral variables of deliberation time and frequency of belief updating. We developed a profile of the best forecasters; they were better at inductive reasoning, pattern detection, cognitive flexibility, and open-mindedness. They had greater understanding of geopolitics, training in probabilistic reasoning, and opportunities to succeed in cognitively enriched team environments. Last but not least, they viewed forecasting as a skill that required deliberate practice, sustained effort, and constant monitoring of current affairs.

When aggregating the probability estimates of many individuals to form a consensus probability estimate of an uncertain future event, it is common to combine them using a simple weighted average. Such aggregated probabilities correspond more closely to the real world if they are transformed by pushing them closer to 0 or 1. We explain the need for such transformations in terms of two distorting factors: The first factor is the compression of the probability scale at the two ends, so that random error tends to push the average probability toward 0.5. This effect does not occur for the median forecast, or, arguably, for the mean of the log odds of individual forecasts. The second factor-which affects mean, median, and mean of log odds-is the result of forecasters taking into account their individual ignorance of the total body of information available. Individual confidence in the direction of a probability judgment (high/low) thus fails to take into account the wisdom of crowds that results from combining different evidence available to different judges. We show that the same transformation function can approximately eliminate both distorting effects with different parameters for the mean and the median. And we show how, in principle, use of the median can help distinguish the two effects.

Five university-based research groups competed to recruit forecasters, elicit their predictions, and aggregate those predictions to assign the most accurate probabilities to events in a 2-year geopolitical forecasting tournament. Our group tested and found support for three psychological drivers of accuracy: training, teaming, and tracking. Probability training corrected cognitive biases, encouraged forecasters to use reference classes, and provided forecasters with heuristics, such as averaging when multiple estimates were available. Teaming allowed forecasters to share information and discuss the rationales behind their beliefs. Tracking placed the highest performers (top 2% from Year 1) in elite teams that worked together. Results showed that probability training, team collaboration, and tracking improved both calibration and resolution. Forecasting is often viewed as a statistical problem, but forecasts can be improved with behavioral interventions. Training, teaming, and tracking are psychological interventions that dramatically increased the accuracy of forecasts. Statistical algorithms (reported elsewhere) improved the accuracy of the aggregation. Putting both statistics and psychology to work produced the best forecasts 2 years in a row.

We describe a hybrid forecasting method called marketcast. Marketcasts are based on bid and ask orders from prediction markets, aggregated using techniques associated with survey methods, rather than market matching algorithms. We discuss the process of conversion from market orders to probability estimates, and simple aggregation methods. The performance of marketcasts is compared to a traditional prediction market and a traditional opinion poll. Overall, marketcasts perform approximately as well as prediction markets and opinion poll methods on most questions, and performance is stable across model specifications.