Superforecasters: Still Crème de la Crème Six Years On
The multi-year geopolitical forecasting tournament sponsored by the research arm of the US Intelligence Community (IARPA) that led to the groundbreaking discovery of “Superforecasters” ended in 2015. Since then, public and private forecasting platforms and wisdom-of-the-crowd techniques have only proliferated. Six years on, are Good Judgment’s Superforecasters still more accurate than a group of regular forecasters? What, if anything, sets their forecasts apart from the forecasts of a large crowd?
A new white paper by Dr. Chris Karvetski, senior data and decision scientist with Good Judgment Inc (GJ Inc), compares six years’ worth of forecasts on the GJ Inc Superforecaster platform and the GJ Open public forecasting platform to answer these questions.
Key takeaway: Superforecasters, while a comparatively small group, are significantly more accurate than their GJ Open forecasting peers. The analysis shows they can forecast outcomes 300 days prior to resolution better than their peers do at 30 days from resolution.
Who are “Superforecasters”?
During the IARPA tournament, Superforecasters routinely placed in the top 2% of accuracy among their peers and were a winning component of the experimental research program of the Good Judgment Project, one of five teams that competed in the initial tournaments. Notably, these elite forecasters were over 30% more accurate than US intelligence analysts forecasting the same events with access to classified information.
Dr. Karvetski’s analysis presented in “Superforecasters: A Decade of Stochastic Dominance” uses forecasting data over a six-year period (2015-2021) on 108 geopolitical forecasting questions that were posted simultaneously on Good Judgment Inc’s Superforecaster platform (available to FutureFirst™ clients) as well as the Good Judgment Open (GJ Open) forecasting platform, an online forecasting platform that allows anyone to sign up, make forecasts, and track their accuracy over time and against their peers.
The data showed:
- Despite being relatively small in number, the Superforecasters are much more prolific, and make almost four times more forecasts per question versus GJ Open forecasters.
- They are also much more likely to update their beliefs via small, incremental changes to their forecast.
- Based on the Superforecasters’ daily average error scores, they are 35.9% more accurate than their GJ Open counterparts.
- Aggregation has a notably larger effect on GJ Open forecasters; yet, the Superforecaster aggregate forecasts are, on average, 25.1% more accurate than the aggregate forecasts using GJ Open forecasts.
- The average error score for GJ Open forecasters at 30 days from resolution is larger than any of the average error scores of Superforecasters on any day up to 300 days prior to resolution.
- GJ Open forecasters, in general, were over-confident in their forecasts. The Superforecasters, in contrast, are 79% better calibrated. “This implies a forecast from Superforecasters can be taken at its probabilistic face value,” Dr. Karvetski explains.
- Finally, the amount of between-forecaster noise is minimal, implying the Superforecasters are better at translating the variety of different signals into a numeric estimate of chance.
You can read the full paper here.
Where Can I Learn More About Superforecasting?
Subscription to FutureFirst, Good Judgment’s exclusive monitoring tool, gives clients 24/7 access to Superforecasters’ forecasts to help companies and organizations quantify risk, improve judgment, and make better decisions about future events.
Our Superforecasting workshops incorporate Good Judgment research findings and practical Superforecaster know-how. Learn more about private workshops, tailored to the needs of your organization, or public workshops that we offer.