Event Trading and the Decomposition of Collective Judgment
An event contract pays one unit if a described event occurs; its price is routinely read as the crowd’s probability, and the reading is wrong in a specific, consequential way. The price is a mixture: dispersed public wisdom aggregated from many small staked judgments, and concentrated private knowledge injected by the few who know. The value of event trading as an instrument depends entirely on whether that mixture can be decomposed. Four forces set the mixture, each derived by holding traditional finance as the control group and turning one dial. Statistically, event markets aggregate estimates of an exogenous fact rather than forecasts of other forecasts, which is why calibration is directly measurable here and only indirectly, through the risk-neutral wedge, measurable for equities; Kyle’s insider-trading model supplies the formal spine. Behaviorally, the motive mixture inverts: expressive flow is the norm, hedging is nascent, and the one motive that lets uninformed retail win in traditional markets, positive-sum investment, has no event-market analogue. Financially, the contract converges to an unknown binary at a known date, insider trading falls outside securities fiduciary law and is nearly model-risk-free in a shrinking gray zone rather than a safe harbor, and retail loss is the structural mechanism by which information is paid for. Computationally, AI agents collapse the cost of informed-looking opinion, scale coordinated manipulation, and launder insider knowledge through anonymous flow, even as frontier AI forecasters still trail liquid market prices and win only by joining the staked aggregation. The conclusion is a design theory: the mixture stays decomposable, and the price trustworthy, only if the venue provides attributed identity, bounded authority, and per-action audit.
The most famous result in the study of collective judgment is usually told as a story about crowds, and it is really a story about stakes. At a 1907 English livestock fair, 787 visitors paid a small fee to guess the dressed weight of an ox, with prizes for accuracy; Francis Galton, expecting to document the ignorance of the average voter, found the middlemost estimate (1,207 pounds against a true weight of 1,198) within one percent of the truth. But Galton’s crowd was no poll: every estimate cost money, was rewarded only for accuracy, and was made independently by people who had chosen to compete. The canonical demonstration of the wisdom of crowds is, on inspection, a demonstration of the wisdom of staked crowds: Hayek’s price system running in miniature at a livestock fair. And alongside the fairgoers guessing for sport stood butchers and farmers whose daily trade was the carcass weight of cattle. Galton’s aggregate was already a mixture: dispersed public judgment and concentrated professional knowledge, pooled into one number that reveals neither ingredient.
Three terms do the framing work. The event contract is the instrument: a claim paying one unit if a described real-world event occurs (a game, an election, an announcement) and nothing otherwise. Event trading is the activity: taking sized positions in such contracts. The event market is the venue. The unit of analysis is the position itself: a stake converts an opinion into a conviction (belief × stake): sized in currency, paid for at entry, priced for exit every moment it is held. Opinion is what you say; conviction is what you hold. A poll collects opinions; a committee negotiates them; a pundit broadcasts them; an event market collects convictions: timestamped, sized, and revocable only at the going price.
The thesis is the generalization of Galton’s line. An event price is the clearing point of a mixture: dispersed public wisdom, carried by many small independent stakes, and concentrated private knowledge, carried by a few large, well-timed ones. Reading it as a clean probability misstates what it aggregates; the value of event trading as an instrument depends entirely on whether that mixture can be decomposed. A price of 0.34 backed by broad, size-diverse, two-sided flow is one epistemic object; the same 0.34 printed by one insider’s conviction hitting a thin book is another; the two are indistinguishable on a price chart. The microstructure classics make the mixture precise: in Kyle’s model, information enters the price only through the informed trader’s flow, camouflaged inside noise flow whose losses fund the informed trader’s profits. Strip the noise flow and price impact diverges; under common priors the Milgrom–Stokey no-trade theorem also bites, and an all-informed market does not trade at all. The expressive money pays the tuition; the concentrated money supplies the information; the published price is their sum, unlabelled (Figure 1).
The paper proceeds as four forces acting on that mixture, one per section, and then a conclusion about design: the statistical force (§2), what a staked aggregate is as an estimator, the Kyle formalism, bet-size decomposition, and the participant’s Bayes-plus-Kelly discipline; the behavioral force (§3), the four motives that populate the book, what stakes repair and what they do not; the financial force (§4), the contract as an instrument, and the hinge of insider flow; the computational force (§5), AI agents, which skew every prior term at once; and the conclusion the forces jointly force (§6): a design theory of trustworthy event prices. Throughout, traditional finance is the control group: each section names the one parameter event markets shift and derives the consequence. This is analysis, not advice: no venue is recommended, and nothing here is a trading tip.
The control group first. Traditional price discovery aggregates estimates of an endogenous value: Keynes’s beauty contest (the investor’s task is to anticipate what average opinion expects average opinion to be), so an equity price is a forecast of other forecasts, all the way down. The event contract shifts exactly one parameter: the object of estimation is an exogenous fact (the game ends, the vote is certified) that no quantity of opinion about opinion can move. The derived consequence is this section’s method: calibration is directly measurable here and far harder for equities. A share has no single realized frequency to score, yet its density forecast can be evaluated through the probability-integral transform (Berkowitz), the equity analogue of the calibration curve, run under the risk-neutral measure rather than the physical one. An election has a realized frequency directly. One caveat carries over from derivatives theory: an event price equals a belief only when the outcome is uncorrelated with the marginal utility of wealth, approximately true for a box score, systematically false for recessions, elections, and rates.
Begin with the estimator. A poll is a sample mean over whoever answered: it weights the decade-long student of the question and the respondent who first considered it during the phone call identically, and it costs nothing to answer carelessly. A staked aggregate differs in three ways. First, self-selection: the market is populated exclusively by people who chose to risk money on this question; participation is a filtered, bonded signal, and Grossman and Stiglitz give the equilibrium version: the expected profit from superior information is the wage that draws the informed into the market at all. Second, conviction weighting: each judgment is weighted by the currency its holder is willing to lose if wrong, self-administered, each participant pricing their own confidence under penalty. Third, proper scoring: Hanson’s market scoring rules make the connection exact for the venues that run one. A market maker running the logarithmic scoring rule is a sequentially shared proper scoring rule, so under myopic trading every trade is a scored probability report, paid on accuracy. The equivalence is a property of the LMSR mechanism and its myopic-trader assumption; it does not carry to the order books where serious volume has consolidated (§4), where non-myopic traders can bluff, moving the price to mislead and profiting on the reversal, rather than report (Chen and coauthors).
What kind of estimate results? Not a mean belief: prices equal mean beliefs only under specific assumptions, and the deviations are systematic. The most robust is the favorite-longshot bias, documented across a century of parimutuel data: longshots overpriced relative to their win frequency, heavy favorites underpriced. The bias is robust in its sign and contested in its mechanism. Snowberg and Wolfers read it as probability misperception rather than risk-love; Ali derives the same tilt from heterogeneous beliefs among risk-neutral bettors; Ottaviani and Sørensen catalog the competing explanations without a settled winner. Even the sign varies: Woodland and Woodland document a reverse favorite-longshot bias in baseball and hockey betting, where the favorites are the overbet side. The venue matters as well. The century of evidence is parimutuel, where bettors set the odds among themselves; on a CLOB exchange the bias attenuates, and part of what survives near 0 and 1 is mechanical, a product of the tick grid and the margin a longshot’s skeptic must post rather than a belief distortion, a caveat Wolfers and Zitzewitz stressed. The bias is the first exhibit for the thesis: the price is the clearing point of a mixture rather than one crowd’s belief, and the tails are where the uninformed component visibly dominates.
The mixture is more than a metaphor: it has a canonical formalization. Kyle’s 1985 model puts exactly three actors in a market: an informed trader who knows the outcome-relevant value, noise traders who trade for reasons unrelated to information, and a market maker who sees only the combined order flow and sets the price competitively. Its equilibrium delivers, as theorems, the three claims this paper needs (Box 1): information enters the price only through the informed trader’s orders; the informed trader’s expected profit equals the noise traders’ expected loss under a zero-profit competitive market maker, so expressive money pays the tuition is that idealized equation in words, weakening to an inequality once makers, fees, and the venue take their share, with noise losses at least covering informed profits (§4); and the insider trades because the noise flow provides camouflage. Remove the noise flow and the construction degenerates. In Kyle’s own comparative static, as noise variance goes to zero, price impact diverges, any informed order jumps the price to full information before it can profit, and trade vanishes. Milgrom and Stokey’s no-trade theorem is the general statement of the same boundary, and it holds only under its stated conditions: a common prior, an ex-ante Pareto-optimal allocation, and common knowledge of rationality. Strip any one and trade returns, which is why an all-expert market need not be silent in practice: heterogeneous priors, genuine disagreement rather than asymmetric information, are the standard alternative microfoundation under which informed participants trade absent any noise flow (Ali; Wolfers and Zitzewitz). Glosten and Milgrom translate the same logic into the bid-ask spread: it must recoup from the uninformed what the informed take, so the spread itself is a running estimate of the mixture’s insider share, and as insider precision approaches 1 the spread widens toward the market-breakdown region where quotes fail, the correct binary benchmark once Kyle’s Gaussian closed forms no longer apply.
Kyle’s model was built for equities, and the transposition to event contracts preserves its qualitative logic while voiding its Gaussian closed forms (Box 1): the event contract’s value v is genuinely exogenous and terminal, so “the informed trader knows v” literally describes the campaign staffer or the team physician (§4). What the model then says is uncomfortable and load-bearing: the accuracy the headline price is famous for and the losses the retail flow is famous for are the same transaction viewed from its two sides. Account-level evidence has measured the transaction at scale: roughly 3 percent of the largest crypto-native venue’s 1.72 million accounts supply the accuracy; on the largest regulated exchange, wealth flows systematically from takers to makers; and the top 1 percent of accounts capture 84 percent of gains. The wisdom of the staked crowd is the wisdom of staked minorities: the skilled few, conviction-weighted, funded by the many.
If the price is a mixture, the analyst’s task is decomposition, and the populations differ most legibly in one observable: individual bet size. Uninformed flow is many small tickets; hedging flow is sized to exposure; informed flow is sized to conviction and bankroll (rational informed sizing is Kelly-shaped, proportional to edge, as Box 2 derives); market-making flow is symmetric and mean-reverting in inventory. The distribution of stake sizes, not just the price, becomes the analytical object, and three instruments follow. Size-conditional calibration: stratify fills by stake size and score each stratum against outcomes. If large stakes are better calibrated than small ones, as the mixture model predicts, size carries recoverable information the price alone blends away; a calibration decomposition across 292 million exchange trades has found exactly such a distinct trade-size component. Stake-weighted versus count-weighted belief: the count-weighted average approximates a poll of the small-ticket population, the stake-weighted average the conviction-weighted belief, and the gap between them is a direct estimate of the uninformed distortion: a favorite-longshot-bias meter computable market by market. Flow-anomaly reading: conviction-sized orders that arrive against the prevailing drift and persist are the informed-flow signature, and measured large-trade order imbalance does predict returns across venues. One caveat sets the threat model for everything downstream. Bet-size forensics catch the naive insider, the one who trades big because he can. Kyle’s own equilibrium describes the strategic one, who sizes his order to β = σu/σv precisely to blend into the noise (Box 1) and is undecomposable by size in equilibrium, his flow scaled to be statistically indistinguishable from noise. The empirical rescue is partial and relocates the signature rather than removing it: informed traders do not vanish, they stealth-trade, concentrating in medium-sized orders that move price out of proportion to their volume (Barclay and Warner; Chakravarty), so the recoverable signal lives in the medium band and its timing and persistence, not in the whale ticket a naive audit hunts for. The decomposition is therefore an estimate against a partially strategic adversary, and §5 shows agents driving that adversary toward the fully undecomposable limit. None of this is computable from the ticker alone; it requires fill-level sizes. Here an irony of venue architecture surfaces: pseudonymous on-chain venues make the full size distribution publicly reconstructible, while regulated exchanges hold account-level data privately; the venue class with the weakest compliance regime is the one whose mixture is most auditable, a tension §6 resolves by design.
The same statistical framework, viewed from the participant’s seat, composes into a single loop (Figure 2). Belief formation: the input is a calibrated probability, manufactured rather than intuited: reference-class base rate, explicit Bayesian updates, the resulting p recorded with its reasoning before looking hard at the price (a belief formed after staring at the market anchors on it, and an anchored belief cannot price the market’s error). Edge: the edge is the difference p − π rather than the belief, and it must survive an adversarial question: why do I know something the market does not? Exactly three answers are defensible: an information edge (true and not yet priced; the insider’s edge, §4); a modeling edge (public information, better processed); or a mixture-reading edge (the price identifiably distorted by non-belief flow). A gap with no nameable source is the market politely informing you that your belief is wrong, not an edge. Sizing: Kelly’s criterion converts edge to stake fraction (Box 2); Thorp’s career is the proof the criterion survives real markets; fractional Kelly insures against one’s own miscalibration; portfolio Kelly allocates jointly across many simultaneous events and their correlations. Three frictions bind: correlated outcomes (one election night resolves dozens of contracts in the same direction), capital lockup (the relevant quantity is edge per unit time; §4 develops the lockup discount as carry), and thin books (sizing must be computed against post-impact prices, Kyle’s λ, now the participant’s own tax).
Scored on calibration, not profit. The loop closes with review, and the participant inherits the anti-resulting discipline of the judgment literature. Over any human-scale sample, profit is dominated by variance (a well-sized portfolio of good probabilities can lose for a quarter, a reckless one can win), so P&L is a noisy, slow, corrupting score. The convergent score is calibration: the participant’s own ledger of recorded probabilities against realized outcomes, Brier-scored, stratified by event class and claimed edge source. It converges far faster than profit and diagnoses the exact failure that P&L merely punishes. The venue keeps half of that ledger for you, since every fill is a timestamped, sized, irrevocable record of conviction.
The market-level version of the same audit is the calibration curve (Figure 3): bucket all prices and plot each bucket against the realized frequency of the events priced there. The empirical record, where its conditions hold, is genuinely strong: the Iowa Electronic Markets’ election prices beat 964 contemporaneous polls 74 percent of the time between 1988 and 2004 (Berg and coauthors), with the advantage growing at longer horizons. The rebuttal belongs in the record too: Erikson and Wlezien show that once the polls are properly projected to election day rather than read raw, they beat the market on the same data, so the IEM’s margin is over naive poll-reading rather than over the best poll-based forecast. But the curve’s signature shape (close to the diagonal mid-range, favorite-longshot deviation at the ends) is the mixture made visible: the uninformed component concentrates in the tails, where the licensed dream lives. Horizon matters as well (prices are well calibrated near expiry and biased at long horizons, consistent with the carry wedge §4 develops), and composition beats quantity: in 2024, a small position-capped venue priced political outcomes more consistently than one trading billions on the same questions (Clinton and Huang). Summarized: the staked aggregate is the best available estimator of the events it prices well, its known biases are mixture signatures, and the bet-size distribution is the instrument by which the mixture is read.
The control group calibrates the door. In traditional markets the dominant retail motive is investment; hedging is the norm; expressive trading is marginal; the informed are criminalized. Event markets keep the categories and shift every weight: investment demand is structurally absent (the missing row of Table 1, this section’s punchline), expressive trading is the norm, hedging is nascent, and the informed are not reached by securities fiduciary law (a shrinking gray zone, §4). Every consequence in this section and the next is a rotation of that one table.
The foil first: the three unstaked instruments a society uses to learn what a group believes. The poll weights every answer equally, costs nothing to answer carelessly, and is a snapshot: nobody revises a poll answer when the facts change. The committee deliberates: agreement is socially rewarded, dissent taxed, and the output converges on what Janis documented as groupthink. The pundit is paid in attention; Tetlock’s data made the correlation between fame and forecasting skill mildly negative. Decisively, the pundit keeps no ledger, so the misses evaporate while the hits are replayed.
Against this foil, who actually answers the market’s door? A mixture of four motives, none of them simply “the informed”, and the decomposition is the working tool of everything downstream (Box 3, Table 1).
Expressive bettors stake small amounts for entertainment, loyalty, and the license to dream: the “face” bet on one’s own team, the pray-for-a-miracle longshot whose few dollars buy a week of pleasant anticipation. Their flow is count-heavy and size-light, concentrates in longshots and emotionally charged outcomes, and is the standard microfoundation for the favorite-longshot bias: lottery-ticket buyers overprice lottery-shaped contracts. Expressive flow is real money but is closer to consumption than to belief; in Kyle’s vocabulary (§2) it is the noise flow: the camouflage the informed trade against, and the tuition that funds the price.
Hedgers hold real-world exposure to the event and buy the contract as insurance. A hedger’s trade is risk transfer, not belief: she may buy YES at 40 cents while privately assessing the probability at 25, because the contract pays exactly when the rest of her life does not. Hedging flow is informative about exposure, not probability; where one side carries more hedgeable pain, it shifts the price off the mean belief predictably: a risk premium wearing a probability costume. In the control group this motive is the origin story of derivatives regulation itself; its event-market counterpart is nascent: the exposures are real, the books mostly too thin for institutional size, and hedging demand is nonetheless the regulated venues’ entire legal theory.
Arbitrageurs and market-makers hold no directional belief at all. They enforce coherence (YES and NO summing to one, the same event priced identically across venues) and they supply the liquidity against which everyone else trades. Their flow makes the price consistent without making it right: arbitrage propagates information between markets but injects none about the outcome. Where the control group’s plumbing is nanosecond-complete, event-market arbitrage is fragmented, capital-constrained, and bears a risk with no clean traditional analogue: resolution basis risk. The “same” event on two venues is not the same contract, because resolution wording and oracles differ. Arbitrageurs enforce coherence, not truth, and here even coherence arrives late.
Informed traders (including, at the limit, outright insiders) trade because their probability materially differs from the price for reasons they could defend: private information, superior modeling, or superior reading of the other three flows. Theirs is the only flow that carries outcome information; it is the flow Grossman and Stiglitz say must be paid to exist; and its signature is size, timing, and persistence: conviction-sized orders, clustered near information events, leaning against the price rather than with it. It is Kyle’s informed trader, walking through a real door. In the control group this trader is constrained by criminalization and by endogenous value: knowing the earnings is not knowing the price reaction. The event insider faces neither (§4), and his signal precision approaches 1: maximal leverage per bit of private knowledge.
And the missing motive: the row that is not there. In traditional markets the dominant retail motive is none of the above: it is investment. Equities are positive-sum through the drift of an expanding economy; a passive, permanently uninformed investor wins by buying a share of an expanding pie and waiting, which is why index funds work. Event contracts have no such row: zero-sum minus fees, no drift, no beta to harvest: there is no index fund for event markets, and there cannot be. The one motive that lets the uninformed win in traditional finance is structurally unavailable, and the only honest retail motive left standing is consumption. §4’s retail-always-loses arithmetic is this row seen from the exit.
| Motive | Traditional markets | Event markets | Consequence for the price |
|---|---|---|---|
| Expressive | Marginal and episodic: meme-stock manias (GameStop, 2021) are the named exception; equities are abstractions, and nobody loves a discounted cash flow | The norm: trading as consumption; the product is having an opinion about the world, sized in currency | The liquidity subsidy that solves no-trade (§2); a predictable distortion (motivated reasoning, identity betting, lottery preference → favorite-longshot bias), decodable, not just noise |
| Hedging | The norm and the origin story: farmers, airlines, pensions; hedging demand legitimized derivatives regulation itself | Nascent and aspirational: the exposures are real (weather, elections, regulatory outcomes) but the books are too thin for institutional size | The missing legitimizer; benign premium-paying uninformed flow, and the regulated venues’ entire legal theory |
| Arbitraging | Nanosecond-complete plumbing; the law of one price enforced by machines | Fragmented, capital-constrained; plus resolution basis risk, with no clean traditional analogue: the “same” event on two venues differs in wording and oracle | Arbitrageurs enforce coherence, not truth, and prices stay incoherent longer |
| Informed | Constrained by criminalization and by endogenous value: even a true insider faces model risk, since knowing the earnings is not knowing the price reaction | Outside securities fiduciary law, a shrinking gray zone (§4), with binary decisive knowledge: the team physician knows rather than estimates | Maximal leverage per bit: insider signal precision approaches 1 in Kyle terms, so the adverse-selection tax is structurally higher than in equities |
| Investment / savings the missing row | The dominant retail motive: equities are positive-sum through growth: the passive uninformed investor wins by buying a share of an expanding pie | Structurally absent: zero-sum minus fees, no drift, no beta to harvest: there is no index fund for event markets | The formal statement of why retail structurally loses here yet can win in equities (§4): the one motive that lets retail win is unavailable; the honest retail motive left is consumption |
Each unstaked failure has a structural repair, and the repairs are worth stating as mechanisms rather than virtues (Figure 4).
Cheap talk becomes costly signal. Spence’s condition for a signal to separate types is that it be more expensive for the pretender than for the genuine article; Crawford and Sobel’s theorem is that when talk is free and interests diverge, little information survives transmission. Punditry is a cheap-talk equilibrium. A trade is a Spencean signal: expressing the conviction requires exposure to being wrong: cheap for the informed (positive expected value), expensive for the poser (negative). The stake is the message rather than decoration on it.
Groupthink is inverted, not exhorted against. In a committee, agreement with the emerging consensus is rewarded and dissent is taxed: socially, immediately, regardless of eventual truth. A market inverts the matrix. Agreeing with the consensus buys at the consensus price and earns approximately nothing even when right: the profit was already in the price. The maximum payoff goes to the trader who is correct precisely where the crowd is most wrong. Groupthink is charged for rather than lectured against, and dissent is paid rather than merely tolerated. No unstaked institution has this property, and it is the paper’s title made mechanism.
Cascades carry bounties. In an information cascade, each actor rationally ignores their private signal and copies the visible actions of predecessors; a unanimous-looking crowd may embody almost no information. Cascades feed on two conditions: acting with the crowd is safe, and acting against it is unrewarded. A market attacks the second directly: if the cascade has pushed the price away from your private signal, the divergence is a posted bounty on defection, growing as the cascade deepens.
Opinions get a maintenance mechanism. A pundit’s old prediction sits in the archive unrevised, costing nothing as it goes stale. A price is a standing offer to be corrected, and a position is marked to market continuously, so staying wrong has a running cost, not a one-time embarrassment. The order book is a conviction ledger that reconciles itself: its current line is always the crowd’s live, funded estimate, never its archived one.
The behavioral ledger has a second column, and the mixture thesis requires it: stakes change the payoff matrix, not the human. Three residues persist under full stakes. Favorite-longshot bias: a century of parimutuel data is a century of staked data, so the distortion survives stakes. On the misperception reading it lives in how humans perceive small probabilities, which a wager prices in rather than corrects, though the mechanism stays contested and the sign reverses in some sports (§2). Herding on the price itself: the market breaks information cascades by paying defectors, but it erects a new cascade substrate in the same gesture: the price is the most visible action of all, and a trader can rationally ride a cascade they expect to continue; markets do bubble, and the incentive argument is a direction, not a guarantee. Overconfidence: money does not repair what reputation does not; even frontier AI models, set loose on live markets, exhibited uniform overconfidence (§5). The stake filters motives and prices confidence; it does not upgrade cognition.
One honest complication belongs here. A careful comparison of a real-money and a play-money market forecasting the same NFL season found no significant accuracy difference. Read carefully, this refines the stake rather than refuting it: the play-money market retained most of the structure (self-selection, conviction weighting in a scarce scoring currency, contrarian reward, continuous correction) and removed only the currency. The operative ingredient is the full structure: a sized, scored, corrigible, self-selected record of conviction with asymmetric reward for correct dissent. Money makes the scoring real and is likely necessary where reputational incentives are thin or insiders must be attracted (§4); but the mechanisms, not the banknote, are the theory; and anything that fakes the structure while dodging the cost, as §5’s agents can, attacks the mixture at its behavioral root.
The control group once more. The event contract’s nearest traditional relative is the binary option: a cash-or-nothing digital claim. Four parameters shift. There is no underlying to delta-hedge: the replication argument that disciplines a digital option’s price has nothing to grip, so the event contract is an incomplete-market instrument whose price is disciplined only by the flow this paper has been decomposing. The legal regime is the CFTC’s event-contract framework, not the securities law that criminalizes the informed: the hinge below. Settlement runs on contract wording and oracles, so resolution risk is part of the instrument itself. And the instrument is terminal: an equity never expires; every event contract dies on a date. Same payoff diagram; four dials turned.
The control group’s own theory says what this instrument is: the Arrow–Debreu security made tradable: the primitive state-contingent claim from which traditional finance builds its entire pricing edifice, and which traditional finance itself can almost never trade directly. Event markets trade the theoretical atom of finance, naked, without the market-completeness apparatus the theory assumes around it: no spanning portfolio, no replication, just the atom and its order book. One inherited pathology does not transfer: short-selling is native and symmetric, and pessimism is expressed by buying NO, with no borrow to locate.
The fourth dial deserves its own subsection. The bond is the foil: it converges to a known value at a known date (pull to par), and traditional finance’s cleanest trades, basis and convergence, exploit precisely that known terminal value. The event contract inverts the trade: it converges at a known date to an unknown binary: pull to unknown par. The date is certain; the destination is the entire question.
Second consequence: information sensitivity peaks at expiry. The greeks are only an analogy here, since there is no underlying to differentiate against. The analogy is still exact in its one claim: a binary contract’s price sensitivity to new information concentrates near expiry at uncertain outcomes. With hours to run, a contract at 50 is a coin whose every new fact swings the price violently. Late-stage price violence is therefore structural, not behavioral, and it fuses with the insider forensics below: the last-minute insider bet is certainty entering where information sensitivity is maximal, buying the largest possible repricing with the smallest possible warning. That is why insider alpha concentrates in the final minutes, and why timing-versus-information-release is the audit that matters most.
Third: carry. A position pays at resolution, so holding it forgoes the riskless rate and every intervening opportunity; the observable price is therefore belief × discount factor, and the discount is measured: settlement discounts show a maturity-dependent term structure, and calibration decays at long horizons in exactly the pattern the discount predicts. A far-dated 95-percent event trading at 88 is an interest rate, not a miscalibration; even a bias-free market quotes discounted state prices, not probabilities. Assembled with §2 and §3, the decomposition reaches its full form: price = physical probability ± risk premium − carry − fees + expressive skew + insider flow. The two-component mixture of §1 is the headline; this is the ledger.
Fourth: resolution clustering. Expiries cluster on event nights: a single election night settles dozens of markets as one correlated draw, and Kelly’s independence assumption fails precisely at expiry. Portfolio risk in event trading is expiry-clustered in a way no equity portfolio is; §2’s correlated-outcomes friction is this fact from the participant’s seat.
The scale is no longer academic: combined monthly volume on the two largest venues ran under $5 billion as late as September 2025 and reached $44.8 billion by June 2026. Reported volume should never be read as liquidity (mint-and-burn mechanics alone inflate outcome-token headline volume roughly 2.5×), but the direction is unambiguous: event trading has become a financial market, and it inherits financial-market questions.
Three market structures implement the staked aggregate, and the choice among them is mostly a choice about liquidity and resolution (Table 2). The parimutuel pool is the mechanism in its simplest form: all stakes on a question pool together, the winning side divides the pool pro rata, and resolution can be a numeric oracle read: no counterparty matching and no market maker, so it functions at any participation level, at the cost that odds are final only when the pool closes. The central limit order book (CLOB) is the exchange model: binary contracts traded continuously against resting bids and asks, producing the live, continuously corrected price, at the cost of needing enough two-sided flow to keep the book populated. The automated market maker solves exactly that: Hanson’s logarithmic market scoring rule lets a sponsor quote both sides of an arbitrarily obscure question at a loss bounded in advance, though generic constant-product designs bleed to arbitrage on binary claims, which is why serious event-contract volume has consolidated on order books. Production venues outsource the liquidity subsidy: incentivized makers are paid to quote tight two-sided markets, which is why market-making is the account-level data’s strongest predictor of profitability, and, read through §2, the venue literally purchasing noise-absorption capacity.
Resolution is the instrument’s other load-bearing wall. Every contract is a promise to pay on an event description, and event descriptions are incomplete contracts in precisely the economists’ sense: no finite wording anticipates every state of the world. A trader must then price two entangled uncertainties: the event, and the adjudicator’s reading of the words. The damage is measured: semantically near-identical contracts differing only in resolution wording sustain persistent 2–4 percent cross-platform price gaps; and where adjudication is decentralized, the adjudicator becomes an attack surface: in March 2025 a roughly $7 million market was swung through its oracle’s dispute process at a cost in the low hundreds of thousands. The practical guidance: name the resolution source before trading opens; prefer numeric criteria; where words are unavoidable, publish the dispute procedure with the contract. A market’s forecasting value is bounded above by the completeness of its resolution clause.
| Parimutuel pool | Order book (CLOB) | Automated market maker (LMSR / CFMM) | |
|---|---|---|---|
| Price formation | Pool ratio; final at close | Continuous double auction | Scoring-rule curve; continuous |
| Counterparty | The pool (peer-to-pool) | Another trader (peer-to-peer) | The market maker’s curve |
| Liquidity requirement | None, works at any size | Two-sided flow; venues pay makers to supply it | None, sponsor subsidizes |
| Operator risk | None (fee on pool) | None (commission) | Bounded loss, set ex ante |
| Continuous correction | Weak; odds drift until close | Full | Full |
| Natural resolution fit | Numeric oracle read | Adjudicated event contracts | Any; sponsor defines |
| Canonical setting | Racetracks; on-chain binary pools | Regulated event exchanges | Subsidized corporate / crypto-native markets |
Event trading did not colonize event classes at random. The verticals where it thrives (sports, elections, major public events) share one property that finance lacks: exogenous, public ground truth (Table 3). The final score and the certified vote count are produced outside the market, unaffected by the market’s own price, verifiable by anyone at resolution time. Financial outcomes are different in kind: the price of an asset largely is the fact rather than a forecast of some external one, and beliefs about it feed back into it; Soros called the loop reflexivity. Finance also has no vacancy: options, futures, and swaps are already the deepest event markets for financial quantities. Sports is the vertical where the staked instrument most visibly outperforms its unstaked ancestor, the bookmaker, along a spectrum of information tolerance. At one end, most retail bookmakers declare a line, build a margin into the odds, and manage risk by limiting or banning the customers who win too often, systematically expelling the informed. At the other end sits Pinnacle, itself a bookmaker, which runs a low-margin, high-limit book that welcomes sharp action and prices off it, evidence that the business model rather than the license sets the tolerance. Exchanges sit toward the welcoming end: the line is discovered endogenously, peer-to-peer matching replaces the overround with a commission, and informed flow is admitted because it makes the price accurate. Even an exchange keeps a casino to protect, though. Betfair levies a Premium Charge of up to 60 percent on its most consistently winning accounts, taxing exactly the informed flow it nominally welcomes, a bookmaker’s instinct inside an exchange’s architecture. The real axis is how much information a venue will tolerate before it protects its recreational money. Hold that axis, because the rest of this section shows every venue, however structured, eventually finds its limit.
| Vertical | Ground truth | Reflexivity | Incumbent instruments | What an event market adds |
|---|---|---|---|---|
| Sports | Exogenous, public, instant (the final score) | Negligible (integrity risk aside) | Bookmakers: declared lines, overround, winners banned | Endogenous line discovery; no vig in the price; informed flow welcomed |
| Politics & major events | Exogenous, public, scheduled (certified counts, official announcements) | Low to moderate (prices can feed narratives) | Polls, committees, pundits | All of §3’s structural repairs; a continuous probability instead of episodic snapshots |
| Finance | Endogenous: the outcome is itself a price | High (reflexivity) | Options, futures, swaps: deep and regulated | Little: the vertical is already priced |
Here is the hinge, and it must be stated without flinching. Event outcomes are often knowable by someone. The injury report is the team physician’s, the internal polling is the campaign staffer’s, the unpublished statistic is the ministry employee’s. And the event contract’s payoff structure makes that knowledge uniquely monetizable: discrete binary resolution converts insider knowledge into alpha with almost no model risk. An equity insider must still price how much the news moves a continuous price through a noisy market; the event insider holds a claim that goes to exactly 1 or exactly 0 on a known date. Knowing the outcome is the terminal value itself rather than an edge in a model, and the only residual risks are detection and the fill.
The legality nuance is real and mostly unappreciated. Securities insider-trading law is built on theft: trading on material nonpublic information breaches a fiduciary duty to the source of the information. An event contract on a non-financial event has no issuer and no shareholder: there is no corporate informational property to steal, and CFTC event-contract regulation does not map the securities regime onto it. Trading an event contract on inside knowledge of the event is therefore not reached by the securities-law theory of fiduciary breach. That is a long way from legal. The same conduct is exposed to the CFTC’s anti-fraud-manipulation authority under Rule 180.1, which reaches trading on the basis of misappropriated material nonpublic information; to wire-fraud liability where a duty is breached to obtain the information; to venue rulebooks that already prohibit trading on MNPI, Kalshi’s among them; and, in sports, to foreign integrity statutes that criminalize it outright, including the UK Gambling Act 2005 cheating offence and comparable Australian provisions. It is a shrinking gray zone, and the May 2026 case marks the contraction: the US Department of Justice brought its first criminal insider-trading case centered on an event market. Manne’s dissent (insider trading accelerates price discovery) has run against the securities consensus since 1966; Hanson carried the argument to event markets explicitly. And the empirical record confirms the flow exists: hours before the 2025 Nobel Peace Prize announcement, the eventual laureate’s contract moved from under 4 percent to above 70 percent, and the Norwegian Nobel Institute called the trading criminal exploitation of leaked information; roughly 1,950 accounts in the largest crypto-native venue’s full history trade in patterns consistent with inside information.
The duality is the thesis in one actor. Kyle’s model says the insider is how truth enters the price: a venue whose insiders trade freely produces the earliest, most informative prices available anywhere. The same trades are how legitimacy exits the market: a last-minute skewed surge in a Nobel market, visible after the fact, tells every recreational counterparty what they were: the exit liquidity. Both descriptions are true of the same fill: the market’s greatest informational asset is its greatest trust liability, and no amount of moralizing resolves the duality, because it is the mechanism working exactly as Kyle describes. What the duality demands instead is decomposability: the price is worth more when the insider component can be seen, and the venue survives only if the noise flow believes it will be protected from what the decomposition reveals.
Forensics: reading the insider component. The decomposition tools of §2 specialize here into an audit practice (Figure 5). Timing against information release: insider flow clusters before the announcement it anticipates; a price move whose bulk precedes the public information event carries the signature. The Nobel episode is the canonical print, and terminality supplies the reason this audit is decisive: certainty entering where information sensitivity is maximal. Size anomalies: conviction-sized orders, out of scale with the account’s history and the market’s ambient ticket size, leaning against the drift and persisting, which is the signature of the naive insider; the strategic insider sizes to blend (Box 1) and stealth-trades in the medium band (§2, Barclay and Warner; Chakravarty), so this audit is a lower bound on the insider share rather than a full separation. Decomposition by bet size: the stake-weighted versus count-weighted gap and size-conditional calibration (§2) estimate how much of the price the large-and-right population is carrying. Venue architecture decides who can run the audit: on-chain venues expose every fill publicly but attribute nothing; regulated exchanges attribute internally but publish nothing. Neither has both attribution and openness: the gap §6 turns into a design requirement.
The structural consequence: retail always loses in aggregate. An event contract is zero-sum minus fees. There is no underlying enterprise compounding value, no dividend stream, no drift that lets all long-term holders win together: every dollar of trading profit is another participant’s trading loss, before the venue’s take. Combine that with the mixture and the conclusion is arithmetic: if the informed and the makers profit (and the account-level data says they do), then the aggregate retail book loses, necessarily, always. This is a structural feature, not a market failure to be fixed in a later product iteration: retail losing is the mechanism by which information enters the price, Kyle’s tuition inequality (Box 1) settling in real accounts, the wedge shared by informed, makers, and the venue. Nor is the asymmetry technological: the professional’s edge is capital and parallelization: Kelly-optimal play compounds many small edges across many thin markets simultaneously (Box 2). Hand every retail participant a frontier model and the power law survives, because the binding constraints were never information processing. The recreational trader’s expected loss is the price of admission: the market’s subsidy for the price everyone else reads for free.
The fee arithmetic binds regardless. Racetrack takeout of 15 to 20 percent made most players negative-expected-value regardless of skill; event-market fee loads are far smaller, but the identity is the same. Zero-sum fixes the dollar-weighted mean gross edge at zero, and with the documented right skew (the top 1 percent capturing 84 percent of gains) the median participant’s gross edge sits below zero before a cent of fees. So the mean after-fee return is exactly minus the fee, and the median is more negative still. Skill only decides which tail of the arithmetic you occupy.
The same arithmetic is this paper’s one paragraph on the gambling boundary. The stake that makes the record of conviction credible is, for the expressive population, a consumption good with a well-documented capacity for harm: expressive betting shades into problem gambling, and the favorite-longshot bias is in part a tax on hope collected from those least able to pay it. The claims here are about instrument quality; the welfare verdict belongs to the policy literature. The contribution to that debate is one clarifying fact: the mixture is a map of who bears the cost: the informed are compensated, the hedgers insured, the arbitrageurs flat, and the expressive population pays for the instrument everyone else reads.
Venue-legitimacy dynamics: paying to protect the noise. If the informed eat the uninformed, and the uninformed fund the market’s existence (§2’s no-trade boundary), then a venue’s commercial viability depends on retaining noise flow that knows it is noise flow, and adjacent industries have repeatedly paid real money to do exactly that. Online poker banned heads-up displays and introduced anonymous tables to keep recreational money at the table; casinos eject card counters; securities exchanges buy surveillance at industrial scale, because the perception of a rigged market drives away the flow that makes the market, rather than because manipulation is common. Event markets are running the same gauntlet a decade behind: regulated exchanges already operate surveillance and eligibility rules; crypto-native venues, pseudonymous at the wallet level, so far sell openness instead. These are two different instruments (one optimized for a fair game, one for an informed price), and the tension between those targets is exactly what §6’s design theory must resolve, because a venue that protects nothing loses its noise flow, and a venue that polices everything loses its information.
One more financial property closes the section. A published price is continuous, quotable, and legible (one number, where a poll is a crosstab), which gives it a broadcast role: media amplify it, exposed actors reallocate on its movement, and narratives form around the number itself, until “the odds moved” is the story. At scale, the price is partly forecast, partly instruction: Soros’s reflexivity, imported from finance to events. Prices can be self-fulfilling (a candidate priced as viable attracts the coverage and money that make him viable) and self-defeating: a warning priced high enough triggers the response that falsifies it. The structural point sharpens Table 3: exogeneity is a budget rather than a constant, and scale spends it. Sports outcomes are hardest to move (match-fixing is the degenerate case where the market’s money reaches the field); politics is plausibly the most movable, because every link (odds to coverage, coverage to momentum, momentum to donations and turnout) runs through actors who read the number. And one distinction must be stated plainly: manipulating the price attacks the measurement and tends to self-correct by subsidizing informed counterparties; manipulating the outcome attacks the world, and nothing in the market corrects it: a large enough position is, structurally, an incentive to make the event happen. The line between a hedge and a bounty is the holder’s ability to influence the outcome, one more thing a trustworthy venue must be able to see, which is to say: one more argument for §6.
The fourth force is best understood as a multiplier on every term already in the mixture rather than as a new participant type. An AI agent is a participant whose marginal cost of opinion production, account multiplication, and around-the-clock execution is near zero. The control group sets the tempo: traditional finance absorbed its automation arc over roughly forty years, and crucially after surveillance and regulation had matured; event markets are compressing the same arc into a few years, before either exists. The arc has the same two chapters here that it had there: first the quants, then the agents.
The naive verdict says event markets cannot be quantified. Traditional quantitative finance rests on two data pillars: long, homogeneous panels on persistent instruments, and wide cross-sections of correlated instruments sharing factor structure. Event contracts have neither: the instrument dies in weeks, and there is no factor model of one-off events. On the pillars as stated, the asset class is unquantable.
The verdict fails on its unit of analysis. The instrument dies; the game repeats. Every NBA game is a new contract drawn from the same generating process (same rules, same measurable features, same causal machinery), and the panel is the resolved outcomes of thousands of structurally identical instances rather than the price history of one contract. Quantitative event trading is possible exactly where that substitution works, and it has an existence proof at industrial scale: Benter’s Hong Kong horse-racing operation fit conditional-logit models over decades of structurally identical races and compounded them into roughly a billion dollars of winnings. Note what the method needed: recurrence, stable features, and a parimutuel counterparty (precisely the §3 expressive pool) to pay the model.
Generalized, quantability = event recurrence × feature stability, and the verticals arrange themselves along the axis. Sports sit at the proven-quantable end: thousands of recurrences per season, stable rules, rich public features. Macro prints and crypto-native events are semi-quantable: recurring but regime-shifting. One-off geopolitics sits at the far end: n = 1, no reference class that survives scrutiny: unquantable, and therefore insider-dominated, because where no model can manufacture an edge from public data, private knowledge is the only edge left. This is the section’s central observation: the quantability spectrum and the mixture-decomposition spectrum are the same axis. Where the game repeats, models and public data dominate and the wisdom component is large and legible; where it does not, the insider component is the price. Table 3’s ordering is derivable from this one axis.
The industrial chapter does not extend the quant chapter; it attacks the mixture’s observability itself, one force at a time.
Against the behavioral force: informed-looking opinion becomes free. §3’s repairs rest on Spence’s cost asymmetry: the signal separates types because faking it is expensive. A frontier model collapses the cost of producing fluent, evidence-flavored, confidence-calibrated prose about any event to fractions of a cent, flooding the discourse around every market with informed-looking noise. The stake still separates (an agent that must post collateral is subject to the same scoring rule as anyone), but the surrounding cheap-talk environment, which traders read for signals and narratives, is now synthetic at scale. Staked aggregation becomes more valuable relative to unstaked commentary, while the commentary itself becomes a manipulation surface.
Against the statistical force: sybil and coordinated flow break the decomposition’s assumptions. Every instrument in §2’s toolkit reads the shape of flow: many small independent tickets mean dispersed public wisdom; size-conditional calibration assumes a stake’s size reflects one participant’s conviction. Agents falsify the shapes cheaply. One principal operating a thousand wallets can manufacture exactly the signature the decomposition attributes to public wisdom; coordinated agent flow can paint a cascade onto the tape; wash trading becomes a programmatic product. The mixture does not merely shift; its observability degrades, because the forensic signatures were always proxies for identity and independence, and agents attack identity and independence directly.
Against the financial force: the insider problem goes from episodic to industrial. §4’s forensics catch the human insider because human insiders are operationally clumsy: one account, one large well-timed order, a visible spike before the announcement. An insider laundering knowledge through an agent swarm places no such print: the position is accumulated across hundreds of wallets, in ambient-sized orders, over hours: the timing signature blurs, the size signature disappears, and concentrated private knowledge arrives at the price wearing the statistical costume of dispersed public wisdom. This is the mixture thesis’s worst case: a price whose insider component has been engineered to be undecomposable. Anonymous venues are structurally defenseless against it; attributed venues are defended exactly to the degree that attribution is real: the design theory’s cue (§6).
And yet: the benchmark evidence says agents join the aggregation rather than beat it. On the standing forecasting benchmark, human superforecasters beat all seventeen frontier models tested; in late 2025 the first system reached parity, and the same technical report found it still underperforms liquid market prices. The best silicon now matches the best humans; both trail the staked crowd. Trading is harder still: set loose on live markets, only two of seven frontier models traded profitably, with overconfidence uniform across all seven: the Bayes half without the Kelly half loses (§2), and fluency in probability talk is not calibration under stakes. The market holds the top tier because its edge is the staked aggregation itself, never a smarter node. A frontier model, however capable, is one more unstaked opinion until it joins the book, and the winning configuration in the evidence is membership, not replacement: the ensemble of the parity-grade AI forecaster with the market consensus beat the market alone. The forward path is agents joining the aggregation: staking beliefs under bounded, capped mandates, their fills landing in the same auditable ledger as everyone else’s. Which mandates, whose caps, and what ledger is the design problem.
Assemble the four forces and one conclusion is forced. The statistical force says the price’s value is conditional: it is the best available estimator exactly when the mixture can be read (§2). The behavioral force says the mixture is permanent: the noise flow is the funding rather than a contaminant (§3). The financial force says the mixture is existential: the insider is at once the price’s information source and the venue’s trust liability (§4). The computational force says the mixture’s observability is under attack: agents can forge precisely the flow shapes the decomposition depends on (§5). The conclusion: a venue’s price stays trustworthy exactly as long as its mixture stays decomposable, and decomposability is a design property the venue must provide, not a market outcome. Traditional finance bought trustworthy prices with full permissioning (know-your-customer identity, licensed intermediaries, exchange surveillance) at the cost of open access. The claim of this section is that the same guarantees can be purchased without it: attribution where the traditional regime demands identity everywhere, bounded authority where it demands a license, per-action audit where it demands an intermediary.
Attributed identity. Every forensic instrument in §2 and §4 is a proxy for two facts the tape does not show: how many principals are behind the flow, and whether their signals are independent. The threat model makes attribution mandatory. §2 already conceded that a Kyle-optimal insider sizes to blend, and §5 showed agents driving that blending to the undecomposable limit, so once size and timing can be forged, the only invariant the forensics can still fall back on is who and how many. Attribution (a persistent, sybil-resistant mapping from accounts to principals) is what makes “many small independent tickets” mean dispersed wisdom rather than one agent swarm, and it is the only durable answer to §5’s laundering attack. It is derived from the threat, the fact the decomposition reduces to when its statistical signatures fail against a strategic adversary. Attribution need not mean publicity: the mapping can be held by the venue and audited by a regulator while wallets stay pseudonymous in public. What matters is that someone accountable can decompose.
Bounded authority. Participants (human and increasingly agentic) should act under explicit, sized, revocable mandates: position caps, rate limits, scoped permissions per market class. Bounds serve three masters at once: they keep bet size meaningful as a conviction signal (a cap regime is why the small position-capped venue’s prices were more consistent than the billion-dollar book’s); they cap the blast radius of a confused or adversarial agent, the precondition for admitting agents to the book at all; and they enforce §4’s hedge-versus-bounty line, since a bounded position cannot pay for outcome manipulation. The tension has to be named rather than buried, because it cuts against §2: the same cap that blunts a manipulator also throttles the conviction-sized informed order that §2 says carries private truth into the price fastest. A position cap therefore buys manipulation-resistance at the cost of some informational efficiency, an explicit design tradeoff. The optimal cap is an interior point, neither zero nor unbounded: a venue chasing the most informative price caps loosely, one chasing the fairest game caps tight, and Box 4’s condition (i) is how either checks that informed flow still sets the margin under the cap it chose. The mechanism is scoped, capped, time-bounded, revocable grants of authority, checked per action: rather than give the agent the wallet, give it a proxy holding a mandate, with every action authorized against the mandate and recorded.
Per-action audit. Every fill timestamped, sized, attributed, and retained, so that the §4 forensics can be run ex post by the venue, a regulator, or, in aggregate form, the public. The exchange-surveillance precedent says venues will buy this anyway to protect their noise flow (§4); the design claim is stronger: the audit trail is the instrument rather than overhead on it, because a price whose flow cannot be decomposed is just a number. The current landscape offers the pieces but never together: on-chain venues have public per-action records without attribution; regulated exchanges have attribution without openness. The trustworthy configuration needs both, with bounded authority connecting them.
Instrument, never decision-maker. Because this paper borrows so much of Hanson’s machinery, it must be maximally clear about the program it is not advancing. Hanson’s futarchy proposes governance by markets: values are voted, beliefs are bet, and the conditional-market comparison selects the policy: the market is the decision-maker. This paper’s use is markets inside a decision practice: the price is an instrument reading (the strongest available prior on a contested empirical question), deconfounded per the mixture and weighted per event class by Box 4’s conditions. The difference is a different failure surface. Futarchy inherits every pathology this paper has catalogued at the level of institutions: a distorted price does not merely misinform the decision, it is the decision. Instrument-use inherits the same pathologies as measurement error, which a practice can detect, weight, and survive. Futarchy bets the constitution on the mixture staying clean; this paper only ever bets a prior.
Agents as bounded participants. One product implication follows, and it runs against the reflexive one. §5’s evidence is that agents trail the staked crowd, and the design theory admits them only as bounded participants. The two combine into a single rule: an agent earns its place in the book by a sourced edge, a nameable information, modeling, or mixture-reading advantage (§2), staked under a capped, attributed, audited mandate, its fills scored on calibration like everyone else’s. The winning configuration in the evidence was the agent-plus-market ensemble, never the agent alone (§5), so a product that promises autonomous agents which simply beat the market is selling the one configuration the evidence rules out. What a venue can honestly offer is the harness: bounded authority, attribution, and audit that let a sourced edge compete while stopping an unsourced one from laundering.
Conclusion. The paper began with Galton’s ox and the observation that the canonical wise crowd was a staked crowd, and, on second look, already a mixture of fairgoers and butchers. Everything since has been that second look, run to its end. An event price is the clearing point where dispersed public wisdom and concentrated private knowledge settle their accounts: the noise funding the market’s existence, the informed collecting the noise’s losses as the wage of making the price true. What this paper adds to the mechanisms it inherits (from Galton and Hayek through Kyle, Kelly, Hanson, and Tetlock) is the assembly, and the design theory as its resolution: attributed identity, bounded authority, per-action audit. The stake is where the knowledge lives. The mixture is what the stake buys. And whether the mixture can be decomposed is a choice venues make rather than a property of markets, and the rest of us should read their prices accordingly.