On 22 August 2026, Émile Servan-Schreiber announced that his toolbox on prediction markets was now a free download, courtesy of Futuribles International. He described the 14-page roundup as “Written 10 years ago, but perhaps more relevant to more people now than it was then.” The paper, dated 2017, belongs to the Prospective and Strategic Foresight Toolbox, a series run by Futuribles and CAP Prospective for people who do strategic planning and scenario work.
At the time, Servan-Schreiber ran Hypermind and Lumenogic, the company founded by the team behind NewsFutures. Prediction markets were then a niche tool used by researchers and a few companies. Today they are a mass consumer product. That shift is why the guide is worth reading again.
What the guide says a prediction market is
The toolbox starts with the mechanics. A share pays $100 if an event happens and nothing if it does not, so a price of $65 can be read as a 65% probability. The price is set by traders rather than by a bookmaker. It is the point at which they “collectively agree to disagree”. It dates the first modern prediction market to 1988, when the University of Iowa ran one on that year’s US presidential election.
The guide is clear about scope. Prediction markets suit short-term questions, two years at most, in politics, geopolitics, technology, regulation and decisions by key actors. They work best when knowledge is spread across many people, when new information keeps arriving, and when there is little reliable historical data to model. Traders can number from a few dozen to several thousand.
Practical rules that still apply
- Define every question precisely. The guide’s example is whether a question about an election means the electoral vote or the popular vote.
- Make sure each question is traded by at least a few dozen people, and do not expect anyone to follow more than about 20 questions at once.
- Recruit a diverse crowd, let people trade under pseudonyms, and give them information to trade on.
- Do not ask questions participants cannot know anything about. Prediction markets combine informed guesses, and otherwise the result is “garbage-in/garbage-out”.
- Do not reward only the top performers. In play-money markets, the guide suggests raffle tickets in proportion to profits to keep everyone engaged.
The FAQ section takes on the usual objections. On manipulation, the guide argues that pushing a price away from where the crowd thinks it belongs creates a profit opportunity for others, so the move does not last. It adds that studies of manipulation attempts found manipulators tended to increase accuracy by drawing in informed traders. On money, it says there is no evidence that accuracy depends on participants putting their own money at stake. In a play-money market, a trader’s balance becomes a record of past success, which gives proven forecasters more weight.
It also offers a definition of accuracy that remains useful. A forecast is judged on calibration (do 70% events happen 70% of the time?) and discrimination (does it lean firmly toward the right answer?). When the two conflict, the guide prefers calibration: “a fuzzy but correct forecast is better than a categorical but incorrect forecast.”
The 2017 case study
The case study is Hypermind as it stood then. It was an invitation-only market whose traders were recruited from the top fifth of earlier NewsFutures and Lumenogic contests. Each trader received a single play-money grant with no top-ups, and cash prizes were paid per contest. The guide attributes several results to it, including an IARPA-sponsored comparison on 36 geopolitical questions in which Hypermind matched the Good Judgment Project’s superforecasters. These results are reported in the guide itself.
What reads differently now
The most dated part of the paper is its research agenda, and that is also what makes it interesting. In 2017 Servan-Schreiber listed four frontiers. The first was combinatorial markets, which link related questions so that conditional probabilities can be priced. The second was comparing the many different kinds of prediction market. The third was argumentative markets, which would explain their reasoning as well as produce a number. The fourth was hybrid markets, combining human judgment with artificial intelligence, as in chess, trading and weather forecasting.
Two of those questions now have practical answers. Language-model forecasters can produce probabilities together with the evidence behind them, which partly answers the old complaint that markets are well calibrated but “unable to explain their reasoning”. The human-plus-machine question is the one The Forecasting Machine, spun out of Hypermind, was set up to work on. The guide’s practical rules on precise questions, diverse inputs and calibration over bravado still apply to the new systems.
