Forecasting Beyond Prediction Markets
How markets reward informed judgment, what limits their reach, and why forecasting matters beyond the questions people trade.
Scott Alexander’s case for prediction markets begins with a powerful incentive: “either prediction markets will be accurate, or you can get rich quick.” If someone can reliably estimate probabilities better than the market, they can profit by trading on the difference. Their trades then move prices toward the information they possess.
The argument explains why markets reward accurate judgment. It does not establish that every error will be worth correcting. Alexander acknowledges this himself: a conspicuous mispricing can persist when the potential return is too small to justify the trouble of trading.
Prediction markets can produce useful forecasts, but their reach depends on what attracts money, attention and analytical effort. Understanding that mechanism helps explain both their achievements and the questions they leave unanswered.
I. Different motives, one price
Consider an illustrative election market. A contract pays $1 if a candidate wins and nothing otherwise. It trades at $0.52, commonly read as a 52% probability of victory. That price expresses an assessment of uncertainty, not an absence of information.
A partisan buys because he expects his candidate to win. A casual bettor buys for the excitement of having money at stake. A portfolio manager buys because the contract could offset losses elsewhere if that candidate takes office. These participants need not share a motive, an information source or a forecasting method.
Their orders create opportunities for someone with a better estimate. An analyst studies turnout, polls and historical polling errors, and puts the candidate’s chance at 58%. At 52 cents, the contract offers an expected gain of six cents per share before costs, provided the analyst’s estimate is sound. Buying moves the price toward that estimate, although the available profit shrinks as the price rises.
The analyst is deliberately forecasting. The market’s other participants may be expressing conviction, seeking entertainment or buying insurance. The useful feature of the mechanism is that they do not all need to be trying to produce an accurate public forecast for their interaction to yield one.
That does not mean every trade improves the price. Informed traders can be wrong, biases can reinforce one another, and a correct election outcome cannot by itself establish that the probability quoted beforehand was well calibrated. The question is whether the mechanism works repeatedly, across many events.
II. Where the information comes from
The familiar explanation is the “wisdom of crowds”: independent errors cancel out when enough judgments are combined. A market, however, does not simply average everyone’s opinion. Participants choose how much to trade, when to trade and whether to participate at all.
Gómez-Cram, Guo, Jensen and Kung (2026) investigate this distinction using Polymarket transactions. They identify a persistently skilled minority, around 3% of accounts, whose trades predict subsequent prices and final outcomes. These traders respond to public news, correct inconsistent prices and trade against behavioral mistakes. The broader crowd supplies much of the volume but comparatively little information.
This is evidence for concentrated skill in the market they studied. It does not imply that exactly 3% of traders drive every prediction market, or that everyone else contributes nothing of value.
Distinguishing skill from luck
A profitable record alone cannot tell us whether a trader has an informational advantage. Someone can make money by chance, particularly over a short history. The paper uses a sign-randomization test to compare realized trading performance with a benchmark constructed by randomizing trade direction.
The intuition is to hold the trading opportunities fixed and ask how often randomized choices would perform at least as well. Suppose 50 of 10,000 randomized records match or exceed the actual result. The resulting fraction, 0.005, measures how unusual the record is under that particular benchmark. It is not the probability that the trader lacks skill.
Under a valid continuous null model, p-values are uniformly distributed. A concentration near zero can therefore provide evidence against the no-skill benchmark, while a concentration near one can indicate systematically poor performance. Interpreting either tail still depends on the test’s assumptions and how the trading records were selected.
Illustration of the test’s logic using simulated data, not a reproduction of the paper’s measurements. Bar heights and tail sizes are illustrative.
A stronger check is whether the apparent skill survives beyond the observations used to identify it. In an interview about the research, the authors emphasize that the skilled group’s performance persists. That distinction matters: identifying yesterday’s winners is much easier than identifying traders whose future decisions will remain informative.
How a minority can move prices
A small informed group can influence prices disproportionately if its trades respond consistently to relevant information, while less informed trading is more dispersed. The following simulation illustrates that possibility by assuming the noise is centered around zero and informed trades move in the right direction.
Illustrative simulation, not observed trades. Noise cancellation and informative trade direction are assumptions of this example, not guarantees about real markets.
Real biases need not cancel. Traders can share the same misconception, and informed traders can disagree. The empirical result is narrower and more interesting than the claim that a crowd inevitably finds the truth: in this study, information enters prices disproportionately through a small group with persistent skill.
III. When correction is worth the effort
An incentive to correct an error is not the same as a guarantee that someone will correct it. Research takes time, trading incurs costs, and capital has alternative uses. The opportunity must be large enough to pay for all three.
The economics of correction
Suppose a trader estimates an event’s probability at 70% while its contract trades at 50 cents. The apparent gap is substantial, but the trader still needs to know how many shares can be bought near that price, how reliable the estimate is and how long the position may need to be held.
A large pricing error with only a few dollars available to trade may not justify hours of research. Conversely, a small discrepancy can be valuable when a trader can take a sufficiently large position. Trading volume alone does not settle the question: volume records past activity, while liquidity concerns the ability to transact now without moving the price substantially.
There is no universal dollar threshold above which a market becomes accurate. The cost of investigating the question, the available prices and the competitive environment all matter. The relevant constraint is whether the expected reward can support the work required to improve the forecast.
This leaves a gap between questions that are important to answer and questions that are profitable to trade. A decision can matter enormously to one organization without attracting a broad betting audience. Its importance does not automatically fund a liquid market or the research needed to price it.
Time and persistent bias
Long horizons can make the economics harder. A position held until resolution ties up capital, so an otherwise attractive expected gain may compare poorly with alternatives available over the same period. Traders can sell earlier, but only if there is a buyer at an acceptable price. Long-horizon forecasting is possible; the route from better judgment to a realized return is less direct.
Bias introduces another complication. A 2026 study by Nam Anh Le, revised in August, finds persistent underconfidence in political markets on Kalshi and Polymarket: prices are compressed toward 50% relative to observed outcomes. It also finds that calibration varies with domain and time to resolution. The result challenges the idea that every market price can be interpreted in the same way.
That finding does not establish that partisan traders continuously cause the compression. It identifies a pattern that the simple correction story does not explain away. Understanding when a price is informative requires examining the market and its participants, not merely noting that someone could profit if it were wrong.
IV. What markets establish, and what remains open
Prediction markets provide a powerful way to reward and observe forecasting skill. The research on persistent traders suggests that accurate judgment can survive beyond a lucky run, and that markets can transmit that judgment to people who never trade.
But the skill and the mechanism that rewards it are different things. A market recruits analytical effort through the prospect of trading profits. That is an effective arrangement when enough participants, information and capital meet around a well-defined question. It is a less obvious fit for questions that require expensive research, concern a narrow audience or take years to resolve.
This is the limit of treating prediction markets as a general solution to forecasting. Their coverage follows the economics of participation, which need not match the importance of the decisions people face. Better infrastructure may broaden that coverage, but it does not make the two identical.
The broader opportunity is to organize and evaluate forecasting effort wherever it is useful, including where a betting market cannot support it. Prediction markets show one way to make informed judgment valuable. They leave open how much further that judgment could reach if we funded the research directly and tested it against outcomes.