Do prediction market odds equal probability? Not quite
A US exchange can list a new prediction market by filing a form saying the contract complies with the law, and start trading the next day. No approval required.
- Prediction market prices are well calibrated overall: studies of thousands of settled markets find outcomes occurring at close to their implied frequencies, with accuracy that beats individual experts and polls.
- The best-documented distortion is the favorite-longshot bias: cheap contracts win less often than their prices imply and expensive contracts win slightly more, so buyers of long shots earn systematically negative returns.
- Capital lock-up is the least discussed distortion: a contract paying $1 in six months is worth less than its probability today because the money is committed and earning nothing, which pushes long-dated prices below fair value.
- Calibration varies by domain and horizon, with political markets showing compression toward 50% at long horizons, attributed to opposing partisan bets cancelling instead of informing.
- Fees, spreads, and the maker-taker split move realized returns meaningfully on instruments priced in cents, and resolution risk sits underneath everything as the possibility that a correct forecast still fails to pay.
Behind that speed sits a trapdoor written into Dodd-Frank, three undefined words, and a rulemaking the CFTC opened this June to finally settle what they mean.
The most useful sentence ever written about prediction markets is that a contract trading at 70 cents implies a 70% probability, and the most useful next sentence is that this is an approximation with known, measurable errors. Both halves matter. The first is why journalists, analysts, and increasingly institutional data buyers treat these prices as forecasts: the mapping is real, and the empirical record supporting it is better than most critics assume. The second is why traders who read the price as literal truth lose money in patterned, predictable ways. Research covering hundreds of thousands of settled contracts across the largest venues now supports a precise account of where the mapping holds and where it bends, and the answer is not that markets are wrong but that a price is a market-clearing number produced by capital under constraints, not a probability produced by an oracle. This guide walks the evidence: the calibration record, then the five distortions, then how to read a price properly.
The mapping, and why it mostly works
Start with the good news, because it is stronger than the skeptical framing usually allows.
Calibration studies plot implied probabilities against realized frequencies: take every contract that traded at roughly 30 cents, check how often those events actually happened, and see whether the answer is close to 30%. Across large samples of settled markets, the resulting curve tracks the ideal diagonal closely. One analysis of thousands of markets on the largest regulated venue found overall accuracy above 90% across probability ranges, with the curve hugging the diagonal and no evidence of gross systematic error. Academic work examining more than 300,000 contracts reached a compatible conclusion: prices are informative, and they improve as markets approach settlement, which is exactly what an efficient information aggregator should do as uncertainty resolves.
Comparisons to alternatives are the second part of the case. Aggregated market prices have generally outperformed individual expert forecasts, single polls, and simple statistical models, because the mechanism rewards being right with money and punishes confident error, which is a stronger incentive structure than reputation. Market efficiency has also been improving as the sector grows: spreads on the leading venue compressed sharply as volume expanded, which mechanically improves price quality. For context, crypto.news has explained where the liquid markets live and why the venue structure matters for market quality.
So the base case is that these prices deserve to be taken seriously as probability estimates. The rest of this guide is about the five ways they deviate, each of which is measurable and each of which points the same direction: the deviations mostly hurt the participant who reads the price naively.
Distortion one: the favorite-longshot bias
The best-documented bias in the literature, imported from a century of horse-race betting research, is that markets overprice unlikely outcomes and underprice likely ones.
The evidence in prediction markets is now substantial. Studies of large Kalshi samples find that low-priced contracts win far less often than needed to break even, while high-priced contracts win slightly more often and deliver small positive returns. One analysis found that events priced above 80% occurred about 84% of the time, several points below what their prices implied, meaning even the favorites side of the bias produces a modest shortfall against expectations at that end of the range. The pattern shows up across politics, entertainment, and economic data releases, and across trade sizes and volumes, which argues against it being an artifact of one market type.
The explanations are behavioral and structural in combination: people systematically overestimate small probabilities, a finding that predates prediction markets by decades; cheap contracts offer lottery-like payoff profiles that attract optimistic buyers; and limited arbitrage capital means the mispricing is not fully competed away. For a participant, the practical implication is uncomfortable and simple: buying long shots at five or ten cents is, on the historical record, a systematically losing strategy, and the sellers of those contracts have been the ones collecting.
Distortion two: capital lock-up
The most underappreciated distortion has nothing to do with psychology. It is arithmetic about time.
Buying a contract at 70 cents commits 70 cents until settlement, earning nothing in the meantime. If settlement is a week away, the cost of that commitment is negligible. If settlement is a year away, the buyer has forgone a year of risk-free return on the capital, which at prevailing rates is a meaningful percentage of the stake. Rational participants therefore pay less than the true probability for long-dated contracts, and recent work formalizes this as settlement discounting: in collateralized markets where capital sits locked until resolution, the price-as-probability mapping is incomplete, because these venues are information aggregators embedded in capital markets, not frictionless probability oracles.
The practical consequences run in two directions. For a reader treating the price as a forecast, long-dated contracts systematically understate the true probability, and the effect compounds with the horizon. For a trader, the discount is not an anomaly to exploit but the market correctly pricing the cost of committed capital, which means an apparent edge on a distant contract may be entirely consumed by the opportunity cost of getting there. Any comparison between a prediction market price and a poll or model output should account for this, and almost none do.
Distortion three: liquidity and domain
Calibration is not uniform across markets, and the variation is systematic enough to have been decomposed.
Volume concentrates heavily: political and macroeconomic contracts have accounted for a majority of trading on the largest venues, which means those markets have the tight spreads, the professional participation, and the price quality that the calibration studies mostly measure. Thin markets on obscure questions inherit none of that, and a 40-cent price in a book with a fifteen-cent spread carries far less information than the same number on a Fed decision.
Domain matters beyond liquidity. Research examining calibration across knowledge domains, horizons, and trade sizes found that a handful of components accounted for the large majority of variation, with political markets showing pronounced underconfidence: prices compressed toward 50%, understating the probability of favored outcomes, at nearly every horizon and most strongly among the largest traders. The proposed mechanism is bilateral cancellation, in which opposing partisan bets pull prices toward the middle without adding information, and the same pattern replicated on a structurally different venue, which strengthens the finding.
There is also a category where calibration is close to meaningless: questions with no historical base rate. A market on whether an unprecedented technological milestone occurs by a distant date has nothing to anchor to, and its price reflects sentiment among a small self-selected group. Those markets are entertainment dressed as forecasting, and they should be read accordingly.
Distortion four: fees, spreads, and who you trade as
On instruments priced in cents, transaction costs are not a rounding error, and the research shows they fall unevenly.
Analyses of the maker and taker split find that participants providing liquidity earn better returns than those taking it, for two compounding reasons: makers obtain better prices by definition, and takers generally pay the fees. Layer the favorite-longshot bias on top and the worst realized outcomes concentrate among takers buying cheap contracts, which is also the most intuitive behaviour for a new participant. The gap is measurable in the return data across price deciles.
The spread deserves separate attention because it is the cost most often ignored. A two-cent spread on a 65-cent contract consumes roughly 3% of the position immediately on a round trip, which against an expected edge of a few percentage points can erase the trade’s entire rationale. Spread quality has improved substantially with volume, but it varies enormously by market, and checking it before sizing is the single highest-return habit available.
Distortion five: resolution risk
The last distortion is the one that turns a correct forecast into a loss, and it is structural, not statistical.
A contract pays according to its stated resolution criteria as adjudicated by its named source or process, and that adjudication can diverge from what an ordinary observer concludes happened. On regulated venues the source is typically a designated authority, which makes disputes rare but not impossible where wording is ambiguous. On blockchain-based venues, settlement runs through decentralized oracle processes with proposal, challenge, and token-holder voting stages that this publication examines in detail, and there the divergence risk is materially higher and has produced real disputed payouts. That is the risk underneath every price.
The correct way to hold this is as a haircut on every price. A contract at 90 cents is not a 90% chance of being paid; it is a roughly 90% chance the event occurs multiplied by the probability that the resolution process pays it as expected. In liquid markets with objective single-source criteria, that second factor is close to one. In ambiguously worded or contentious markets, it is meaningfully lower, and it is entirely absent from the headline number.
What the calibration research cannot tell you
Before assembling the method, one honest caveat about the evidence base, because the studies cited above have limits that their headline numbers conceal.
The samples are historical and venue-specific. The largest datasets cover a regulated exchange over a period running from 2021 through 2025, an era in which prediction markets were smaller, more concentrated among sophisticated participants, and dominated by categories with clean resolution sources. Calibration measured on that population may not describe a market that has since added tens of millions of retail accounts through brokerage distribution, expanded aggressively into sports, and grown volumes by an order of magnitude. More retail participation could improve calibration by adding diverse information or worsen it by adding correlated sentiment, and the honest answer is that nobody yet knows which dominates at current scale.
Selection also shapes what gets measured. Calibration studies necessarily examine markets that resolved, which excludes contracts delisted, withdrawn, or voided, and those are disproportionately the ambiguous or contested ones where the price-to-probability mapping would have performed worst. The measured record is therefore a record of the well-behaved subset, and the true error rate including resolution failures is worse than the curves show.
Regime change is the third limit. Calibration is a property of a market’s participant mix, incentive structure, and information environment, all of which are shifting fast: new venues, new distribution, institutional data buyers, leveraged product variants, and a legislative environment that could remove entire categories. Findings proven on one configuration do not automatically survive into the next, which is why the coming election cycle is the most informative calibration test the sector has faced, and why any confident claim about accuracy should be dated.
None of this undermines the base case. It sharpens it: prediction market prices have a good measured record on a specific historical population under specific conditions, and the correct posture is to use that record as evidence while treating the current, much larger, much more retail market as an ongoing experiment whose results are not yet in.
How to read a price properly
Assemble the five and a usable method falls out.
Treat the price as a strong prior, never a fact. Adjust upward for long-dated contracts to account for the capital lock-up discount. Discount extreme prices toward the middle, since long shots are overpriced and heavy favourites are slightly overpriced too. Weight the reading by liquidity, taking prices from deep, professionally traded markets seriously and thin ones as sentiment. Read the resolution criteria and apply a haircut where the wording admits argument. And when trading rather than reading, account for fees, the spread, and whether you are making or taking, because those costs land before any edge does.
None of this argues against the instruments. The calibration record is genuinely good, better than most alternatives, and improving with volume. It argues for reading them the way a professional reads any market-implied number, an inflation breakeven or an options-implied volatility: as information produced by capital under constraints, containing real signal and predictable distortions, and worth more to the person who knows which is which.
One last practical note, aimed at the readers who consume these prices without ever trading them, which is now most of the audience. Prediction market numbers increasingly appear in political commentary, market research, and media dashboards as substitutes for polls, and the substitution is usually presented without any of the qualifications above. A responsible citation of a market price does three things: it names the venue, since calibration differs by market structure and resolution architecture; it names the date and horizon, since the same question priced a year out and a week out carries different distortions; and it treats the number as one estimate among several, never the answer, because the research showing markets beat individual experts does not show them beating the combination of markets, models, and polls read together. Crypto.news has also covered who is buying these numbers as exchanges, sportsbooks, and data buyers fight over the value of market-implied probabilities.
The strongest version of the case for these instruments is that they add a real, financially disciplined signal to a forecaster’s toolkit. The weakest version, and unfortunately the most common in circulation, is that a number from a screen settles a question. The distance between those two readings is what this guide has been about, and it is entirely made of the five distortions above.
Frequently asked questions
Does a 70-cent contract mean a 70% probability?
Approximately. Calibration studies across thousands of settled markets find implied probabilities track realized frequencies closely, with overall accuracy above 90% across price ranges. The mapping is a good first approximation with documented deviations at the extremes, over long horizons, in thin markets, and after fees.
What is the favorite-longshot bias?
The tendency for cheap contracts to win less often than their prices imply and expensive ones to win slightly more. Research on large samples finds low-priced contracts deliver systematically negative returns while high-priced contracts yield small positive ones, with one analysis showing events priced above 80% occurring about 84% of the time. Buying long shots is, on the record, a losing strategy.
Why do long-dated contracts trade below their true probability?
Because capital is locked until settlement and earns nothing meanwhile. Committing money for a year to a contract paying $1 has a real opportunity cost, so rational buyers pay less than the fair probability, an effect recent research formalizes as settlement discounting. Any comparison of a long-dated market price to a poll or model should adjust for it.
Are prediction markets more accurate than polls or experts?
Generally yes, in the aggregate. Market prices have outperformed individual expert forecasts, single polls, and simple statistical models across many studies, because participants are financially rewarded for accuracy and penalized for confident error. The advantage is largest in liquid markets and smallest in thin ones with no historical base rate to anchor prices.
Which markets should be trusted least?
Thin ones, distant ones, and unprecedented ones. Wide spreads mean low information content; long horizons introduce the lock-up discount and, in political markets, documented compression toward 50%; and questions with no historical base rate, such as unprecedented technological milestones, have nothing anchoring their prices beyond the sentiment of a small self-selected group.
How much do fees and spreads matter?
Considerably, on contracts priced in cents. A two-cent spread on a 65-cent contract costs roughly 3% on a round trip, which can exceed a realistic edge. Research also finds liquidity providers earn better returns than takers, who both pay fees and receive worse prices, with the gap widest among buyers of cheap contracts.
What is resolution risk?
The possibility that a contract fails to pay as expected because of how it resolves rather than what happens in the world. Contracts settle against named sources and pre-written criteria, so ambiguity can produce outcomes that surprise participants, and on blockchain venues using decentralized oracle voting the risk is materially higher. Every price should be read with a haircut for it.
How should a careful reader use these prices?
As a strong prior rather than a fact: adjust long-dated prices upward for capital lock-up, discount extreme prices toward the middle, weight by liquidity, read the resolution criteria, and subtract transaction costs before assuming an edge. Treated that way, prediction market prices are among the most useful public forecasts available. This is educational information, not investment advice.
Disclaimer: This article is for information and educational purposes only and does not constitute financial or investment advice. Research findings cited reflect published studies of historical data and do not predict future accuracy, and trading event contracts carries risk of total loss of amounts invested. Always do your own research. Information is accurate as of July 27, 2026.
