On 20 July 2017, Émile Servan-Schreiber announced that Lumenogic, the company that ran the Hypermind prediction market, was joining a new research programme of the US Intelligence Advanced Research Projects Activity (IARPA). Its goal was to find out how human judgment and machine models could be combined to forecast geopolitical events. Lumenogic was part of a consortium led by Raytheon BBN Technologies, one of three teams selected to compete.
To illustrate the idea, he shared a 1997 cartoon by Philippe Andrieu: a small figure cranking a hand-built machine, a dog, and a giant mechanical digger, all after the same buried bone. For Servan-Schreiber, it showed how a mix of different kinds of intelligence can solve a problem together.
What IARPA wanted
IARPA funds high-risk research for the US intelligence community. Its programme page for the Hybrid Forecasting Competition (HFC), run by programme manager Seth Goldstein, set out the problem. Human forecasts can suffer from cognitive biases and do not scale easily. Statistical and computational models scale well but are often poorly suited to unusual or newly emerging questions. A hybrid system might keep the strengths of each and offset their weaknesses. The aim was to help the intelligence community “radically improve the accuracy and timeliness of geopolitical forecasts”.
When IARPA announced the awards on 18 August 2017, Intelligence Community News reported that the four-year programme would cover topics such as foreign elections, interstate conflict, disease outbreaks and economic indicators. The questions were to become more specific and arrive at a faster pace over the course of the programme. MITRE, Good Judgment Inc and Cultivate Labs would test the systems independently, using thousands of volunteer forecasters.
The three teams
IARPA awarded contracts to teams led by HRL Laboratories, Raytheon BBN Technologies and the University of Southern California. Raytheon BBN’s subcontractors included Lumenogic, Wright State, Tufts University and Ipsos Public Affairs. In a July 2018 release, Raytheon said its contract was worth $14.5 million. The principal investigator was Lance Ramshaw, and Lumenogic and the Wright State Research Institute contributed to the technology research.
In a January 2018 post, Servan-Schreiber described the competition from his side. The three teams’ hybrid systems would be judged on thousands of real geopolitical questions over four years. He listed the rivals as including people from Harvard, MIT, HRL and Microsoft Research on one team, and USC, Stanford and Columbia on the other. That list is his own; IARPA’s published subcontractor list differs in its details.
What Lumenogic brought
Lumenogic’s contribution was its experience in collective forecasting, which pools the judgments of many people into a single probability. It had run two prediction markets: NewsFutures from 2000 to 2010, and Hypermind from 2014. In a prediction market, participants trade contracts on future events, and the trading price can be read as the crowd’s estimate of the probability. From 2012 to 2015, Lumenogic had also worked with the Good Judgment team in IARPA’s earlier forecasting tournament, the Aggregative Contingent Estimation (ACE) programme.
That collaboration produced a 2017 paper in Management Science, co-authored by Servan-Schreiber with Pavel Atanasov, Philip Tetlock, Barbara Mellers and others. It compared prediction markets with “prediction polls”, in which forecasters give probabilities directly and their answers are combined statistically. Prediction polls are the approach Hypermind later built into its Prescience platform.
- Human forecasts handle new and unusual questions well but are slow to gather and prone to bias.
- Machine models process large amounts of data quickly but struggle when there is little history to learn from.
- Hybrid systems try to feed one into the other, for example by showing forecasters model outputs, or by weighting human and machine inputs according to how accurate each has been.
What came of it
The Raytheon BBN team’s system was called Prescience. According to Hypermind’s own history, the IARPA work led to the development of its Prescience crowd-forecasting platform. In 2018, Lumenogic regrouped all its activities under the Hypermind brand. Prescience has since been used by the Johns Hopkins Center for Health Security to forecast infectious disease outbreaks from 2019, by EDF for in-house projects, and in 2025 by Glimt, a Swedish government crowd-forecasting platform supporting Ukraine.
In January 2020, Servan-Schreiber wrote that he and Maurice Balick had just returned from Washington, where they presented Hypermind’s latest research results on combining artificial and collective intelligence for intelligence-agency forecasts. We have not found published accuracy results for the Raytheon BBN team, so this article makes no claim about how its system ranked.
