On 3 August 2026, Émile Servan-Schreiber shared an analysis of every market Hypermind has resolved since it opened in May 2014. He linked the full report and data, dated 28 July. The question is whether a prediction market where nobody stakes their own money can price the future as well as Polymarket and Kalshi, the two big real-money platforms. The report’s answer is that it can, within the limits of the comparison.
What Hypermind measured
The dataset covers 1,141 markets with a clean winner-take-all resolution, from May 2014 to July 2026. Of these, 694 are binary and 447 have more than two outcomes, and together they drew 981,180 trades. Seventy per cent are about elections and geopolitics, 21% about the economy, and the rest about science, technology, health and sports. Every trader starts with the same play-money endowment and gets no refills. Small yearly cash prizes are paid in proportion to play-money profit, €72,000 in total across the markets analysed. The trade-level data are published as CSV files.
Two measures matter. A Brier score is the squared distance between a forecast and the outcome, so lower is better. For a market with several outcomes it runs from 0 (perfect) to 2. Calibration asks whether events priced at 70% happen about 70% of the time. The report summarises it as mean absolute calibration error (MACE): the average gap, in percentage points, between predicted and observed frequencies.
- All 1,141 markets: Brier 0.322 against 0.580 for chance, a 45.6% improvement. The eventual winner was the favourite for more than half the market’s life in 77.7% of markets.
- Binary markets: 0.249 against 0.500 for chance, with 84.1% correctly forecast.
- Multi-outcome markets: 0.435 against 0.703 for chance, with 67.6% correctly forecast.
- Calibration error in 5-point bins: about 1 percentage point.
The report checks for a status-quo shortcut. Of the binary markets, 76% were yes/no questions, and 62% of those resolved “no”. Always answering “no” would have scored 0.767, far worse than chance. On that subset Hypermind scored 0.222.
The comparison with real-money markets
Hypermind does not have Polymarket’s or Kalshi’s trade data. It uses Brier.fyi, an independent open-source project, which scores 259 Polymarket and 183 Kalshi markets at their midpoint using the single-outcome Brier score. The mixes differ: Kalshi is reported to be mostly sports by volume, while Polymarket has historically been split between politics, crypto and sports. The report therefore re-weights each platform’s category scores to Hypermind’s own mix.
- Polymarket: 0.165 on its own mix, 0.155 re-weighted to Hypermind’s mix.
- Kalshi: 0.199 on its own mix, 0.151 re-weighted.
- Hypermind, scored on the same single-outcome basis: 0.119.
The report does not claim a win from these figures. It calls the platforms “very likely in the same ballpark”, because the questions are not the same.
Calibration at market midpoint, measured with data from Brier.fyi and Calibration City, puts Hypermind at 1.8 percentage points of error, Polymarket at 2.0 and Kalshi at 1.6. Kalshi’s figure is dominated by one bin: 88% of its 347,373 scored markets sit between 45% and 50%. The report argues that Hypermind’s residual error falls within the margin explained by sampling noise. By contrast, it calls the miscalibration on the two real-money platforms “statistically real”.
Nothing in this analysis says real money hurts accuracy.
Hypermind accuracy report, July 2026
What it does not show
In the comment thread, Nicolas Espinoza asked whether more traders bring better scores. Servan-Schreiber replied that within one market, accuracy should rise with the number of traders, with diminishing returns. Across markets, Brier scores depend at least as much on how hard the question is.
Two days later he shared a reading of the report by Scott Longley in the newsletter Compliance+More. Longley listed the limits. The sample has few sports markets, the comparison covers only a few hundred real-money markets, and it is Hypermind analysing Hypermind, though the data are open. His verdict was: “Its defensible conclusion is parity, not victory.” He called for a head-to-head study on identical questions, especially sports, and for regulators to separate three questions: whether markets are accurate, whether money improves them, and what public value they create.
Hypermind’s evidence would appear to shift the burden of proof.
Scott Longley, Compliance+More, 4 August 2026
That is the argument Servan-Schreiber has made since 2003, when an early play-money market matched a real-money exchange over an American football season. The new report rests that argument on twelve years of public data, and invites others to check it.
