On 28 September 2026, Émile Servan-Schreiber, who has run commercial prediction markets since 2000 and co-founded The Forecasting Machine, announced he would speak later that week at a seminar organised by MIT’s Dražen Prelec, the inventor of the Bayesian Truth Serum. He described it as touching his three favourite subjects: crowd forecasting, AI forecasting and forecasting with the Bayesian Truth Serum. He posted the full abstract of his talk, “Lessons from 25 years of commercial crowd forecasting”.
The seminar is the workshop of the Lürssen Foundation Principal Fellowship Programme 2026, titled “Future of Forecasting: Collective and Artificial Intelligence”. The event week runs from 26 September to 3 October. It has three parts: a two-day workshop at Terminal in Rijeka on 1 and 2 October, research by a core group of fellows, and public colloquia in Rijeka and Bremen.
An early announcement on 20 May framed the workshop as a look beyond prediction markets. It argued that crowd-forecasting tournaments, the Bayesian Truth Serum and AI forecasting had caught up with markets as ways to quantify uncertainty, and said the Prelec workshop would try to move that broader forecasting frontier forward.
Who is in the room
The programme’s permanent fellows are Prelec, a professor of management at MIT Sloan, and Danica Mijović-Prelec, co-founder of the Sloan Neuroeconomics Laboratory. The invited fellows are Rava da Silveira (Basel and Zurich), Steve Fleming (UCL), Yonatan Loewenstein (Hebrew University), John McCoy (Wharton) and Servan-Schreiber, listed under the School of Collective Intelligence at Mohammed VI Polytechnic University. Themistoklis Sapsis of MIT gives the keynote.
The workshop is built around three core questions: whether aggregated judgments can beat individual expertise; how AI can “democratize forecasting tools while potentially narrowing the diversity of viewpoints”; and how asking people to predict others’ beliefs can improve aggregation. The first session, on Thursday 1 October, is titled “Wisdom of Crowds: Forecasting and Collective Intelligence”. Prelec opens it with the Bayesian Truth Serum, McCoy follows with elicitation and aggregation, and Servan-Schreiber’s talk is scheduled for 10:15.
The four lessons
The abstract makes four claims. This article reports what it says the talk would argue. It does not report on how the talk went.
- Accurate crowd forecasting does not require real-money wagering. The abstract says Hypermind’s play-money market matches Polymarket, and that the Good Judgment Project and Glimt produced useful forecasts without any wagers.
- Brier scores are a poor way to explain accuracy to decision makers. The talk promises alternatives.
- The Bayesian Truth Serum makes long-term forecasts, and forecasts that are hard to resolve, possible. Its value as a way of combining probability forecasts is still unproven.
- AI forecasting is about to transform practice through confidentiality, speed, cost, rolling forecast windows and narrow topics. The remaining difficulty is connecting how decision makers think about uncertainty with well-posed, resolvable questions.
Money, Brier scores and truth serums
The first lesson draws on an analysis Hypermind published in July 2026. It covers 1,141 play-money markets resolved between 2014 and 2026, compared with Polymarket and Kalshi. The report finds the platforms broadly at parity (the full report and data are public). The Good Judgment Project won a US intelligence forecasting tournament in the early 2010s by polling forecasters rather than running a market. Glimt, run by the Swedish Defence Research Agency, aggregates public forecasts to support Ukrainian decision-making.
The second lesson is about communication rather than measurement. A Brier score is the average squared gap between the probabilities given and what happened, with each outcome counted as 1 if it occurred and 0 if not. Lower is better, and 0 is perfect. It is a useful scoring rule, but a number like 0.16 tells a minister or a board little about whether to trust a forecast. Servan-Schreiber’s argument, in the abstract, is that forecasters need other ways to show accuracy.
The third lesson is about Prelec’s own invention. In his 2004 Science paper, respondents give their own answer and also predict how others will answer. Answers that turn out more common than the group predicted score higher. Because it does not need a resolved outcome, the method suits questions that will not be settled for years, or ever. The abstract says that its value for combining probability forecasts has not yet been established.
Where AI fits
The fourth lesson is the territory The Forecasting Machine works in. According to the abstract, AI forecasters can be confidential, fast and cheap. They can update forecasts on a rolling basis and cover topics too narrow to attract a human crowd. What is left is translation: turning a decision maker’s worry into a set of MECE questions. MECE means mutually exclusive and collectively exhaustive, so that exactly one answer comes true and each question can be clearly resolved.
Commenting on the announcement, Benjamin Schatter wrote that turning a vague worry into a well-posed question is probably the hardest part. Other workshop sessions touch the same issues: Fleming on metacognition in humans and machines, a talk on peer predictions, and one on “forecasting monoculture” when markets forecast markets.
