At 50%, a Manifold market asking "Will I...?" resolves YES about 39% of the time. Nothing else in the paper looks as immediately tradeable. The authors never attempt the trade, though, and the paper offers no evidence that the gap would pay.
A crowd prices personal goals
Chen, Zheng and Zhu begin with the planning fallacy: people overestimate their own follow-through. In lab studies, observers judging someone else rely on base rates and make less optimistic predictions. A prediction market should bring those observers together and reward accurate forecasts. Manifold, a play-money platform, lets anyone create a binary market. Users often ask about their own goals, such as going to the gym three times this week or finishing a thesis draft by Friday. The authors use these markets to ask how well a crowd prices one person's future.
The researchers collected all 194,347 markets created from 2021 to 2026, then pulled every order for the sampled markets: 1,346,194 order records. Their definition of a self-market requires a forward-looking, first-person question beginning with "Will I", "Can I", "Shall I" or "Am I going to". External markets contain no first-person word. The resulting sample includes 6,150 self-markets opened by 1,660 creators. It also contains 6,272 external markets opened by other creators and 4,486 external markets opened by the self-market creators themselves.
Self-markets ended YES 45.0% of the time, while their average final price was 49.4%. That leaves a -4.4 pp gap (SE 0.5), compared with -0.9 pp for external markets from other creators. After controls and price-bin fixed effects, the difference is -2.2 pp (SE 0.5). Comparing markets within creators who opened both types gives -1.9 pp (SE 0.7, p=.006). The errors cluster in the middle: among final prices between 20% and 80%, self-markets have a -9.4 pp gap and external markets a -1.2 pp gap.
Who moved prices away from outcomes?
Observers account for most of it. They largely failed to sell. At opening, self-markets were underpriced against external markets at the same price, with a difference of +3.0 pp. By one day after creation, that same-price difference had reversed to -2.4 pp (p=.021). It reached -5.3 pp a week before close.
During the first day, human observers pushed external markets 4.5 pp toward NO. For self-markets that did not describe an unwanted outcome, they pushed only 0.8 pp toward NO. The 3.6 pp difference measures relative support for YES through a missing sell-off. When the person never traded, observers left self-markets unsold, moving prices 0.7 pp toward YES (SE 0.5). External markets with similar YES rates (38% and 37%) dropped 4.3 pp. The difference was +5.0 pp (SE 0.9).
The person supplied much of the eventual correction, often near the end. Creators accounted for 38% of movement toward the realized outcome in self-markets, compared with 9% in external markets. Of that contribution, 47% arrived on the final day. Even with the creators' help, total movement toward the outcome reached 27.8 pp in self-markets against 37.3 pp in external markets.
Four judgments between the result and a trade
The appendix sets out the classification rule, including its word lists. The authors reconstruct prices using every order, including cancelled and unfilled orders, in placement order. Executions carry their fill times. This reconstruction makes a difference: final prices match the platform in 99.4% of markets, compared with 91.9% when using only filled orders that were never cancelled. Orders subsequently cancelled accounted for 17.0% of purchase executions. Working from the released data, two research assistants reproduced 23 reported quantities to within 0.02%.
Four judgment calls still stand between these results and a trade.
Close means the earlier of scheduled close and resolution. When a market resolves early, landmarks measured backward from close therefore depend on a time the trader learns afterward. We did not find a version anchored to scheduled close. Landmarks measured forward from creation avoid this problem. At one day after creation, the raw self-market difference is -3.7 pp (SE 1.0).
The "person never traded" split yields the striking -7.6 pp a week before close. As the authors explain beneath their landmark table, the split uses the market's entire history. A trader at that landmark cannot know whether the creator will stay out through the final price.
Settlement presents another choice. YES or NO resolutions account for 92.0% of self-markets and 90.8% of external markets. Treating all remaining outcomes as NO or as YES puts the self-market gap between -7.9 and +0.1 pp. Valuing them at their last price gives -4.1 pp.
Prior interaction and the 50-market threshold for "experienced" observers also depend on the sample. Both are counted within sampled markets, so another draw changes both measures.
Creators pick winners 4 points more often and earn nothing extra
Creators bought the eventual winning side 4.0 pp more often than strangers trading the same market (SE 1.2). Restrict purchases to more than a day before close, and the edge is +1.6 pp (SE 1.4). After the purchase price, it is -0.7 pp per share (SE 1.7). Creators knew more. By the time they bought, prices already reflected most of that information.
Anyone considering selling self-markets faces the same arithmetic. The paper's closest approximation to a return measure compares purchases within the 8,611 observers who bought both types of market. Their self-market YES purchases returned -0.3 pp per share (SE 1.8) relative to their own external YES purchases at the same price. For NO purchases, the difference was +0.8 pp (SE 1.8). Neither difference is detectable.
Execution would also happen in thin books: the median self-market had 9 observers and 383 mana in observer volume. The creator usually settles the market, too. Coders could not reliably determine whether other people could verify the outcomes (κ=0.22).
What can recalibration repair?
Final self-market prices recorded a Brier score of 0.105, compared with 0.049 for external markets. The Murphy decomposition separates calibration error from the ability to distinguish outcomes. It attributes most of the difference to resolution, at 0.146 versus 0.195; reliability was 0.004 versus 0.001. Self-market prices did less to separate successes from failures. A monotone transform cannot repair that weakness.
A recalibration fitted before 2025 and tested on 1,297 later self-markets treated those markets separately. It reduced the Brier score by 1.5% (change -0.0021, 95% CI -0.0058 to 0.0022). Only pooled recalibration of day-one prices produced an improvement with an interval excluding zero: 2.5% (CI for the change -0.0114 to -0.0008).
The abstract reports the 1.5% improvement and explains its small size in the same sentence: prices mainly failed to distinguish successes from failures. It then locates the missing information with the person.
We did not reconstruct the data and ran no test of our own. Substituting equity or crypto markets would lose the mechanism, which rests on one person controlling the outcome and settling the market.
The paper delivers a clean measurement of crowds failing to apply a base rate to people. It leaves no evidence that the gap pays.