Prediction Markets

Contribution of Szofi Varga

Introduction

In 2008 twenty-two scholars, most of them economists and among them Kenneth Arrow, Thomas Schelling, Vernon Smith and Robert Shiller, published a short policy statement in Science arguing that US regulators should remove the legal barriers to prediction markets (Arrow et al. 2008). Prediction markets, in their definition, are venues for trading contracts whose payoffs depend on the outcome of uncertain events. A contract that pays $1 if a candidate wins an election and nothing otherwise, trading at 53 cents, can be read as the market assigning that candidate a 53% chance of winning.

The case rests on two properties. The first is that such markets collect and aggregate information. What is relevant to an outcome is typically dispersed among many actors, and a market in which almost anyone can participate, and in which the prospect of profit and loss rewards the search for better information, is an effective way of bringing it together in a single price. The second is that the resulting forecasts are accurate: the authors point to growing evidence that prediction markets forecast with lower error than conventional methods. In the Iowa Electronic Markets’ vote-share markets for the 1988–2000 US presidential elections, the average absolute error in the week before the election was 1.5 percentage points, against 2.1 for the final Gallup poll; 150 days out it was around 5 points, at a horizon where polls read as forecasts erred considerably more.

On the strength of these two properties the authors argued that prediction markets could improve decisions well beyond forecasting — in business investment and in fiscal and monetary policy — and that the main obstacle was legal: federal and state anti-gambling laws cast enough doubt on such markets to deter their establishment in the United States. They proposed that regulators clear these barriers, beginning with a safe harbour from the CFTC for small-stakes markets run for research or internal use.

This wiki examines both properties and how the markets have developed since. Section “Theoretical Foundations” sets out the theoretical foundations, beginning with Arrow’s own work on state-contingent claims (Arrow 1964); whether prices aggregate information as the authors describe is taken up in Sections “Prices as information aggregators” and “Accuracy”, and whether they can be read as probabilities in Section “Interpreting a price as a probability”. Section “Historical Perspective” describes the Iowa Electronic Markets, Section “Current Framework” how current platforms operate and are regulated, Section “Market Microstructure” how the mechanics of trading shape prices, Section “Applications” how researchers and hedgers use them, and Section “Challenges” the problems that accuracy does not solve.

Theoretical Foundations

Event contracts as Arrow–Debreu securities

One of the earliest theoretical foundations for prediction markets is the system of state-contingent claims in Arrow (1964) (originally published in French, 1953). With $S$ states of nature and $C$ commodities, Arrow distinguishes two prices for the same good: $p_{sc}$, paid today for a claim delivering one unit of commodity $c$ if and only if state $s$ occurs, and ${\overline{p}}_{sc}$, the ordinary spot price of commodity $c$ paid after state $s$ has been revealed. Equation (3) (equation indices taken from the original paper) links them: $$p_{sc}=q_s{\overline{p}}_{sc},$$ where $q_s$ is the price of the $s$th elementary security—the contract paying one monetary unit if state $s$ occurs and nothing otherwise. This is precisely a binary event contract. Theorem 2 shows that $S$ such securities plus $C$ spot markets replicate the full system of $SC$ contingent-commodity markets, i.e. binary contracts span all state-contingent risk with $S+C$ markets instead of $SC$.

Equation (6) determines $q_s$ in equilibrium: $$q_s=\frac{\sum _{i=1}^I\sum _{c=1}^C{p}_{sc}{x}_{isc}^{*}}{y},s=1,\dots ,S.$$ Applying the feasibility condition (1), ${\sum }_{i=1}^I{x}_{isc}^{*}=x_{sc}$, the individual index drops out and this reduces to $$q_s=\frac{\sum _{c=1}^C{p}_{sc}{x}_{sc}}{y},$$ the market value of everything delivered in state $s$, as a share of the total money stock $y={\sum }_{i=1}^I{y}_i$. It is therefore an aggregate equilibrium price, not a belief. It coincides with the probability of state $s$ only under risk neutrality (linear $U_i$) and homogeneous beliefs; in general $q_s$ equals ${\pi }_s$ times a tilt given by marginal utility in state $s$ relative to its average, so claims on states where consumption is scarce trade above their odds. Arrow moreover allows subjective, individual-specific probabilities ${\pi }_s^i$, so even net of the risk tilt the state price aggregates disagreement rather than revealing any single belief.

A further result is equation (11), $$\sum _{s=1}^S{q}_s=1.$$ Arrow derives this from (6) and the budget constraints and then notes that it closes an arbitrage loophole: buying one of every security synthesises a sure monetary unit, which must therefore cost exactly one, so hoarding cash offers no advantage. This is the coherence condition familiar from prediction markets—prices across a mutually exclusive and exhaustive set of contracts sum to one. It equals exactly $1$ here because the model has no interest or time discounting; with a rate $r$ it would be $1/(1+r)$.

The mathematical formalisation of extracting probabilities from derivative securities is given by Breeden and Litzenberger (1978), whose title—“Prices of State-Contingent Claims Implicit in Option Prices”—makes the debt to Arrow explicit. Let $C(K,T)$ denote the price of a European call with strike $K$ and maturity $T$ written on an underlying $S$, let $r$ be the risk-free rate, and let $\mathbb{\mathbb{Q}}$ denote the risk-neutral measure, with density $q(\cdot )$ for $S_T$. Differentiating the call price with respect to the strike gives $$\frac{\partial C(K,T)}{\partial K}=-e^{-rT}\mathbb{\mathbb{Q}}(S_T>K),$$ so that a digital (binary) call paying one unit if $S_T>K$ has price $-\partial C/\partial K$, and the risk-neutral probability of the event is $$\mathbb{\mathbb{Q}}(S_T>K)=-e^{rT}\frac{\partial C(K,T)}{\partial K}.$$ Differentiating a second time recovers the entire risk-neutral density: $$\frac{{\partial }^{2}C(K,T)}{\partial K^2}=e^{-rT}q(K),\text{equivalently}q(K)=e^{rT}\frac{{\partial }^{2}C(K,T)}{\partial K^2}.$$ Given a sufficiently fine grid of traded strikes, the second derivative is approximated by a butterfly spread, $$C(K-\mathrm{\Delta }K,T)-2C(K,T)+C(K+\mathrm{\Delta }K,T)\approx e^{-rT}q(K)(\mathrm{\Delta }K{)}^2,$$ which pays approximately one unit if $S_T$ falls in a narrow band around $K$ and nothing otherwise—that is, an Arrow security on that band, synthesised out of options.

In prediction markets neither differentiation is necessary. A binary contract on an event $A$ is the digital claim, so its price is $e^{-rT}\mathbb{\mathbb{Q}}(A)$ directly; there is no strike grid to differentiate across, and in practice far too few strikes for a second difference to be numerically stable. Where a venue lists a strip of contracts on adjacent ranges of a price or a yield—Treasury yields and Bitcoin being the usual examples—the strip spans an exhaustive partition of the state space, and its prices give the discrete risk-neutral distribution outright, subject to the same coherence condition $\sum _s{q}_s=1$ noted above.

As in Arrow’s framework, what is recovered is $\mathbb{\mathbb{Q}}$ rather than the physical measure: the two differ by the risk tilt, and separating them requires additional structure of the kind assumed in Ross recovery.

Prices as information aggregators

A popular argument in favour of prediction markets is that they aggregate the wisdom of crowds, that through the buying and selling of event contracts traders have the opportunity to convey their beliefs, and the resulting price aggregates the knowledge of many, which somehow leads to accurate predictions. The underlying thesis behind this perspective can be traced back to the seminal paper of Hayek (1945), in which he argues against central planning offices and in favour of free market activities to achieve optimal allocation of resources. His central claim is that for centralised institutions to work efficiently, they need knowledge to be at their disposal, and therefore that knowledge should be easy to collect, measure and standardise - which is only true about scientific knowledge, however, that is a small and specific type of knowledge, and leaves very informative sources untapped. He argues that one of the most important types of knowledge is the knowledge of time and space, that of special circumstances that is learnt from daily frictions, of solving ad-hoc problems. And while it is essential for the stable functioning of systems, it cannot be collected or channelled into statistics, and therefore cannot be analysed and utilised by central planning entities.

He discusses that while this “intimate” knowledge is limited to specific circumstances, the interconnected nature of the economic system facilitates these specific “knowledges” through the most important source of information - prices. Therefore to achieve equilibrium it is not necessary for everyone to be aware of global phenomena affecting supply and demand conditions, but it is sufficient if each individual optimises within their own scope once observing prices. He uses the example of how people don’t need to be aware whether a price increase in tin is caused by a supply shortage or a demand spike - the price itself carries enough information for them to adjust their consumption in a way that brings the system back to equilibrium.

Hayek’s argument is informal; it asserts that prices transmit dispersed information but does not model how private signals become impounded into a single price, nor whether the resulting price can be trusted to carry all of it. Grossman (1976) supplies that mechanism. He studies a market with a safe and a risky asset and $n$ types of informed trader, where trader $i$ observes a private signal $y_i=P_1+{ϵ}_i$ — the unknown future payoff $P_1$ corrupted by an idiosyncratic noise term ${ϵ}_i$ that is independent across traders. Under constant-absolute-risk-aversion utility and normality, each trader submits a mean-variance demand, and because the market-clearing price is driven by aggregate demand it comes to depend on the whole signal vector only through its average $\overline{y}=\frac{1}{n}\sum _i{y}_i$. The equilibrium price is a strictly increasing linear function of that average, $P_0={\alpha }_0+{\alpha }_1\overline{y}$ with ${\alpha }_1>0$, so any trader who merely observes $P_0$ can invert it and recover $\overline{y}$ exactly. Reading the price is reading the consensus signal.

The consequence is that a trader’s own signal is worthless once the price is known. An individual signal $y_i$ carries noise of variance one, whereas the average $\overline{y}$ has noise variance $1/n$: pooling independent signals cancels their errors, so $\overline{y}$ is a strictly sharper estimate of $P_1$ than any $y_i$. Grossman shows further that $\overline{y}$ is a sufficient statistic for the signal vector, meaning that conditional on $\overline{y}$ an individual’s own $y_i$ conveys no additional information about $P_1$. The price therefore does not merely transmit information, it aggregates it, delivering to every participant, for free, a statistic that dominates the costly signal each of them could have acquired privately. This is the strongest possible formalisation of Hayek’s claim — the price is fully revealing — but it is precisely this strength that makes it self-undermining: if the price already supplies a better signal than the one a trader would pay for, no trader has any reason to pay.

Crucially, the noise that survives this construction is the wrong kind to rescue the incentive. Idiosyncratic signal noise is exactly what averaging destroys — the more of it there is, the more valuable the price’s aggregation becomes — so it cannot protect the informed. What restores an incentive is a second, non-informational source of noise. Grossman notes that if the total supply of the risky asset is itself random, the price becomes a function of two unknowns, the consensus signal and the supply shock, and a single observed price can no longer be inverted to recover $\overline{y}$: a high price might reflect a favourable signal or merely a low realised supply, and the observer cannot tell which. The price is then only partially revealing, and it is in this gap that a trader who has paid to observe the fundamental knows something the price alone does not, and can trade on it profitably.

This is also where the model turns against the very intuition it began by formalising. Hayek’s tin example presumes that the only sources of a price change are fundamentals — a supply shortage or a demand spike — both of which call for the same response, so it genuinely does not matter which one it is. Once non-fundamental supply noise is admitted, a price rise may instead be pure noise, calling for no response at all, and the trader now does need to know why the price moved. Grossman states the reversal directly: with noise in the system, “each individual wants to know why the price has risen.” The wedge between what the price says and why it moved is the same wedge that leaves a return to being informed; the two observations are one.

Grossman and Stiglitz (1980) make this systematic by letting the fraction of informed traders $\lambda$ be determined endogenously. Informed traders pay a cost $c$ to observe a signal $\theta$ about the return $u=\theta +ϵ$; uninformed traders condition only on price; and a per-capita random asset supply $x$ prevents the price from ever becoming fully revealing, since the uninformed cannot separate a high price caused by a favourable $\theta$ from one caused by low supply. The informativeness of the equilibrium price, measured by the squared correlation ${\rho }^2$ between price and $\theta$, is pinned down by the free-entry condition that informed and uninformed earn equal expected utility, yielding the compact relation $1-{\rho }^2=(e^{2ac}-1)/n$ in terms of the information cost $c$, the coefficient of risk aversion $a$, and the informativeness $n$ of the underlying signal. The two limiting cases make the knife-edge explicit. As $c\to 0$ the price becomes fully revealing but no one is compensated for information; and if the supply noise vanishes ($\mathrm{V}\mathrm{a}\mathrm{r}x\to 0$) the price again reveals $\theta$ exactly and, by their Theorem 5, no competitive equilibrium exists at all. The only consistent outcome for interior parameters is an equilibrium degree of disequilibrium: the price must stay partially uninformative so that the return to being informed exactly covers its cost. An informationally efficient price must leave a residual return to the informed, and that return is the equilibrium price of the information the market aggregates.

For prediction markets the implication is direct, and it qualifies the wisdom-of-crowds narrative rather than endorsing it. The naive version of that narrative is pure Hayek: the price aggregates everyone’s information, so one need only read the price. The Grossman–Stiglitz correction is that a market which did this perfectly would give no participant any reason to bring information to it in the first place. A real, functioning prediction market must therefore be leaving some return to informed traders, which means its price is not fully revealing, which means that price is a probability proxy carrying a residual noise wedge. Costly information acquisition, together with the non-informational trading that lets the informed profit, is thus not a defect of a prediction market but the very mechanism that makes it worth anyone’s while to bring information to it — a point that recurs in Section “Accuracy”, where the accuracy of contemporary platforms is traced to the trading of an informed minority rather than to the crowd as a whole. Moreover, one should remember that the probabilities reflected by prices are not physical, but rather the risk-neutral probabilities.

The wisdom of crowds

Another popular historical argument which is often mentioned when talking about prediction markets is the wisdom of crowds. This dates back to 1907, to an article published by Francis Galton in Nature (Galton 1907). He conducts a brief experiment about an ox auction where spectators have the opportunity to enter into a contest about guessing the weight of the ox as precisely as possible. He then tallies the tickets and performs some summary statistics on them, leading to the discovery that the median guess was $1207$ lbs, while the correct weight was $1198$ lbs. Therefore the median guesser, the vox populi, as he describes, was very accurate, making only a $0.75\mathrm{\%}$ error. Importantly, the crowd did not solely consist of cattle experts, but mostly bystanders. The conclusion Galton draws from this is that

“The average competitor was probably as well fitted for making a just estimate of the dressed weight of the ox, as an average voter is of judging the merits of most political issues on which he votes

$(\dots )$

”,

indicating that collective intelligence is the appropriate way of decision making regarding public issues as well.

For the interested reader we recommend the book of James Surowiecki, The Wisdom of Crowds (Surowiecki 2004), in which he elaborates on the conditions necessary for collective intelligence to flourish, its limitations, amongst other subtopics.

The application to prediction markets is straightforward: through the buying and selling of event contracts, traders push prices towards equilibrium which then reflects the opinion of the entire crowd. As will be demonstrated in Section “Accuracy”, the results of prediction markets are in line with Galton’s findings: collective intelligence proves to be a very good predictor, sometimes outperforming experts.

Historical Perspective

The Iowa Electronic Markets

One of the most important predecessors of today’s prediction markets is the Iowa Electronic Markets (IEM), initiated in 1988 by the University of Iowa Henry B. Tippie College of Business (Iowa Electronic Markets, n.d.). The IEM is a small-scale, real-money futures market that exclusively serves teaching and research purposes. It charges no trading commissions, and while it is specifically targeted as a teaching-enhancing tool, participation is not restricted to students: the political markets are open to traders worldwide, whereas the earnings and returns markets are open only to academic traders. Trading accounts are funded with between $5 and $500.

It facilitates real betting activities with contracts on election outcomes, economic indicators and stock returns. Similarly to other event contracts, each contract pays $1 if the outcome is realised and $0 otherwise.

The IEM facilitates market and limit orders, as well as bundles, i.e. a collection of contracts spanning all outcomes of a market that is bought or sold as a unit. While the platform does not allow short selling explicitly, one could replicate a short sale by the following action: imagine buying a fixed-price bundle of each contract listed in a market, which has a total cost of $1, then selling the contract the trader wishes to short. In this case the price paid for the bundle can be interpreted as a margin deposit that covers every possible contingency, which is precisely why the exchange can dispense with short sales altogether.

Research by Berg et al. (2008) utilising IEM data shows that the platform provides good predictions for presidential elections: relative to contemporaneous polls, IEM markets produce lower prediction errors on both short and long horizons, and perform especially well on long horizons.

Current Framework

This section describes how contemporary platforms operate (Section “Mechanisms”), what their prices can be taken to mean (Section “Interpreting a price as a probability”), and how they are regulated (Section “Regulatory overview”).

Mechanisms

A prediction market lists event contracts: claims on a specified future event that pay $1 if the event occurs and $0 otherwise. Each market consists of two complementary contracts, YES and NO, which trade at prices between $0 and $1. The price of the YES contract is commonly read as the market-implied probability of the event; Section “Interpreting a price as a probability” examines when that reading is justified. This subsection describes the arrangements that produce the price: the structure of the venue, the listing of markets, the life-cycle of an order from submission to settlement, and the fees charged along the way.

Exchange structure

The platforms discussed here are organised as exchanges rather than as bookmakers. A sportsbook sets its own odds and is the counterparty to every wager it accepts. A prediction market platform does neither: participants trade with one another, prices are determined by their supply and demand, and the platform earns its revenue from trading fees (Section “Fees”) rather than from taking the other side of its customers’ positions.

Two consequences recur throughout this wiki. First, since there is no house to set a price, a price exists only where participants are willing to quote one, so the depth and quality of a market depend on who chooses to supply liquidity (Section “Liquidity provision and the maker–taker structure”). Second, since positions remain tradable until the market closes, a participant is not committed to an initial position in the way the holder of a fixed-odds betting slip is (Section “Holding and closing a position”).

Market listing

Markets are listed by the platform. Proposals originate either with the platform itself or with users, and are reviewed by the platform’s contracts team before listing. A market is defined by two separate elements: a question with a binary answer, and a set of resolution rules specifying how that answer will be determined (Section “Resolution and settlement”). On Kalshi, which operates as a designated contract market under CFTC oversight, contracts are listed through the CFTC’s product certification process; KalshiEX LLC (2026) is an example of such a filing.

Order submission and execution

Most platforms match orders through a central limit order book (CLOB); see Harris (2003) for a practitioner treatment and Foucault et al. (2013) for a formal one. Figure 1 traces a single order from submission to settlement; this and the two subsubsections that follow describe its stages in turn.

Complementary books. The YES and NO books are not independent. Since one YES and one NO contract together pay exactly $1 whatever the outcome, a bid for YES at price $p$ is economically equivalent to an offer of NO at $1-p$. A trade can therefore arise from two opposed views as well as from a buyer meeting a seller: a bid for YES at $0.40 is matched by a bid for NO at $0.60, the two buyers jointly contribute the $1 that the pair will pay out, and a new YES–NO pair is in effect created. The first trade in a newly listed market necessarily takes this form, since no contracts yet exist to be sold.

Writing $a$ and $b$ for the best ask and best bid on YES, the best ask and best bid on NO are $1-b$ and $1-a$, so that $$a^{\mathrm{Y}}+a^{\mathrm{N}}=1+s,b^{\mathrm{Y}}+b^{\mathrm{N}}=1-s,$$ where $s=a-b$ is the quoted spread. Buying both sides costs $1 plus the spread, selling both yields $1 less the spread, and the two midpoints sum to exactly one. This is the coherence condition $\sum _s{q}_s=1$ of Section “Event contracts as Arrow–Debreu securities” in its two-state form, holding up to the spread.

Market orders. A market order executes immediately against the best quotes on the opposite side of the book: a buy at the best ask, a sell at the best bid. If the quantity available at the best price is smaller than the order, the remainder executes at the next price level, and so on — the order walks the book. Suppose the best ask is $0.30 for 100 contracts and the next is $0.35 for 200. A market buy of 200 contracts executes 100 at $0.30 and 100 at $0.35, at an average price of $$\overline{p}=\frac{100\times 0.30+100\times 0.35}{200}=0.325,$$ 2.5 cents above the best ask. The shortfall grows with the size of the order relative to the depth at the top of the book.

Limit orders. A limit order specifies a maximum purchase price or a minimum sale price $p$. If it is marketable — a buy at or above the best ask, or a sell at or below the best bid — it executes on arrival against resting orders priced at $p$ or better. Any quantity that cannot be executed on arrival is handled according to the order’s time in force. Platforms usually support four:

• good-till-cancelled (GTC): the remainder rests in the book until it is filled or cancelled;
• good-till-date (GTD): as GTC, but the order expires automatically at a specified time;
• fill-or-kill (FOK): the order executes in full on arrival or is cancelled in full;
• fill-and-kill (FAK, also known as immediate-or-cancel): whatever can execute on arrival does, and the remainder is cancelled.

A resting order may be filled in several parts by successive counterparties, and may be cancelled by its owner at any time before it is filled.

Makers and takers. An order that rests in the book supplies liquidity, and its owner is the maker of any trade in which it is later filled; an order that executes against a resting one consumes liquidity, and its owner is the taker. A market order is always a taker. A limit order can be either, or both, if part of it executes on arrival and the remainder rests. The distinction determines fees (Section “Fees”) and is the organising variable of Sections “Liquidity provision and the maker–taker structure” and “Informed trading and adverse selection”.

Choice of order type. The two order types trade price risk against execution risk. A market order secures execution, given sufficient depth, but not the price, as the example above shows. A limit order secures the price but not execution: one placed far from the best quotes may never be filled.

Figure 1. Life-cycle of an order on a central limit order book, from submission to settlement. Diamonds are decisions; grey boxes are executions, dashed boxes cancellations. (a) $p$ is the limit price. Time in force governs only the quantity that cannot execute on arrival, so a fill-or-kill order that cannot be filled in full is cancelled without any part executing. Every fill opens, adds to or reduces a position. (b) An offsetting order is an ordinary order and re-enters at the top of panel (a). The resolution branches correspond to Kalshi’s centralised process and Polymarket’s UMA optimistic oracle; see Sections “Resolution and settlement” and “Resolution”.

Holding and closing a position

Trading is continuous until the market closes. A participant can therefore close a position at any point before then by an offsetting trade — selling the contracts held, or equivalently buying the complementary side — and so realise a gain or limit a loss before the outcome is known. This distinguishes prediction markets from fixed-odds betting, where a wager is held to settlement unless the bookmaker offers to buy it back, on terms the bookmaker sets.

Resolution and settlement

The final stage is resolution. Once a market is resolved, trading ends: each winning contract is redeemed for $1 and each losing contract expires worthless. Although every market poses a binary question, the rules by which it is answered are specified separately. They name the sources that determine the outcome, the date from which the market becomes eligible for resolution, and the treatment of ambiguous cases. Determining the outcome is not always straightforward, and platforms allocate the decision differently. Kalshi resolves centrally: the outcome is determined by the exchange itself. Polymarket resolves through UMA’s optimistic oracle: a proposed outcome, backed by a bond, stands unless it is disputed within a set window, in which case it is decided by a vote weighted by staked UMA tokens. Section “Resolution” describes both mechanisms in detail and examines how each can fail.

Fees

As on securities exchanges, fees depend on whether an order supplies or consumes liquidity. Makers pay very small fees (Kalshi) or receive rebates (Polymarket), while takers generally pay somewhat more. The aim is to encourage resting orders and so deepen the book, which should in principle improve price discovery; Section “Liquidity provision and the maker–taker structure” examines how far the incentive explains who actually quotes. Fees are a function of the number of contracts, the contract price and a fee rate that varies across market categories. Charges for transferring cryptocurrency are set by third parties rather than by the platforms; Kalshi charges 2% on deposits made by debit card.

Interpreting a price as a probability

The claim that a contract trading at 40 cents means a 40% chance is the one that makes prediction markets interesting to anyone outside them. It is also the claim most often asserted and least often tested. This subsection sets out what would make it true, what stands in the way, and what the evidence shows. Two related questions are handled elsewhere and it is worth separating them now: whether these prices are good forecasts, and where that quality comes from, is the subject of Section “Accuracy”; whether the number can be read as a probability at all is the subject here.

What would make the claim true

The property at issue is calibration. Let $Y\in \{0,1\}$ be the outcome and $\hat{p}$ the price. A forecast is perfectly calibrated if $$\mathbb{\mathbb{P}}(Y=1\mid \hat{p}=p)=p\text{for all }p\in [0,1],$$ that is, if among all contracts priced at 40 cents the event occurs 40% of the time. Since $\hat{p}$ is continuous, this is assessed over bins, and plotting realised frequency against quoted price gives a reliability diagram, in which perfect calibration is the 45° line. The standard summary statistic is the Brier score (Brier 1950), $$\mathrm{B}\mathrm{S}=\frac{1}{N}\sum _{i=1}^N(p_i-Y_i{)}^2,$$ which is zero for a perfect forecast and 0.25 for always saying 50% against a 50/50 base rate.

The reason to care about calibration specifically, rather than about accuracy, is that it is what makes a probability claim falsifiable at all. A single probabilistic statement about a one-off event cannot be checked; it is only by pooling many instances in which the same probability was quoted, and appealing to the law of large numbers, that the claim acquires empirical content. Without calibration a probability forecast is an unfalsifiable opinion; with it, it makes a testable prediction about long-run frequencies (Kagan and Baiocchi 2026).

Two theoretical obstacles

The first is Manski (2006). When traders hold heterogeneous beliefs, the equilibrium price is determined by the marginal trader and by the distribution of stakes, and is not in general a sufficient statistic for the average belief of participants. It is not that the price means nothing; it is that the step from price to “what the participants think” requires assumptions about risk preferences and position sizing which the market itself does not supply. Those assumptions can be stated: with logarithmic position sizing and a symmetric distribution of beliefs, the average-belief interpretation is recovered exactly (Wolfers and Zitzewitz 2006). The objection is therefore conditional rather than fatal, but the conditions are substantive and are rarely checked.

The second is the one Section “Event contracts as Arrow–Debreu securities” closed on and is independent of the first. What a traded price recovers is the risk-neutral probability $\mathbb{\mathbb{Q}}$, which equals the physical probability $\mathbb{\mathbb{P}}$ only under risk neutrality. The two differ by a factor given by marginal utility in the state concerned, so a claim paying $1 in a state investors care about should trade above its physical probability. For most contracts the tilt is plausibly small; for macroeconomic ones it is not, as Section “Academic application examples” argues. A market can therefore be perfectly calibrated against $\mathbb{\mathbb{Q}}$ and still not report $\mathbb{\mathbb{P}}$.

The evidence

Calibration has now been measured at scale. use the complete resolved history of Kalshi — 2,243,741 markets across eleven categories from the platform’s launch in 2021 to mid-2026 — and find prices that are, in aggregate, very well calibrated, and increasingly so as resolution approaches. Pooled Brier scores fall from roughly 0.087 at a three-month horizon to below 0.02 at close; naive accuracy rises from 88.3% to 97.2% over the same range; and reliability diagrams track the 45° line closely in nearly every category. This is the largest such exercise conducted on a regulated exchange, and it is the strongest available support for the literal probabilistic reading. However, it should be noted that the paper is an internal research report published by Kalshi, moreover, it does not address the difference between risk-neutral and physical probabilities.

Some of the interesting findings of the paper are the following: Participation, and its limits. Calibration improves near-monotonically with both dollar volume and the number of unique traders. Event volume below $10,000 gives a Brier score of 0.0635 at close, against 0.0099 above $200,000. But the improvement is bounded by how much genuine uncertainty remains: three-month markets never reach a Brier score of 0.05 however many traders participate; one-month markets reach it only at about 1,900 traders, which describes 1% of markets at that horizon; one-week markets at 901; and one-day markets at 20. No amount of trading substitutes for the passage of time.

Category and clock. Politics and Economics markets fall below a 0.05 Brier score at volumes of a few thousand dollars, while Sports and Mentions need roughly two orders of magnitude more. And the measured picture depends on a choice that looks purely administrative: anchoring the horizon to the true occurrence of the underlying event, rather than to the platform’s market-close timestamp, improves apparent calibration in Sports and removes a bias in Elections. The authors believe this re-scaling is novel, and it is a caution for anyone computing horizons from exchange metadata.

How much this settles

However, there is a tension with Section “Accuracy”, which should not be left unremarked. Calibration improving near-monotonically with the number of traders reads naturally as support for the wisdom-of-crowds account: more participants, better prices. Yet the trader-level evidence in Section “Accuracy” finds that roughly 3% of accounts supply the informational content and that the rest, on net, move prices away from the outcome. These are not strictly contradictory — a market that attracts more traders may attract more skilled traders, and greater volume also means tighter spreads and less execution noise, either of which would improve calibration without any contribution from the median participant. But the reconciliation is an assumption rather than a finding, and Kagan and Baiocchi (2026) say as much, naming the joint relationship between participation, trader skill and calibration as the open question their results raise.

The defensible summary is therefore narrower than the slogan. A liquid Kalshi contract in mid-life, in a category with a reasonable information-discovery window, is a well-calibrated forecast of its own resolution criterion, and demonstrably so. It is a risk-neutral probability rather than a physical one. Whether it is also the average belief of its traders is a separate claim requiring assumptions that Manski (2006) showed are not innocuous.

Regulatory overview

No jurisdiction has legislated a bespoke regime for prediction markets. Every regulator that has addressed them has instead asked which existing perimeter they fall into — financial instrument, gambling product, or neither — and the jurisdictions surveyed here have answered that question in different ways. The European Union has left the allocation unresolved between securities and gaming authorities; the United States has federal and state regulators each claiming the whole field; the United Kingdom has divided the category between two regulators that publicly agree on where the boundary sits.

Current as of 17 August 2026. Litigation in the United States and national enforcement in Europe were both active at the time of writing; the dated claims below should be re-checked before citation.

United States

The United States took the opposite path from the European Union: rather than absorbing prediction markets into an existing prohibition, it established federal jurisdiction over them through litigation, and then spent two years contesting how far that jurisdiction reaches. The result is a market that is federally lawful, commercially large, and under sustained attack from state regulators.

The 2024 turning point. The federal framework originates in KalshiEX LLC v. CFTC, and it began as a defeat for the regulator. In 2023 the CFTC ordered Kalshi, a designated contract market (DCM), to withdraw contracts on which party would control Congress, invoking the Commodity Exchange Act’s “Special Rule” (§ 5c(c)(5)(C)), which allows it to prohibit event contracts involving “gaming” or activity unlawful under state or federal law. Kalshi sued and won: on 12 September 2024 Judge Jia Cobb (D.D.C.) held that “gaming” means the playing of a game and does not extend to elections, and that since elections are not themselves unlawful, the contracts did not involve unlawful activity. The D.C. Circuit denied the CFTC an emergency stay on 2 October 2024, finding its concerns about election integrity and manipulation “fail to rise beyond the speculative level.” Election contracts traded openly through the 2024 presidential cycle, and the CFTC dropped its appeal in May 2025 — establishing by litigation, not statute, that event contracts on a registered exchange are federally regulated financial instruments.

Federal accommodation and political economy. The agency’s posture then reversed: having failed to exclude these products, the CFTC moved to supervise them, issuing a staff advisory and advance notice of rulemaking in March 2026 and a formal proposed rule on 11 June 2026. This is the substantive contrast with Europe, where ESMA applied an existing retail prohibition instead. The reversal also coincided with a change of administration, and the accommodation has been explicit. Federal investigations into Polymarket were closed in 2025. Donald Trump Jr. became a paid strategic adviser to Kalshi in January 2025 and joined Polymarket’s advisory board in August 2025, after his venture firm 1789 Capital invested a reported “double-digit millions”; Trump Media announced its own event-contract product with Crypto.com in October 2025. On 28 May 2026 the President posted that CFTC exclusive jurisdiction was “critically important,” praising its chairman and attacking opposing state officials by name (Crypto Times 2026). Critics, including state attorneys general, read this as regulatory capture; the agency’s defenders respond that its position follows directly from the 2024 rulings. Both readings are available on the public record.

The preemption fight. Federal accommodation collided with state gambling authority once platforms moved from elections into sports, where the volume is much higher. States issued cease-and-desist orders — Nevada, New Jersey, Maryland, Massachusetts, Arizona, Ohio and others — and the platforms sued, arguing that the CEA’s exclusive jurisdiction over DCM trading preempts state gambling law. Results diverged sharply:

• KalshiEX LLC v. Flaherty, No. 25-1922 (3d Cir., 6 April 2026) — affirmed a preliminary injunction against New Jersey, holding sports event contracts are likely “swaps” and that the CEA likely occupies the field of DCM trading regulation. Judge Roth dissented, calling them “virtually indistinguishable from online sportsbooks” (Holland and Knight 2026).
• Ohio (10 March 2026) — sports contracts are not swaps; relief denied.
• Tennessee (20 February 2026) — preliminary injunction granted, contradicting Ohio.
• Ninth and Fourth Circuits — argued April and May 2026, pending.

The Third Circuit ruling is narrower than its headlines suggest: a preliminary-injunction standard requiring only a “reasonable chance” of success, expressly not a merits determination. Federal oversight has therefore not replaced state-by-state regulation — that displacement is the contested question, not a settled outcome, and a circuit split makes Supreme Court review a possibility.

States mobilise. On 28 July 2026, 44 state attorneys general filed a joint comment letter opposing the CFTC’s proposed rule, arguing the agency “goes well beyond its statutory authority” and would “drastically expand federal regulatory authority in an area of major economic and political consequence that states have traditionally regulated” (The Block 2026). The framing — “States have long regulated gambling, including sports bets. The federal government has not.” — invokes the major questions doctrine. Three days later New York sued Kalshi for operating an unlicensed gambling operation, seeking treble damages, forfeiture and restitution, and objecting that the platform admits users aged 18 to 20. The CFTC intervened against the state, its chairman objecting that New York sought “to force an unprecedented sudden shutdown” (PYMNTS 2026) — a federal financial regulator litigating to shield a platform from state law enforcement, which measures how far the jurisdictional question has hardened.

Conduct and consumer protection. A separate line of pressure concerns marketing, originating in a Wall Street Journal investigation in mid-2026. The Journal reviewed over 1,100 videos posted between December 2025 and May 2026 across TikTok, YouTube and Instagram, drawing more than 140 million views, in which paid creators displayed apparently profitable Polymarket trades. Roughly 70% showed bets never placed on the live platform; in 118 videos celebrating some $900,000 in winnings, blockchain records indicated the positions would have lost more than $166,000. Creators, paid $2,000–$3,000 a month, filmed on password-protected clones of the site that simulated trading, and initially did not disclose the payments (Tech Startups 2026). Senators Curtis and Schiff requested a CFTC investigation on 26 June 2026, and on 12 August 2026 New York City Council Speaker Julie Menin opened an inquiry into deceptive advertising and marketing to young adults, writing to Polymarket, Kalshi, Gemini Titan and Coinbase (CBS News 2026).

This matters beyond the conduct itself. On-chain verifiability is what exposed the fabrication — the same transparency treated in Section “Market Microstructure” as an analytical advantage also functions as a forensic audit trail. And it shifts the regulatory question from whether these instruments are swaps to how they are sold to retail users, the ground European regulators had already occupied.

Spillover into market structure. Federal status has also enabled expansion beyond event contracts. On 18 June 2026 CME Group sued the CFTC, challenging its approval of Kalshi’s perpetual bitcoin futures contract and the policy statement letting other DCMs self-certify similar products. CME argues these are swaps rather than futures — they involve funding-rate payments and lack a set delivery date — so treating them as futures evades swap dealer registration, margin and reporting requirements (Dechert LLP 2026). Kalshi’s perpetuals crossed $1 billion in volume within a week of launch. The suit is not about prediction markets as such, but shows the same dynamic: an accommodating CFTC, and established parties litigating to constrain it.

European Union

The European Union has no bespoke regime for prediction markets. Unlike the United States, where the CFTC has moved toward a tailored framework for event contracts, EU law contains no standalone legislation on the category. The regulatory question in Europe is therefore not what the rules for prediction markets are, but which existing perimeter — financial services or national gambling law — a given contract falls into.

MiFID II and the financial instrument test. The threshold question is whether an event contract is a “financial instrument” within Section C of Annex I to MiFID II. Three points matter most: C(4), derivatives on financial underlyings; C(9), financial contracts for differences; and the catch-all in C(10), covering cash-settled derivatives on climatic variables, freight rates, inflation or other official economic statistics, and “any other derivative contracts relating to assets, rights, obligations, indices and measures” that have the characteristics of derivative financial instruments. The decisive variable is the nature of the underlying: an event contract referencing an enumerated financial variable — whether an equity index closes above a given level, for instance — is likely to qualify, while a contract on a sporting result or a purely political outcome falls outside the perimeter. The Commission has confirmed this reading in its MiFID II Q&As, stating that a sports or political CFD not linked to a Section C(10) variable would not be covered (Norton Rose Fulbright 2026).

Where the underlying is financial, event contracts closely resemble binary options: both pay a fixed amount or nothing depending on whether a specified condition holds at expiry. The Commission has confirmed that binary options meet the MiFID II definition, and ESMA defines a binary option as a cash-settled derivative paying “a predetermined fixed monetary amount or zero” contingent on specified events. This distinguishes them from contracts for differences, whose payoff varies proportionally with the underlying. Classification nonetheless remains fact-sensitive: a fixed payoff on a discrete binary event arguably lacks the value-tracking characteristic of a derivative financial instrument.

Consequences of classification. A platform offering qualifying contracts is subject to the full MiFID II regime — authorisation, conduct of business, appropriateness and suitability, product governance and best execution — and binary options are not “non-complex” instruments, so they cannot be distributed execution-only without an appropriateness assessment. More restrictively, ESMA’s May 2018 product intervention decision under Article 40 MiFIR prohibits marketing, distribution or sale of binary options to retail clients; the measure has since been made permanent at national level across the Member States. Retail access to financially-referenced event contracts in the EU is thus effectively closed.

ESMA’s July 2026 statement. On 3 July 2026 ESMA issued a public statement applying the national binary options intervention measures directly to event contracts (European Securities and Markets Authority 2026). It confirms the underlying-based test — only contracts referencing Section C(4)–(10) underlyings are financial instruments — and states that “the commercial name provided by firms (e.g. ‘event contracts’) is irrelevant” to classification. Firms must analyse product substance themselves. The statement is a supervisory convergence measure rather than new law, but it removes much of the ambiguity the perimeter question previously carried.

National enforcement. Contracts outside the MiFID II perimeter are left to national gambling law, where Member States retain broad discretion and there is no sector-specific EU legislation. It is through this channel, not through securities enforcement, that most restrictions have actually been imposed. Since late 2024 a substantial share of the EU has closed off retail access to Polymarket, the largest platform, on the ground that event contracts on political and sporting outcomes constitute unlicensed betting:

• France — formal warning November 2024, voluntary geoblocking, then an ISP blocking order in July 2026.
• Belgium — full block, February 2025.
• Poland — blacklisted, existing positions close-only, January 2025.
• Portugal — full block, January 2026, on the basis of a statutory prohibition on betting on political events.
• Hungary — block framed as temporary, January 2026.
• Spain — ISP block, May 2026.
• Italy — blocked October 2025, revoked December 2025, reimposed July 2026.
• Germany — declared unlicensable and publicly warned against, September 2025, but not blocked.
• Netherlands, Greece, Romania, Czechia — restrictions or blocks reported, 2025–2026.

Three cases illustrate the range. In France, the regulator obtained voluntary geoblocking in 2024, found the platform still drawing over 200,000 unique French visitors in June 2026, and ordered ISPs to block the site and its mirrors (iGaming Business 2026). In Italy, the platform was blocked in October 2025 under Article 4 of Law No. 401/1989, which prohibits collecting bets without a concession, enforced through the ISP-blocking power in Article 102 of Law Decree No. 104/2020; the block was lifted in December 2025 only after the operator withdrew real-money trading for Italian users — no court ruled on the merits — before being reimposed in July 2026 (Coraggio 2026). Germany shows the opposite failure mode: the Fourth Interstate Treaty on Gambling permits licensing only for defined sporting events, so election betting is not merely unlicensed but unlicensable, yet the regulator issued a consumer warning rather than a blocking order (Focus Gaming News 2025), and markets on the September 2026 state elections remained live and visible to German users (Deutsche Welle 2026).

The contrast is the practical lesson of this period. Where a national authority holds and uses ISP-blocking powers, access closes quickly; where it relies on payment friction and warnings, a crypto-settled offshore platform continues to operate on contracts the law plainly prohibits. Legal clarity and effective enforcement are separate variables.

No unified EU picture. There is no single EU-wide register, list or guidance stating where prediction market platforms may lawfully operate or be accessed. No major platform holds a licensed European operation (Finance Magnates 2026), and because gambling competence is national while financial classification is contract-by-contract, any assessment of availability must be made jurisdiction by jurisdiction. Nine European regulators issued a joint statement against unlicensed platforms in June 2026, but coordination of this kind is not harmonisation. Cassar (2026) notes the absence of harmonised terminology and EU-wide guidance, and proposes a structured “Prediction Test” classifying a contract by asking in sequence whether it falls within MiFID II, whether it is prohibited as a binary option or on public policy grounds, and, by exclusion, whether national gambling rules apply. No such test exists in EU law, and the practical map remains fast-moving: statuses range from full blocks to close-only positions to view-without-trading, and change on a timescale of months.

Two features of this landscape matter analytically. First, enforcement has fallen on precisely the contracts that sit outside the MiFID II perimeter — political and sporting underlyings, which the Commission places beyond Section C. The EU therefore restricts prediction markets from both directions, financially-referenced contracts by the retail binary options prohibition and non-financial ones by national gambling law, leaving little space in between. Second, no rule allocates a contract to the securities or the gaming regulator ex ante, so platforms cannot invoke a financial characterisation to escape the gambling regime while remaining unlicensed under both. This is the mirror image of the United States, where CFTC exclusive jurisdiction over event contracts as swaps, reinforced by the Third Circuit’s April 2026 ruling, aims to resolve by preemption what remains unallocated in Europe.

United Kingdom

The United Kingdom resolves the classification problem that Europe left open and the United States is still litigating: it splits the category in two and assigns each half to a different regulator, according to the nature of the underlying event. Both have now stated their positions explicitly.

Non-financial contracts: the Gambling Commission. Contracts on elections, sporting results and other real-world outcomes fall under the Gambling Act 2005. In a blog post of 4 February 2026, the Commission’s director of strategy stated that prediction market platforms as currently structured fall within the statutory definition of a “betting intermediary” — the classification already applied to betting exchanges, licensed in Great Britain since 2000, which makes the product category considerably less novel here than in the United States (Enright 2026). Unlicensed operators should not target UK consumers, and criminal offences attach to operating without a licence.

Financial contracts: the FCA. When the UK transposed MiFID II, HM Treasury extended financial services regulation to binary options referencing financial instruments, indices, economic statistics, commodities, interest rates or climatic variables (Cleary Gottlieb 2026). The FCA’s March 2026 perimeter report confirms the split: prediction markets on non-financial events belong to the Gambling Commission, while those referencing financial or climatic events constitute binary options (Financial Conduct Authority 2026). That classification is decisive, because the FCA’s permanent ban on selling, marketing and distributing binary options to retail consumers, in force since March 2019 (Financial Conduct Authority 2019), applies to them. Such contracts therefore cannot lawfully be offered to UK retail consumers at all.

The resulting position. Which half a platform’s contracts occupy determines not just the applicable rulebook but whether the product is offerable at all. Non-financial contracts have a compliance path — a Gambling Commission licence as a betting intermediary. Financial contracts have none for retail, and no major platform held a UK licence as of mid-2026.

Enforcement. Clarity of classification has not produced control of access. Reporting in July 2026 found UK users reach Polymarket straightforwardly via a VPN and cryptocurrency deposits; a consumer advocate quoted there said working out how took about ten minutes, and that the platform’s warnings against VPN use are not, and realistically cannot be, policed (The Guardian 2026). Polymarket said such access breaches its terms of service and that it was strengthening on-chain monitoring and third-party compliance tools. Volumes on UK political events, including individual byelections, suggest the circumvention is not marginal. The UK thus reproduces the German pattern: a prohibition legally unambiguous and practically porous, because a crypto-settled offshore platform remains reachable whatever the rulebook says.

Comparison. The UK reaches substantively the same retail outcome as the EU, but by a clearer route. Where EU law leaves classification unallocated between securities and gaming authorities, and US federal and state regulators each claim the whole field, two UK regulators have publicly agreed where the boundary sits, so operators can determine their position in advance rather than discover it through enforcement. What the UK shares with both is the gap between that position and its practical effect.

India

India represents a third model, distinct from both the European split perimeter and the American jurisdictional contest: prediction markets are prohibited not as mis-classified financial instruments but as a subset of online money gaming, and the question of whether their underlying event is financial never arises.

The instrument is the Promotion and Regulation of Online Gaming Act 2025, which took effect on 1 May 2026 and prohibits online money gaming generally. The Ministry of Electronics and Information Technology (MeitY), rather than a securities or gaming regulator, administers enforcement. On 25 April 2026 the Ministry issued an advisory to VPN service providers, warning that users were continuing to reach “illegal and blocked prediction market and online betting platforms” and that enabling such access carried legal consequences. Polymarket became inaccessible to Indian users on or around 22 May 2026, with blocking orders issued to internet service providers (CoinDesk 2026). Kalshi withdrew rather than waited: on 17 June 2026 it added India to its restricted jurisdictions list, amending its member agreement to prohibit trading by users “domiciled in, organised in or located in India” (Reuters 2026).

Two details are worth noting for comparison. First, the Indian ban required no classification analysis at all — where EU and UK regulators had to determine whether a contract referenced a financial underlying, the Indian statute captures the activity by its economic form, so the distinction that structures European regulation is simply absent. Second, MeitY addressed the circumvention channel directly by warning VPN providers, which is precisely the vector that left the German and British prohibitions porous. A CFTC-regulated platform withdrew from the market voluntarily, an outcome no European measure has produced.

Market Microstructure

Section “Mechanisms” described the mechanics of prediction markets: the order book, order types, fee structures, and the resolution process. This section asks how these features affect the prices one observes on prediction markets.

Knowing the underlying processes, for example, how orders are matched, tells us how a price is generated, however, it does not tell us whether the resulting number can be read as a probability, who pays for the market to exist, or whose beliefs the price actually reflects. Those are questions about microstructure: the study of how the mechanics of trading shape the prices that emerge from it.

Prediction markets inherit the standard conceptual framework for these questions from equity markets, where it was developed by Kyle (1985) and Glosten and Milgrom (1985). Readers who want that theory in full will find an accessible treatment in Harris (2003), written for practitioners, and a formal one in Foucault et al. (2013). For those interested, we warmly recommend the course of Prof. Rindi, “Market structure and information based models”.

What follows uses the results from the literature above but concentrates on the places where event contracts behave differently from the instruments the theory was built for. There are three such differences, and all of them come from the payoff structure: the price is bounded between $0 and $1, the contract ends at exactly $0 or exactly $1 rather than at some value along a continuum, and there is no underlying asset in which a position can be hedged.

The section runs from mechanism to outcome: first the data, then liquidity, informed trading, and price discovery, and finally the two questions the literature cares most about — whether these prices are accurate, and whether they are biased.

Measurement: what the data show

Prediction markets provide a very rich bundle of information for researchers and the public, but different platforms have different advantages as data sources.

Kalshi records, for every trade, which side crossed the spread to initiate it. This is a very important source of information, which is usually not published in equity markets, so researchers must guess it from the pattern of prices and quotes, and every estimate built on that guess inherits its error. This information is crucial because in every trade one side is paid for providing liquidity — the market maker resting a limit order — while the other pays for it — the liquidity taker crossing with a market order. The trade record alone does not say which counterparty was which. Getting this wrong would not result in a slightly incorrect number in the calculation of the effective spread, the standard measure of what a trade cost the impatient side, but a sign flip: the same trade would appear either to have earned the market maker money or to have lost it.

Kalshi also publishes its quoted bid and ask prices, and because every contract ends at $1 or $0, the profit or loss of the trader who supplied liquidity on any given trade can be calculated exactly rather than approximated (Bartlett and O’Hara 2026). Equity data offer none of these things.

Polymarket data has a different advantage: because it settles on a public blockchain, every trade is visible and can be traced to the wallet that made it. Researchers can therefore follow individual traders across markets and over time. Gómez-Cram et al. (2026) use this to study 1.72 million accounts across 210,322 markets and $13.76 billion of volume, observing every trade each account ever made. Data of that kind are rare in finance, where individual behaviour is usually inferred from aggregates, and it is what makes the trader-level research possible.

Polymarket’s public record has a limit however. It shows what was settled, not who initiated it, and inferring the initiator from the public order-book feed turns out to be barely better than guessing, only being correct around 59% of the time, against roughly 81% for the equivalent procedure on equities (Dubach 2026). Estimates that depend on knowing who initiated a trade are therefore less reliable on Polymarket than on Kalshi, and the problem is discussed properly in Section “Applications in research”.

Liquidity provision and the maker–taker structure

An equity market maker who accumulates inventory holds an asset whose value moves continuously and which can be hedged, at a cost, in a correlated instrument. An event contract market maker who accumulates inventory holds a claim that will be worth exactly 0 or exactly 1 at a known date, with no underlying in which to offset the exposure. Inventory risk is therefore not the risk of adverse drift but of a discrete terminal jump, and it cannot be laid off.

This has consequences for liquidity provision: Bartlett and O’Hara (2026) report that the large equity market-making firms (Citadel Securities, IMC, Hudson River Trading) stayed away from Kalshi over their sample period, deterred in part by the lack of hedging instruments, while the institutional makers that did enter came from sports-betting rather than equity-market-making backgrounds.

The exchange operates two tiers of provision: a gated Market Maker Program under which designated firms accept two-sided quoting obligations in exchange for reduced fees and adjusted position limits, and an open Liquidity Incentive Program rewarding ordinary members for posting limit orders. Only about 5% of matched bids come from large institutional makers, the remainder from some two thousand smaller participants, most of whom are individuals. Why they are willing to do it is not fully explained by the exchange’s incentives. Palumbo (2026) finds that Kalshi’s rebate programmes do not account for how persistently some participants quote, which suggests those traders hold advantages in capital, information, or tolerance for risk that make the activity worthwhile without a subsidy.

Two patterns in the price of liquidity follow from the 0-1 payoff structure.

The first is that liquidity is far more expensive at the edges of the probability range. Spreads on contracts trading below $0.10 are roughly an order of magnitude wider than on contracts priced near $0.50 (Dubach 2026). A market maker who sells a contract at three cents collects those three cents and is exposed to losing ninety-seven if the event happens: about thirty-two times the risk of the reward. Quoting that side at all requires being compensated for the asymmetry, and the compensation is the width of the spread. This will matter in Section “Known biases”, because some of what is usually described as a behavioural bias toward longshots may simply be the price of liquidity where liquidity is most dangerous to supply.

The second is that liquidity disappears when it is most needed. On Polymarket’s five-minute Bitcoin contracts, shortly before settlement, depth in the prediction market falls while depth on the underlying spot exchange rises (Dai et al. 2026). The divergence is a perfect example of the effects of the continuous-versus-binary payoff difference. A late push that moves Bitcoin by a few basis points costs the spot market maker caught on the wrong side a few basis points, and part of that comes back when the price reverts moments later, as it usually does. The prediction-market maker’s contract does not care how far the price moved, only which side of the strike it ended on. The same few basis points, if they cross that line, turn her contract’s worth from $1 to $0, and the reversion afterwards is worth nothing to her, because the contract has already settled. One maker loses a proportionate amount temporarily; the other loses everything, permanently — therefore it is logical for market makers in the prediction market Bitcoin contracts to withdraw from liquidity provision.

Informed trading and adverse selection

The counterpart to liquidity provision is the risk of quoting against someone who knows more. In Glosten and Milgrom (1985) a market maker facing a mixture of informed and uninformed counterparties cannot distinguish them ex ante and must set a spread wide enough that gains from uninformed trades cover losses to informed ones; the spread is the price of that ignorance.

Event contracts make the problem worse in a specific way. An informed equity trader knows a share is worth a little more than the market thinks. An informed event-contract trader often knows something much starker — not that a contract is worth more than $0.30, but that it is going to settle at $1. And where the equity dealer expects to buy and sell her way back to a flat position, collecting the spread as she goes, the prediction-market maker frequently cannot: when informed traders are all pushing the same way, she is left holding the position until it resolves, and then loses the entire distance to $0 or $1 rather than a few cents (Bartlett and O’Hara 2026), which cannot be covered by normal spreads.

The theory therefore predicts that liquidity providers should abandon the markets where informed trading is worst. Interestingly, we observe the opposite. On Kalshi they quote those markets, and earn roughly twice as much per contract there as elsewhere.

Kalshi lists both single-name markets, which turn on a specific company or person, and broad-based markets, which turn on aggregates like inflation or an index level.

While the outcome of an election is not something about which one could hold insider information, somebody may well know whether a particular executive is about to resign, so private information is concentrated in the first group and dispersed in the second, while on every other structural aspect they are identical.

Comparing them isolates the effect of information concentration on its own. Single-name markets do show substantially more informed trading, and the price impact it produces is permanent rather than temporary: prices move and stay moved, which is the signature of information arriving rather than of inventory being worked off (Bartlett and O’Hara 2026). What keeps makers in those markets is a mistake on the other side of the book. Traders taking liquidity on Kalshi’s single-name markets pay too much for YES — not only on longshots, but across the whole probability range. A market maker there is right most of the time and wrong expensively, and it is the steady overpayment by optimistic traders that covers the expensive occasions. The arrangement is self-reinforcing in an uncomfortable way: the markets with the most concentrated private information are also the markets that attract the most optimistic uninformed money, so the surplus grows in step with the losses it has to absorb, and spreads never need to widen. Trading stays cheap in exactly the markets where it is most dangerous to make.

Whether a maker can see this coming in real time is a separate question. test it with a measure of how one-sided recent order flow has been — the intuition being that when trades start arriving overwhelmingly on one side, someone probably knows something. Adapting it to event contracts requires care, because a contract whose outcome is becoming obvious will mechanically attract one-sided flow without anyone being informed at all. Restricting attention to contracts still priced in the middle of the range, where that problem does not arise, one-sided flow does predict maker losses on single-name markets and does not on broad-based ones.

Evidence of informed trading turns up on Polymarket too, and in a particularly clean form. classify accounts by skill first and then look at what each group’s trading predicts. The market makers in their sample show a revealing pattern: their buying predicts where the price goes next, but predicts the opposite of where the contract eventually settles. They successfully accommodate the flow arriving at them in the short run, and are systematically on the losing side of it by the time the market resolves. They make money from the spread, not from being right, which is perfectly in line with what the theory suggests.

Price discovery and the limits of arbitrage

If these markets worked as the theory in Section “Event contracts as Arrow–Debreu securities” describes, two things would follow. Within a platform, the prices of a set of contracts covering every possible outcome would add up to the payout, because any gap is free money for whoever closes it. Across platforms, the same contract would cost the same everywhere.

The first condition holds loosely and can fail when the platform makes closing the gap unattractive. Fees on profits, fees on withdrawals, and limits on position size all reduce what an arbitrageur can earn and how large a correction they can make, and where those frictions have been severe the inconsistencies have persisted rather than being competed away (Stershic and Gujral 2020). The general point is worth holding onto: a price only equals a probability when someone is both able and motivated to correct it when it does not, so a market’s fee schedule and position limits are part of what determines whether its prices can be read as probabilities at all.

The second condition fails outright, and more interestingly. During the 2024 US presidential election, the same contracts traded at economically meaningful different prices on Polymarket, Kalshi, PredictIt and Robinhood, and the gaps did not close quickly; Polymarket tended to move first when trading was active (Ng et al. 2026). Ordinarily one would expect capital to flow toward the cheaper venue until the difference vanished. Here it could not, because the barrier is legal rather than economic: the access restrictions, collateral requirements and geographic blocks catalogued in Section “Regulatory overview” physically prevent most traders from being on both venues at once. A trader in Europe blocked from Polymarket cannot arbitrage it against Kalshi however obvious the discrepancy.

This is one of the more important points in the section, because it means regulation is not merely a constraint imposed on these markets from outside, but it directly affects price formation through limiting cross-platform arbitrage. It also has a practical consequence for anyone using the data: there is no such thing as the market-implied probability of an event, only Polymarket’s and Kalshi’s, which reflect different populations of traders under different rules.

A final case turns the usual question around. Prediction markets are normally asked whether they reveal information about the world. But where a contract settles on something its own holders can move, the market can degrade the information it is supposed to reflect. On Polymarket’s five-minute Bitcoin contract — which pays out according to the Bitcoin price at the close of a five-minute window — spot trading spikes in the final seconds before settlement, and the price move reverses within seconds afterwards (Dai et al. 2026). Genuine information would not reverse; a shove intended to drag the settlement price across the line does. The mechanism is a matter of horizon: over five minutes there is almost no time for ordinary price discovery, so a single well-timed push can decide the outcome, whereas over fifteen minutes it is diluted by everything else that happens. The signature is sharp in the five-minute contract and much weaker in the fifteen-minute one, which makes contract duration a design lever rather than a detail. The resolution problems this creates are taken up in Section “Resolution”.

Accuracy

Prediction market prices are good forecasts. Iowa Electronic Markets prices beat contemporaneous polls over decades, especially at long horizons (Berg et al. 2008); Polymarket prices track realised frequencies closely across more than 124 million trades (Reichenbach and Walther 2025); and state-level contracts appear to have had an edge over poll aggregation in the 2024 US presidential election, where the Electoral College makes correlated state-by-state probabilities the thing that actually matters (Cutting et al. 2025).

The interesting question is why, and this wiki has already given two incompatible answers.

Section “The wisdom of crowds” gave the Galtonian one. A crowd guessing the weight of an ox is collectively accurate because individual errors are independent and cancel out; a prediction market aggregates beliefs the same way, and its price is accurate because it averages away the noise in any individual view. On this account accuracy is a property of the crowd, and more participants make it better.

Section “Prices as information aggregators” gave a very different one. If the price already told you everything the informed traders know, reading it would be free and learning anything would be a waste of money — so nobody would bother, and the price would stop being informative. A market that aggregates information must therefore leave the people who gather it enough profit to make gathering it worthwhile. That profit has to come from somewhere, and the only available source is the other traders. Following this logic, accuracy is not a property of the crowd, nor a product of information accumulation, but rather is produced by a minority, and uninformed traders pay for it.

These predict different things about who moves prices, which makes them testable, and the evidence clearly supports the second framing.

This could already be inferred from the IEM’s 2020 presidential market, where Gruca and Rietz (2023) found the usual behavioural mistakes among traders (selling winners too early, holding losers too long, overvaluing what they already owned) but showed that the orders which actually set prices came from a much more disciplined subgroup. Market efficiency came from a small subgroup, while most of the participants were not trading efficiently. found the modern version of this on Polymarket: only about a third of accounts make money, and most importantly, the same accounts keep making it. Under the crowd account there is no reason for winners to repeat, since everyone is equally informed and outcomes are noise. Persistent winners mean somebody has an edge.

settled the question by finding a way to tell skill from luck. With over a million accounts, plenty of traders will profit by chance alone, so realised profit proves nothing. Their test asks what each account would have earned had it made the same trades in random directions, and treats an account as skilled only if its actual profit is extreme against that benchmark. On Polymarket, roughly 3% of accounts qualify. Everyone else divides into traders who profited without demonstrable skill, traders who lost, and a small group who lost more consistently than chance would explain. Together these make up about 97% of accounts and supply about 70% of the volume.

The most important finding for determining where accuracy comes from is that only the skilled minority’s trading actually moves prices toward the correct answer — their buying predicts both the next price move and the final outcome, and the more of the volume they account for, the closer the price gets to the truth. Everyone else’s trading either predicts nothing or points the wrong way.

While one could suspect that this minority are insiders, this is not the case. Accounts showing the signature of inside information are few, and account for a fraction of a percent of volume; almost none of them overlap with the skilled group (Gómez-Cram et al. 2026). What the skilled actually do is they react to scheduled news faster than everyone else, they take the free money when prices on related contracts stop adding up, and they trade against the crowd’s predictable mistakes. Insider trading proper is treated in Section “Insider trading”.

So prediction market accuracy is real, but it does not work the way the wisdom-of-crowds story suggests. Prices are accurate because a small number of skilled traders are paid, out of the losses of a large number of unskilled ones, to make them accurate. The crowd’s role is not to supply the information, but to supply the money that makes gathering information worthwhile, and the trading noise that lets the informed profit without giving themselves away. That is a more fragile arrangement than it first appears, because it needs both groups: a market that lost its skilled traders would lose its accuracy, and a market that lost its crowd would lose the traders.

Known biases

A price can be informative and still be predictably wrong. The best-known example is the favourite–longshot bias, first documented at racetracks (Thaler and Ziemba 1988): unlikely outcomes are priced too high and likely ones too low, so that contracts trading at five cents win less often than one time in twenty, and contracts at ninety-five cents win more often than nineteen times in twenty.

The pattern appears in prediction markets. Why the bias exists is genuinely unsettled, and the competing explanations are worth separating because they are not the same kind of claim.

The oldest is about preferences. Traders enjoy the small chance of a large payoff, or systematically overestimate how likely rare events are, and so pay more than a longshot is worth.

The second is about costs. On Kalshi, much of the measured bias comes from fees and spreads rather than from mispriced beliefs: traders who take liquidity pay both, lose considerably more than those who supply it, and the gap is widest on cheap contracts (Bürgi et al. 2026). A trader here can hold perfectly accurate beliefs and still show a losing record on longshots.

The third is about liquidity, and follows from Section “Liquidity provision and the maker–taker structure”. Spreads on low-priced contracts are roughly an order of magnitude wider than in the middle of the range (Dubach 2026). Someone buying a longshot pays that spread, and the shortfall in their realised returns would look exactly like a favourite–longshot bias even if their beliefs were flawless and they had no taste for gambling at all.

These are not mutually exclusive, and sometimes empirical findings contradict each other even when done on the same platform.

Moreover, the most detailed study of Kalshi finds a miscalibration that is not the favourite–longshot bias in any recognisable form: traders overpay for YES across the entire probability range, favourites included, and only in the single-name markets, while broad-based markets are priced about right (Bartlett and O’Hara 2026). That is a different shape of error with a different role, since as Section “Informed trading and adverse selection” described, this particular overpayment is what funds market making where informed trading is worst.

A last kind of bias has nothing to do with the price level. A contract’s price is not equally trustworthy over its life. Early on, when little has traded and little is known, prices are unreliable simply for lack of information; late on, they can be distorted by traders positioning for settlement (Reichenbach and Walther 2025), which is consistent both with how much volume arrives in a market’s final hours and with the settlement-window manipulation described in Dai et al. (2026). For anyone using these prices as data, the practical rule is that a liquid contract’s mid-life price is the most defensible number to use, and the prices on either side of it should be treated with caution.

Applications

Prediction markets are used by two distinct groups of people (besides, of course, the traders themselves). Researchers treat them as data: an input to a question about how voter preferences move together across states, the inflation expectations of the market, or how political risk is priced into equities. In the other group, firms and individuals treat them as instruments: something to hold in order to change an exposure they already have.

The two uses make different demands of the same number. The research use needs the price to be an accurate and, ideally, unbiased estimate of a probability; if it is distorted, whatever is estimated from it is distorted too. The hedging use needs almost none of that. It needs clear and predictable resolution, and the market to be deep enough to absorb a position of the required size. A contract can be an entirely serviceable hedge while being a poor probability estimate — indeed, a hedger who believes the price is too high is simply someone who thinks insurance is expensive, which does not stop them buying it. The converse also holds: a well-calibrated contract with insufficient liquidity is useless to anyone with real exposure.

Applications in research

Four properties make prediction market data attractive as an empirical input, and it is worth stating each together with the condition it depends on, because Section “Market Microstructure” has already established that none of them holds unconditionally.

1. The beliefs are financially backed. A participant who is wrong loses money, so there is an incentive to form an accurate view rather than to signal group membership or give the socially comfortable answer (the failure modes that bias survey data). Whether this incentive actually holds from traders can be inferred from trading behaviour: relying on Gómez-Cram et al. (2026), there is only a small group of participants who lose more consistently than what chance would explain and do so persistently. While the lack of skill is compatible with following financial incentives, the persistence of it is less so. It suggests that these accounts either derive utility from betting itself, or bet based on their own political convictions.
2. Prices update continuously. Surveys are conducted at fixed intervals and may lag an event by days; a market reprices within seconds of a data release. The condition is liquidity. Section “Liquidity provision and the maker–taker structure” showed that depth is thin at the edges of the probability range and can evaporate entirely near settlement, so “continuously updated” describes a liquid contract in mid-life and not much else.
3. The contract prices the event directly. Extracting an event probability from options requires differentiating a call price with respect to the strike, as in Section “Event contracts as Arrow–Debreu securities”, and in practice requires a strike grid fine enough for the second difference to be stable. A binary contract is the digital claim, so no transformation is needed. The condition is the one Section “Event contracts as Arrow–Debreu securities” closed on: what is recovered is the risk-neutral probability $\mathbb{\mathbb{Q}}$, not the physical one. This is returned to below, because for macroeconomic contracts the gap is not a technicality.
4. On-chain records are individual-level. Polymarket’s settlement layer makes every trade traceable to a wallet, which gives researchers an unparalleled panel of individual position-taking. The limitation is that this exists on one platform only, and that its population is not a random sample — both Polymarket and Kalshi users can be expected to be tech-native, possessing the necessary skills to navigate the platforms and, in the case of Polymarket, the blockchain infrastructure. These biases have to be accounted for when researchers wish to extrapolate from prediction market data to a more general population.

Academic application examples

Election forecasting. State-level contracts covering the whole Electoral College map outperformed poll aggregation in 2024, on a single cycle (Cutting et al. 2025). A more sophisticated application is the following: prediction market prices are used as an additional measurement equation in a state-space model of latent voter preferences, alongside polls and fundamentals. They identify a two-factor comovement structure across states, and ignoring that correlation biases estimated win probabilities by more than ten percentage points, with Pennsylvania pivotal on both counts (Chernov et al. 2025). The template generalises beyond elections, because a model that treats the price as a noisy signal can absorb the frictions of Section “Market Microstructure” instead of being defeated by them. Its binding assumption is that the measurement error is uncorrelated with the latent state, and the biases of Section “Known biases” are price-level dependent, hence correlated with the very quantity being recovered.

Macroeconomic expectations. Kalshi lists contracts on CPI, core and PCE inflation, unemployment, payrolls, GDP growth, recession and the Federal Reserve target rate, which respond sharply and plausibly to central bank communications and employment releases (Diercks et al. 2026). Against surveys the gain is frequency; against derivatives it is three-fold: no model-dependent transformation is needed to obtain a probability, coverage extends to variables with no liquid derivative (recession being the clearest), and because contracts are written on outcome bins the prices trace a distribution rather than a point, so uncertainty is measured rather than assumed. The risk-neutral caveat bites hardest here. A recession is close to the definition of a high-marginal-utility state, so a claim paying $1 in it should trade above its physical probability for the same reason a put trades above its expected payoff. Read as an expectation, a recession contract price is biased upward by a risk premium — not noise, since it does not average out and it moves with risk appetite.

Policy beliefs. Wallet-level data on the 2025 debate over Federal Reserve independence show that users assigning higher probability to political pressure on the Fed simultaneously expect lower short rates, higher long-term yields and higher inflation, which is in line with what the credibility mechanism would suggest (Eichengreen et al. 2025). The test is only available at the individual level: aggregate prices would show that the market expects both things, not that the same people do, which is what distinguishes a credibility mechanism from two unrelated views held by disjoint groups.

Asset pricing. Changes in political probabilities can enter an augmented Fama–French specification as a factor, on the identifying assumption that a binary outcome has an approximately stable and locally linear effect on a firm’s conditional expected value (Auld 2022). Across six events the model finds support in four, with Mexican equities loading heavily on the probability of a Trump victory in 2016 and Brexit sensitivity varying with domestic revenue exposure. The null on the 2017 UK general election is the informative result: an election with no cross-sectional implications should produce no loading, and a specification that manufactured one would be capturing generic election-period volatility. The advantage over a conventional event study is twofold. An event study takes the outcome date as the event and therefore captures only the final surprise, discarding a revision that ran for months; and a change in probability is a regressor with a meaningful scale, so the loading reads directly as the difference in a firm’s value between outcomes, which a dummy variable cannot deliver.

Portfolio construction. Candidate baskets are built by identifying windows in which odds moved materially toward one candidate and collecting the assets whose returns rose in them (Jain et al. 2025). The reported outperformance of the Trump-linked basket should not be read as evidence for the method: assets are selected on the same returns against which they are then evaluated, over one electoral cycle, so the winner’s basket outperforms close to mechanically, and what the exercise produces is a conditional exposure rather than an alpha. Read instead as a hedging instrument the construction is sound, and it is the synthetic counterpart of the direct hedges in Section “Applications in practice” — a portfolio loading on a political factor is what one shorts to remove that exposure, and a hedger cares most about the correlation.

Belief measurement. Trade records can be used as revealed-preference survey data rather than as forecasts. A Political Betting Leaning Score built from bet direction, price, position size, timing, party alignment and frequency recovers a user’s political orientation. A case study of the 2022 US Senate markets finds supporters of the trailing candidate positioning on belief while the leader’s supporters balance belief against price (Chen et al. 2025). The appeal is that a survey records what someone says and a market records what they will pay. The constraint is selection, in two stages: participation requires the digital literacy to fund and operate an online trading account, and Polymarket additionally requires crypto-nativity. The problem is not that the sample is unrepresentative (almost every sample is, and researchers reweigh) but that the variable selecting into it is not plausibly independent of the variable being measured. Reweighing works when sampling probability, conditional on the covariates, is unrelated to the answer; here the entry requirement is holding cryptocurrency, which there is no reason to treat as politically neutral. If it is not, a partisan score from this population is biased rather than merely noisy, and the bias cannot be removed without knowing a correlation the design cannot observe.

Using prediction market prices responsibly

The papers above collectively imply a short set of checks that anyone using a prediction market price as data should be able to answer.

1. Which platform, and on what grounds? document economically meaningful and persistent price differences across Polymarket, Kalshi, PredictIt and Robinhood on identical 2024 contracts. As Section “Price discovery and the limits of arbitrage” argued, the segmentation is legal rather than economic, therefore it is likely to persist and cannot be closed by markets becoming more efficient. There is consequently no such thing as the unique market-implied probability of an event, and a choice of venue is a choice of trader population and rulebook that needs to be explicitly stated.
2. Can arbitrage operate? show that PredictIt’s fee on profits, fee on withdrawals, position limits and participation limits together left linked contracts violating no-arbitrage conditions for extended periods, while equivalent deviations on the IEM were smaller and shorter-lived. A price equals a probability only when someone is both able and motivated to correct it when it does not, so the fee schedule and the position limits should be taken into consideration before interpreting prices as probabilities.
3. Where in the contract’s life is the price read? locate inaccuracy at both ends: early, when thin trading and sparse information make prices unreliable, and late, when positioning for settlement distorts them. The defensible observation is a liquid contract’s mid-life price, as Section “Known biases” concluded.
4. Is it $\mathbb{\mathbb{Q}}$ or $\mathbb{\mathbb{P}}$ that the application needs? For a forecasting exercise the distinction is often tolerable. For anything that treats the number as a belief — an expectation, a sentiment measure, a subjective probability — the risk tilt is a first-order problem, and most acute for macroeconomic contracts.
5. Does the estimand require knowing who initiated the trade? This is the question deferred from Section “Measurement: what the data show”. Kalshi records, for every trade, which side crossed the spread. Polymarket does not, and inferring it from the public order-book feed is correct only around 59% of the time, against roughly 81% for the equivalent procedure on equities (Dubach 2026). Anything whose sign depends on trade direction — effective spreads, market-maker profit and loss, order-flow imbalance measures — is therefore unreliable on Polymarket, which is why the effective-spread and maker profit-and-loss estimates in Bartlett and O’Hara (2026) rest on Kalshi data. Estimates that need only prices, positions or wallet identity are unaffected, which is why the trader-level work of Gómez-Cram et al. (2026) and Chen et al. (2025) can use Polymarket data without inheriting the problem. The platform choice is thus partly determined by the estimand and not by convenience or data availability.
6. Is the price exogenous to what it is being used to measure? argues that a publicly visible probability can function as a coordination device: it may move donors, journalists, investors and political actors, and thereby the outcome itself. A price can be well calibrated in expectation and still be endogenous to the event it prices. For a forecasting exercise this is harmless. For any causal claim — and in particular for any design that uses a probability change as a shock or an instrument — it is not, and the design needs an argument that the feedback channel is closed.

Applications in practice

The opening of this section drew a distinction that now does some work. A hedger does not need the price to be right. What a hedger needs is a payoff correlated with an exposure they already hold, in a size that matters, over a horizon that matches the risk, and with basis risk small enough to leave the hedge worth having.

Those requirements are the requirements of insurance, and that is the most useful way to understand what is happening here. The practical use of event contracts is not a new kind of speculation with a commercial justification attached; it is parametric risk transfer, conducted through a derivatives exchange instead of an insurer. This subsection therefore leads with the mechanism and uses the reported cases as illustrations, because the cases are individually small and collectively few.

Corporate and internal prediction markets

Internal corporate markets are not treated at length here. For practical and confidentiality reasons they are thin, low in volume and subject to high price impact, which limits both what they can be used for and what can be learned from them.

The mechanism: parametric cover without insurable interest

An insurance contract in the classical sense indemnifies: the insurer establishes what the policyholder actually lost and pays that amount, which requires a claims process, an adjuster, and proof of loss. A parametric contract does something different. It pays a fixed sum when a defined trigger is observed in a specified data source, regardless of what the holder lost. Catastrophe bonds and weather derivatives work this way, and so, exactly, does a binary event contract. The insurance trade press treats Kalshi’s weather markets as parametric protection without qualification (Artemis 2026). The appeal of this solution is that a complete set of state-contingent claims spans risks that instruments defined on continuous underlyings cannot reach. Conventional derivatives reference prices; insurance references losses; an event contract references a fact. That lets it cover exposures for which no insurance product exists and no underlying can be traded. Because the trigger is a public fact, settlement needs no adjuster and no proof of loss, and happens at once (if resolution mechanisms function well).

Two things are given up in exchange, and they are what the rest of this subsection is about.

The first is indemnity, and its absence is basis risk. A parametric contract pays on the trigger, not on the loss, so the holder is exposed to every way the two can diverge. The second is insurable interest. English insurance law has required since the eighteenth century that a policyholder stand to lose from the event insured, a doctrine adopted precisely to stop insurance contracts being used as wagers on other people’s misfortunes. Neither the Commodity Exchange Act nor a designated contract market’s rulebook imposes any such requirement: a derivative may be bought by anyone (Semafor 2026).

Its absence is what makes the instrument useful. A firm that cannot buy cover because no insurer will underwrite its exposure can buy a contract instead, and can do so without demonstrating anything to anybody. It is also why the counterparty exists at all: the person selling the cover is not an insurer with actuarial models but a trader taking a view, and Section “Accuracy” explains why they are willing to be there.

Its absence also removes the safeguard the doctrine was built to provide. If a hedge is distinguishable from a wager only by the buyer’s exposure, and nothing in the contract records the buyer’s exposure, then the distinction is not enforceable at the level of the instrument. That matters most where the holder can influence the trigger — the configuration insurance law spent two centuries legislating against, and the same configuration that produces the settlement manipulation documented in Section “Price discovery and the limits of arbitrage” and the resolution problems of Section “Resolution”. It also complicates the regulatory argument in Section “United States”, since the hedging use case is the strongest evidence that these are risk-management instruments rather than gambling, and the absence of an insurable-interest requirement is what makes that evidence unverifiable contract by contract.

Three layers at which cover is supplied

The reported cases fall into three structurally distinct arrangements.

Direct. The exposed party buys contracts on the exchange itself. Kalshi’s own showcase of small-business hedges is entirely of this kind: a Manhattan bar buying $5,000 of contracts on a Knicks win to cover a free-tab promotion it had already announced (Kalshi 2026b; Moneywise 2026); a furniture retailer covering a refund promise contingent on a television result; an ice cream shop in Los Angeles taking roughly $20 a day of weather contracts (Kalshi 2026a). The exposure is genuine and the basis risk is often near zero, because the promotion and the contract are written on the identical event. There is no intermediation and no bespoke structuring; the firm is a retail customer.

Brokered and bespoke. A broker identifies an exposure, persuades the exchange to list a contract designed around it, and finds a dealer to take the other side. Western Grazers, a Californian contract-grazing operator facing a statutory change that raised the applicable wage for a goat herder to roughly $240,000 a year (Ag Alert 2026; Agri-Pulse 2026), is reported to have paid a $50,000 premium for a position paying up to $500,000 if California failed to authorise relief before 1 October 2026. Castle Technologies structured it, Susquehanna priced it and took the risk, and Kalshi listed and self-certified the contract to the CFTC (Business Insurance 2026; Quintenz 2026; KalshiEX LLC 2026).

Reinsurance of a conventional policy. The exposed party buys ordinary insurance from a licensed insurer, and the prediction market supplies capacity behind it. The Spanish football club Osasuna took relegation cover through the broker Howden; when the story surfaced, the club stated that it had never traded on Kalshi and had no relationship with the platform, and Kalshi’s account was that the broker had bought the position to reinsure the risk because the exchange priced it better than conventional reinsurers (Fortune 2026; Semafor 2026).

This third layer is the most consequential and the least discussed. It means a firm can be hedged through a prediction market without ever touching one, that the demand reaching these markets need not come from the exposed party at all, and that event contracts may end up competing with reinsurance rather than with retail speculation. It also relocates the insurable-interest question: the club held the interest, the broker held the contract, and the doctrine has nothing to say about a chain arranged that way. reports the industry taking the competitive threat seriously while doubting whether crowd pricing can substitute for actuarial modelling, and notes a further obstacle — regulators grant capital credit for cover placed with certified reinsurers, and it is not clear that a position on a derivatives exchange qualifies.

A caution on this case in particular should be mentioned, namely that the reported figures from different news sources do not agree. However, the underlying mechanism is interesting to consider.

Challenges

The preceding sections have mostly asked whether these prices are accurate, while this one asks what can go wrong that accuracy does not fix. The three problems below share a structure: each concerns something the market cannot correct through trading, because the thing at fault sits outside the mechanism by which counterparties discipline one another.

Resolution

Every event contract needs a rule mapping the world onto $\{0,1\}$. Section “Resolution and settlement” described the two ways platforms supply that rule: centrally, as Kalshi does, or by the UMA oracle, as Polymarket does. The details matter more than they appear to, because a contract is only as good as its map — a platform which has unreliable and disputed resolutions would have a hard time attracting and keeping customers.

There is an important difference between trading manipulation and resolution manipulation. As Abid (2026) discusses, manipulation at the trading layer is self-limiting, because a large directional bet that distorts a price creates a profit opportunity for anyone who thinks the price is wrong, and if such traders have capital the distortion is bounded in size and duration. Manipulation at the resolution layer has no such corrective. Nobody can take the other side of a settlement. This is why aggregate calibration and severe governance vulnerability can coexist on the same platform without contradiction.

The problems concerning the resolution process can be sorted into three main categories:

1. the underlying — the state of the world the contract refers to;
2. the measurement — the instrument or data source that observes it;
3. the adjudication — the process that decides what the measurement showed and what the contract’s words mean.

The first was treated in Section “Price discovery and the limits of arbitrage”, where the settlement-window manipulation of Polymarket’s five-minute Bitcoin contracts is documented (Dai et al. 2026). The other two are the subject of what follows.

Centralised resolution and its failure mode

Kalshi resolves centrally: the exchange’s own staff determine the outcome, and the contract’s rulebook specifies the sources. The failure mode is not capture but drafting. The most important consequence is that there is no possibility for appeal: a centrally resolved contract offers the holder no process at all, which is a different complaint from the one usually made about decentralised resolution but not obviously a smaller one.

The UMA oracle

Polymarket resolves through UMA’s optimistic oracle (UMA Protocol 2026). An outcome is proposed by anyone willing to post a bond, reported in the March 2025 case as $750 in USDC.e (CoinDesk 2025a). If nobody disputes the proposal within a window, it stands and the market settles. This is the optimistic part, and it is the normal path: most markets never reach a vote.

If the proposal is disputed, the question goes to UMA’s Data Verification Mechanism. Voting power is held by staked UMA tokens. A voting period lasts 48 hours and is split in two: during the first 24 hours voters encrypt and commit their votes, and during the second they decrypt and reveal them, with the tally taken by contract at the end of the reveal phase. Two thresholds must be met for a vote to resolve: a floor on total participation (the GAT, a fixed five million UMA) and an agreement threshold (the SPAT, set at 65% agreement among staked tokens); votes failing these roll into subsequent rounds, and become deletable after rolling four times. Stakers earn emissions regardless of how they vote, but a voter who votes against the eventual outcome, or fails to vote at all, is slashed at roughly 0.1% of stake, and the slashed tokens are redistributed to the voters on the winning side (UMA Protocol 2026).

The commit phase matters more than it first appears, and it is worth crediting the design for it. Because committed votes are encrypted, no voter can observe how others are voting while the vote is open. UMA states the purpose plainly: the scheme exists so that votes remain private, protecting voters from being influenced by how others are voting during the polling period. Whatever else is wrong with token-weighted adjudication, the most obvious failure — a visible herd that everyone piles into — is designed out.

What the slashing rule does and does not imply. The rule rewards matching the eventual majority rather than being right, and the documentation is candid about this, describing the mechanism as a commit–reveal Schelling point. It is tempting to conclude that a large holder can therefore induce others to follow, but commit and reveal blocks that: there is nothing to follow, because there is nothing to see. The objection raised by a market participant during the July 2025 dispute — that UMA’s incentives encourage voting with the perceived majority to avoid penalties rather than on factual correctness (CoinDesk 2025b) — is therefore a claim about expectations, not about observation, and it stands or falls on whether a voter can form a view about the majority independently of the facts.

Two situations remain where they can, and they are different problems.

Outright capture. If a single holder controls enough staked UMA to satisfy the agreement threshold alone, no coordination is required and secrecy is irrelevant: the vote is decided by arithmetic. This needs no behavioural story at all, and it is the simplest reading of the March 2025 episode below. The economics are unfavourable in a specific way. Accumulating gradually limits price impact, the tokens are not consumed by the attack, and the attacker, being the majority, is never slashed and is paid out of the stake of anyone who voted against. The only real constraint is that a visible attack damages the value of the asset used to conduct it, which is presumably why the observed case involved a $7 million market and not a much larger one.

Coordination without a focal point. The second situation is the one that needs no attacker, and it is where the Schelling framing actually bites. When the underlying fact is clear, truth is the obvious thing for everybody to converge on, and the mechanism works. When the fact is genuinely ambiguous — which is, after all, the condition under which a dispute arises — there is no truth-shaped focal point to converge on, and a voter’s payoff still depends on matching others. They are then asked not what happened but what a majority of strangers will conclude happened, and the answer turns on whatever reading is most salient rather than most correct. Salience can come from the resolution text, from press coverage, or from a widely observed early signal. Section “When nobody cheats: the problem of definitions” gives a case in which it came from a clause in the contract, and the outcome contradicted forty newspapers.

The distinction is worth holding onto because the two failures have different remedies. Capture is a problem about stake distribution, while coordination failure under ambiguity is a problem about contract drafting and would persist under any voting rule whatever, including a perfectly honest one.

A good example of the first is the March 2025 case. The mechanism was tested and failed. The market asked whether Ukraine would agree to a minerals deal with the United States before April 2025; roughly $7 million was at stake. Between 24 and 25 March the YES price moved from about 9% to 100% and the market resolved YES, although no agreement had been reached — Polymarket itself acknowledged both that the two countries had not come to a mutual agreement and that the market was resolved too soon (The Block 2025; CoinDesk 2025a). The platform described it as an unprecedented governance attack by a large token holder, said it had spent the day in war rooms with the UMA team, declined to compensate holders on the ground that it was not a market failure, and announced no specific remedy. Note that on the reading above this episode requires no herding and no coordination: a holder with sufficient stake and secret ballots wins anyway.

When nobody cheats: the problem of definitions

In 2025 Polymarket listed a market on whether Volodymyr Zelenskyy would be “photographed or videotaped wearing a suit” between 22 March and 30 June. Reported volume was very large: $200 million on one account, $237 million on another (CoinDesk 2025b; Decrypt 2025). On 24 June Zelenskyy appeared at the NATO summit in The Hague in a black jacket and matching trousers. More than forty international outlets described the garment as a suit. An initial YES resolution was disputed and overturned; on 1 July the oracle ruled NO, on the stated ground that there was insufficient “credible reporting consensus”; the outcome was confirmed on 9 July (Decrypt 2025).

The episode exposes two structural issues.

The first is that the contract was incomplete, as every contract on an open future must be. “Suit” has no operational definition, the resolution text supplied none adequate to the realisation that occurred, and no amount of care in drafting eliminates this — it displaces it.

The second is what the record shows afterwards. The market resolved NO. It does not resolve “NO, contested, on a reading of ‘suit’ that forty newspapers rejected.” call this last step epistemic hardening, and it is the point at which the resolution subsection hands over to Section “Epistemic critique”: a disputed adjudication becomes a settled binary fact, and the contestation is not merely forgotten but structurally unrecordable, because the data format does not allow for it.

Attacking the instrument

The third attack surface is the one the literature has not caught up with, and the clearest case is recent.

Polymarket listed contracts on the daily high temperature in Paris, settling on the Météo-France automatic weather station at Charles de Gaulle airport. On the evening of 6 April 2026 the station’s reading rose about 4 °C without meteorological cause. On 15 April it rose 5 °C, from 16.9 to 21.9, in twelve minutes, accompanied by an abrupt fall in humidity, a combination local meteorologists ruled out as natural. An account called xX25Xx turned a $119 stake into $21,398 on the 15 April contract, a position roughly twenty times its typical size, and was subsequently deleted. Around $1.4 million was staked across the two affected contracts, more than double the usual volume for Paris temperature markets. Météo-France filed a complaint with the airport police for “tampering with the operation of an automated data processing system”; observers speculated about a lighter or a battery-powered hair dryer (NPR 2026; Insurance Journal 2026). No tampering has been proven and the investigation was open at the time of writing.

Three things make this more important than its size suggests.

The attack is on the measurement, which is neither of the two surfaces the existing literature models. The underlying was not touched: the weather in Paris was whatever it was. The adjudication was not touched either. A resolution process that is functioning correctly offers no defence against this at all, and may even entrench it, since the reading is exactly the kind of authoritative third-party datum a well-designed contract is supposed to defer to.

The cost asymmetry is extreme and it is the general lesson. The apparatus required was, on the leading speculation, a household appliance. The payoff on one position was two hundred times the stake. The security of a parametric contract is inherited from its data source, and public measurement infrastructure was never built to be adversarially robust, because until recently nothing turned financially on any individual reading. Listing a contract on a public instrument converts it into a financial one and gives it an attack surface, without the operator agreeing to that, being compensated for it, or necessarily knowing.

The externality is not confined to the market. Charles de Gaulle is a working airport and its weather data is used operationally; a meteorologist quoted in the reporting made the obvious point that corrupted data there is dangerous irrespective of anyone’s bet (Insurance Journal 2026). The harm from a manipulated settlement is normally borne by the counterparties, who chose to be there. Here part of it lands on people who have never heard of the market.

This case is worth setting against the previous one. The definitional failure is loud and produces a public dispute; the measurement failure is silent and produces a clean settlement on a false number. It is not obvious which is worse.

What would actually help

The honest summary is that none of the available remedies is free, and that the platforms’ responses so far have been ad hoc: a station swap in one case, a refusal to compensate and no rule change in another.

Redundancy in the data source — settling on a median across several stations rather than one — raises the cost of the sensor attack by roughly the number of instruments, and is the standard answer in parametric insurance. It does nothing about adjudication. Larger proposal bonds and longer dispute windows raise the cost of a frivolous proposal but also delay every honest settlement, and do not touch an attacker who expects to win the vote. Tying oracle security to the notional secured is the fix that addresses the structural problem identified above, and no platform does it. Centralising resolution removes the capture risk and replaces it with an unappealable decision by an interested party.

Epistemic critique

The objections in this subsection are not that prediction market prices are inaccurate. Section “Accuracy” established that they are good forecasts, and nothing here disputes it. The objections concern what a publicly visible probability is, and what it does once it is public.

The number is not what it is read as

The foundational result is Manski (2006), and it is prior to everything empirical in this wiki. When traders hold heterogeneous beliefs, the equilibrium price is not the mean belief of the participants. It is determined by the marginal trader and by the distribution of stakes, so the mapping from a price to any interpretable summary of what participants think requires assumptions the market does not supply. Section “Event contracts as Arrow–Debreu securities” reached the same place from the other direction: the price is a risk-neutral probability, equal to a physical one only under risk neutrality and homogeneous beliefs.

So a perfectly liquid, perfectly honest, entirely unmanipulated market still does not deliver the object that a headline reading “the market gives this 63%” claims for it. This is worth stating first because the remaining criticisms are often heard as complaints about execution. They are not. The gap is in the definition.

How a bet becomes a fact

give the most developed account of how a price acquires its authority, based on digital ethnography, interface walkthroughs and interviews with participants. They call the process prediction laundering and identify four stages, which between them cover most of what the preceding sections of this wiki have established empirically.

Structural sanitisation is the market-creation step: which futures become bettable at all is decided by whoever proposes and approves contracts, and that filter is not neutral. Probabilistic flattening is the collapse of heterogeneous motives — hedging, speculation, political expression, arbitrage, informed forecasting — into a single number that discloses none of them. This is Manski’s point restated as a design critique. Architectural masking is the presentation of a capital-weighted outcome through a front end that reads as crowd consensus. Epistemic hardening is the conversion of a contested resolution into a settled binary record, which Section “When nobody cheats: the problem of definitions” has just illustrated.

Two findings from elsewhere in the wiki supply the quantities their argument needs. show that about 3% of accounts generate the informational content, while the crowd supplies volume; Akey et al. (2026) find that 1% of accounts capture 76.5% of all gains. Whatever the interface conveys, the object behind it is not a consensus of many.

The sharpest illustration is the platforms’ own language. Kalshi describes its markets as leveraging the wisdom of crowds; Polymarket attributes their accuracy to collective intelligence; the chief executive of Robinhood has called them the application of capitalism to the pursuit of truth (Gómez-Cram et al. 2026). The empirical finding in the same paper is that the crowd contributes the money and a small minority contributes the information. The marketing is not merely imprecise, it names the wrong mechanism, and it names the one that sounds democratic.

Reflexivity

raises a different objection: that a visible probability can function as a coordination device and change the event it prices. Donors, journalists, institutional investors and political actors may all condition on a displayed number, and a price can be well calibrated in expectation while still being endogenous to its own subject. He proposes a Signal Credibility Index — combining variance-ratio evidence on price paths, a measure of two-sidedness in order flow, and a discount for concentration — intended to identify when a price is likely to acquire behavioural traction, as distinct from when it is merely accurate.

The concrete case is no longer hypothetical. In April 2026 Fox announced that it would integrate Kalshi forecasts across Fox News Media and its streaming platform (Business Wire 2026). A probability generated by a few thousand active traders, most of whom are unskilled, is thereby broadcast to a mass audience as a news product. Whether that changes outcomes is an empirical question; that it changes the number’s social function is not.

The counter-argument

The critique above is strong but it is not decisive, and a fair treatment has to say why.

The alternative to a market-implied probability is rarely no probability. It is a poll, a pundit, or an editor’s judgement about which possibility to lead with. Each of these has its own aggregation problem, its own selection into the sample, and its own undisclosed motives, and in general none of them is scored. A prediction market price is at least falsifiable after the fact and is in fact routinely scored, which is how Section “Accuracy” could be written. Manski’s objection applies with equal force to a poll aggregate, which is also not the mean belief of anything in particular.

The strongest version of the critique is therefore not that these numbers are bad, but that their form invites a reading their construction does not support: a single number, to two significant figures, updating in real time, with no attached statement of who produced it or how much it moved on how little. make the narrow version of this point in passing — a price is a level forecast and says nothing about how likely it is to change, so a stable 60% and a 60% immediately before a scheduled announcement are different claims presented identically.

Insider trading

Section “Informed trading and adverse selection” described why private information is more dangerous in an event contract than in a share: an informed equity trader knows a stock is worth a little more than the market thinks, whereas an informed event-contract trader may simply know the answer. This subsection asks how common that is, who does it, and — the question that makes it a challenge rather than a microstructure finding — whether anything prohibits it.

How much of it there is

Identifying insiders in pseudonymous data requires proxies, and Gómez-Cram et al. (2026) use three independent ones on the full Polymarket record. The record proxy flags accounts with a per-share win rate above 90%, profit above ten cents per share, at least $1,000 in total profit, and at least 90% of those profits earned in markets where inside information is possible. The timing proxy flags an account opened within seven days of an event, trading that event alone, and falling silent once it resolves. The conviction proxy flags traders who move a price by twenty cents or more in a market they go on to win, and earn most of their profits that way.

Together the three flag 967 accounts. Those accounts appear in 6.0% of markets and are active in 7.5% of time bins, but account for 1.3% of profits and only 0.2% of volume. Where they do trade their effect is large — larger, in fact, than that of the skilled traders who drive prices generally, with an order-imbalance coefficient on the final outcome of 22.54 against 9.82 for the skilled, and a reduction in pricing error of 8.79 basis points against 2.99. The picture is of something decisive locally and negligible in aggregate.

One number does more work than the rest. Of the 967 flagged accounts, eleven are also classified as skilled. Insiders and the skilled minority are very nearly disjoint populations, and it is this that licenses the conclusion reached in Section “Accuracy”: prediction market accuracy rests on traders who process public information quickly, not on traders who already know.

Where it is possible

Not every contract can be traded on privately, and the distinction is sharper than one might expect. classify all 210,322 markets in their sample by whether the outcome could in principle be controlled or privately known by some individual. The result splits the platform almost in two. Culture (77.5% insider-prone by their classification), Politics (72.3%) and Tech (71.6%) are exposed; Economy (44.1%) and Finance (25.7%) partly so; Sports and Crypto, which together dominate the market count, are almost entirely not (1.2% each). Roughly 89% of markets fall in the clean category, which is why insider trading can be locally severe without contaminating the platform’s aggregate calibration.

The exposed categories are exposed for reasons that have no analogue in equity markets. A television result may already be recorded and known to a production company. A market on which person was most searched on a given platform is transparent to that platform’s own staff. A contract on a resignation, a pardon or a military operation is knowable to whoever will make the decision. Corporate earnings markets are insider-prone by construction, since the firm’s own staff and its auditors see the figures before release. The party with the advantage is frequently not a corporate insider in any sense a securities regime contemplates.

A case, and an enforcement action

On 3 January 2026 a United States military operation captured the Venezuelan president. In the days and hours beforehand, three Polymarket accounts took positions of roughly 80,000 shares in a contract on his removal, then trading near ten cents, and realised more than $620,000 between them. All three had been created between 27 December and 3 January; two never traded again. On 23 April 2026 the CFTC filed a complaint alleging insider trading through one of them by an active-duty member of the US Army (Gómez-Cram et al. 2026).

Scale, and a disagreement worth stating

reach a much larger estimate by a different route. Screening every Polymarket contract from February 2024 to February 2026 — more than 93,000 markets and roughly 50,000 wallets — they score accounts on five indicators (bet size, timing relative to the event, profitability, concentration and directional consistency). Flagged traders win 69.9% of the time, more than sixty standard deviations above what chance would produce, and the authors put suspicious profits at approximately $143 million over the period, across episodes including a military strike, the Venezuelan capture, a platform’s own search rankings and a celebrity announcement.

That figure and the 1.3%-of-profits figure above are not obviously reconcilable, and this wiki does not present them as agreeing. The samples, windows and flagging thresholds all differ, and a screen tuned to catch more will catch more. Establishing whether the two estimates are measuring the same thing is a piece of work nobody appears to have done, and it is the first question a reader should ask of either number.

Concentration compounds the problem in a way neither study is designed to capture. documents a single trader operating through a set of linked wallets during the 2024 US presidential election, whose directional positions moved the displayed win probability above the level shown on competing platforms. Whether that trader held private information is unknown and beside the point: the same wallet structure that defeats a concentration screen also defeats an insider screen.

Why the law does not reach it

The reason insider trading belongs among the challenges rather than among the microstructure findings is that, on the current legal position, very little of it is prohibited.

set out the gap. Most event contracts are not securities, so the securities regime and its misappropriation doctrine do not apply. The Commodity Exchange Act reaches insider trading more narrowly than the securities laws do, having been built for markets in which the relevant information is about supply and demand rather than about a discrete external event. And wire fraud, the residual tool, struggles where the information was not obtained in a commercial relationship at all — the service member, the production assistant and the platform engineer are not obviously breaching any duty a court has previously recognised. The authors propose three responses: regulation at the platform level, contract-specific rules, and an expanded misappropriation doctrine.

Two structural features make this harder than the equivalent problem in equities. There is no issuer to impose a blackout period, because for most contracts there is no issuer at all; the entity in a position to restrict trading by people who know the answer often has no relationship with the exchange. And pseudonymous accounts mean that detection runs entirely on statistical inference over trading patterns, which is why every estimate in this subsection is a screen rather than a count, and why the wide disagreement between those estimates should be read as a fact about the difficulty of the measurement rather than as a dispute one side will win.

References

• Abid, Mohamed Nadhir. 2026. “The Polymarket Paradox: Manipulation, Whale Concentration, and Predictive Accuracy in the World’s Largest Prediction Market.” Working paper.
• Ag Alert. 2026. “California Grazing Sector Threatened as Goat Herder Wage Soars to $240K.” California Farm Bureau Federation. July 8, 2026. https://www.agalert.com/california-ag-news/archives/july-8-2026/california-grazing-sector-threatened-as-goat-herder-wage-soars-to-240k/.
• Agri-Pulse. 2026. “Lawmakers Seek Solution to Skyrocketing Goat Herder Wages.” Agri-Pulse Communications. 2026. https://www.agri-pulse.com/articles/25059-lawmakers-seek-solution-to-skyrocketing-goat-herder-wages.
• Akey, Pat et al. 2026. “Who Wins and Who Loses in Prediction Markets? Evidence from Polymarket.” Working paper.
• Arrow, Kenneth J. 1964. “The Role of Securities in the Optimal Allocation of Risk-Bearing.” Review of Economic Studies 31 (2): 91–96.
• Arrow, Kenneth J. et al. 2008. “The Promise of Prediction Markets.” Science 320 (5878): 877–78. https://doi.org/10.1126/science.1157679.
• Artemis. 2026. “Prediction Market Kalshi Partners with The Weather Company for Weather Hedge Settlements.” Artemis.bm. 2026. https://www.artemis.bm/news/prediction-market-kalshi-partners-with-the-weather-company-for-weather-hedge-settlements/.
• Auld, Tom. 2022. “Political Markets as Equity Price Factors.” Cambridge Working Paper in Economics 2264. Faculty of Economics, University of Cambridge. https://www.econ.cam.ac.uk/publications/cwpe/2264.
• Bartlett, Robert, and Maureen O’Hara. 2026. “Adverse Selection in Prediction Markets: Evidence from Kalshi.” Working paper, SSRN 6615739. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6615739.
• Berg, Joyce E., Forrest D. Nelson, and Thomas A. Rietz. 2008. “Prediction Market Accuracy in the Long Run.” International Journal of Forecasting 24 (2): 285–300.
• Breeden, Douglas T., and Robert H. Litzenberger. 1978. “Prices of State-Contingent Claims Implicit in Option Prices.” Journal of Business 51 (4): 621–51.
• Brier, Glenn W. 1950. “Verification of Forecasts Expressed in Terms of Probability.” Monthly Weather Review 78 (1): 1–3.
• Bürgi, Constantin, Wentao Deng, and Karl Whelan. 2026. “Makers and Takers: The Economics of the Kalshi Prediction Market.” CESifo Working Paper 12122. CESifo.
• Business Insurance. 2026. “Goat Herder Hedges Legislative Risk Using Prediction Markets.” Reporting based on a statement issued by Castle Technologies. 2026. https://www.businessinsurance.com/goat-herder-hedges-legislative-risk-using-prediction-markets/.
• Business Wire. 2026. “Fox to Integrate Kalshi Forecasts Across Fox News Media and Fox One Platforms.” April 7, 2026. https://www.morningstar.com/news/business-wire/20260407319986/fox-to-integrate-kalshi-forecasts-across-fox-news-media-and-fox-one-platforms.
• Cassar, Terence. 2026. “Regulating Prediction Markets in Europe Requires a Prediction Test.” Oxford Business Law Blog. March 31, 2026. https://blogs.law.ox.ac.uk/oblb/blog-post/2026/03/regulating-prediction-markets-europe-requires-prediction-test.
• CBS News. 2026. “New York City Council Probing Kalshi, Polymarket and Other Prediction Markets over Marketing Claims.” August 12, 2026. https://www.cbsnews.com/news/nyc-investigates-polymarket-kalshi-prediction-markets/.
• Chen, Hongyin et al. 2025. “Political Leanings in Web3 Betting: Decoding the Interplay of Political and Profitable Motives.” In Proceedings of the 17th ACM Web Science Conference 2025, 96–105.
• Chernov, Mikhail, Vadim Elenev, and Dongho Song. 2025. “The Comovement of Voter Preferences: Insights from U.S. Presidential Election Prediction Markets Beyond Polls.” NBER Working Paper 33339. National Bureau of Economic Research. https://www.nber.org/papers/w33339.
• Cleary Gottlieb. 2026. “Prediction Markets for Those Who Don’t Predict (and Those Who Do): An Overview of the Regulation of Prediction Markets.” April 14, 2026. https://www.clearygottlieb.com/news-and-insights/publication-listing/prediction-markets-for-those-who-dont-predict-and-those-who-do.
• CoinDesk. 2025a. “Polymarket, UMA Communities Lock Horns After $7M Ukraine Bet Resolves.” March 27, 2025. https://www.coindesk.com/markets/2025/03/27/polymarket-uma-communities-lock-horns-after-usd7m-ukraine-bet-resolves.
• ———. 2025b. “‘This Isn’t Decentralized,’ Says Polymarket Power User as Zelenskyy’s Suit Controversy Unfolds.” July 9, 2025. https://www.coindesk.com/markets/2025/07/09/this-isnt-decentralized-says-polymarket-power-user-as-zelenskyys-suit-controversy-unfolds.
• ———. 2026. “India Cracks down on Prediction Markets: Polymarket Goes Dark, Kalshi Could Be Next.” May 22, 2026. https://www.coindesk.com/markets/2026/05/22/india-cracks-down-on-prediction-markets-polymarket-goes-dark-kalshi-could-be-next.
• Coraggio, Giulio. 2026. “Prediction Markets Italy: Polymarket Blocked Again.” GamingTechLaw. 2026. https://www.gamingtechlaw.com/2026/07/prediction-markets-italy/.
• Crypto Times. 2026. “Trump Backs Prediction Markets as Family Holds Ties to Polymarket and Kalshi.” May 27, 2026. https://www.cryptotimes.io/2026/05/27/trump-backs-prediction-markets-as-family-holds-ties-to-polymarket-kalshi/.
• Cutting, Lucas E. et al. 2025. “Are Betting Markets Better Than Polling in Predicting Political Elections?” 2025. https://arxiv.org/abs/2507.08921.
• Dai, David, Ruizhe Jia, and Shihao Yu. 2026. “Settlement Manipulation in Prediction Markets.” July 1, 2026. https://arxiv.org/abs/2606.31675.
• Dechert LLP. 2026. “Perpetual Contracts Update: CME Takes CFTC to Court.” 2026. https://www.dechert.com/knowledge/onpoint/2026/6/addendum-to-perpetual-contracts.html.
• Decrypt. 2025. “Polymarket Rules ‘No’ on $237M Controversial Bet over Zelenskyy’s Suit.” 2025. https://decrypt.co/329210/polymarket-rules-no-237m-bet-zelenskyys.
• Deutsche Welle. 2026. “Germany: Betting on Elections Is Illegal — but Possible.” April 12, 2026. https://www.dw.com/en/germany-betting-on-elections-is-illegal-but-possible/a-76712766.
• Diercks, Anthony M., Jared Dean Katz, and Jonathan H. Wright. 2026. “Kalshi and the Rise of Macro Markets.” NBER Working Paper 34702. National Bureau of Economic Research. https://www.nber.org/papers/w34702.
• Dubach, Philipp D. 2026. “The Anatomy of a Decentralized Prediction Market: Microstructure Evidence from the Polymarket Order Book.” Replication package archived at DOI 10.5281/zenodo.19811426. May 15, 2026. https://arxiv.org/abs/2604.24366.
• Eichengreen, Barry et al. 2025. “Under Pressure? Central Bank Independence Meets Blockchain Prediction Markets.” Working paper, SSRN 5366862. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5366862.
• Enright, Brad. 2026. “Prediction Markets — Here’s What You Need to Know.” Gambling Commission. February 4, 2026. https://www.gamblingcommission.gov.uk/blog/post/prediction-markets-heres-what-you-need-to-know.
• European Securities and Markets Authority. 2026. “Public Statement on the Application of the National Product Intervention Measures on Binary Options to Event Contracts.” July 3, 2026. https://www.esma.europa.eu/.
• Finance Magnates. 2026. “Europe Has No Licensed Prediction Markets. ESMA Just Raised the Entry Bar.” 2026. https://www.financemagnates.com/fintech/europe-has-no-licensed-prediction-markets-esma-just-raised-the-entry-bar/.
• Financial Conduct Authority. 2019. “PS19/11: Product Intervention Measures for Retail Binary Options.” Policy statement. Financial Conduct Authority. https://www.fca.org.uk/publications/policy-statements/ps19-11-product-intervention-measures-retail-binary-options.
• ———. 2026. “Perimeter Report.” Corporate document. Financial Conduct Authority. https://www.fca.org.uk/publications/corporate-documents/fca-perimeter-report.
• Focus Gaming News. 2025. “Regulator Warns Social Betting Platforms Are Not Permitted in Germany.” 2025. https://focusgn.com/regulator-issues-warns-social-betting-platforms-not-permitted-in-germany.
• Fortune. 2026. “Controversy Swirls over Spanish Soccer Club Accused of Using Novel $600,000 Kalshi Wager to Bet on Its Relegation.” Fortune. June 12, 2026. https://fortune.com/2026/06/12/controversy-swirls-over-spanish-soccer-club-accused-of-using-novel-600000-kalshi-wager-to-bet-on-its-relegation/.
• Foucault, Thierry, Marco Pagano, and Ailsa Röell. 2013. Market Liquidity: Theory, Evidence, and Policy. Oxford University Press.
• Galton, Francis. 1907. “Vox Populi.” Nature 75: 450–51.
• Glosten, Lawrence R., and Paul R. Milgrom. 1985. “Bid, Ask and Transaction Prices in a Specialist Market with Heterogeneously Informed Traders.” Journal of Financial Economics 14 (1): 71–100.
• Gómez-Cram, Roberto et al. 2026. “Prediction Market Accuracy: Crowd Wisdom or Informed Minority?” Working paper, SSRN 6617059. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6617059.
• Grossman, Sanford J. 1976. “On the Efficiency of Competitive Stock Markets Where Trades Have Diverse Information.” Journal of Finance 31 (2): 573–85.
• Grossman, Sanford J., and Joseph E. Stiglitz. 1980. “On the Impossibility of Informationally Efficient Markets.” American Economic Review 70 (3): 393–408.
• Gruca, Thomas S., and Thomas A. Rietz. 2023. “Peering into the Black Box: Trader Strategies in the Iowa Electronic Markets.” Working paper.
• Harris, Larry. 2003. Trading and Exchanges: Market Microstructure for Practitioners. Oxford University Press.
• Hayek, Friedrich A. 1945. “The Use of Knowledge in Society.” American Economic Review 35 (4): 519–30.
• Holland and Knight. 2026. “Federal Appeals Court: CFTC Jurisdiction over Sports Event Contracts Likely Exclusive.” 2026. https://www.hklaw.com/en/insights/publications/2026/04/federal-appeals-court-cftc-jurisdiction-over-sports-event-contracts.
• iGaming Business. 2026. “ANJ Orders ISPs to Block Polymarket over Illegal Gambling Concerns.” 2026. https://igamingbusiness.com/prediction-markets/anj-orders-isps-block-polymarket-illegal-gambling-concerns/.
• Insurance Journal. 2026. “France Probes Weather Data Tampering After Surge in Polymarket Bets.” April 23, 2026. https://www.insurancejournal.com/news/international/2026/04/23/867026.htm.
• Iowa Electronic Markets. n.d. “About IEM.” https://iemweb.biz.uiowa.edu/about-iem/. Accessed: 2026-08-12.
• Jain, Prerna et al. 2025. “Election Arbitrage During the 2024 U.S. Presidential Election.” Working paper.
• Kagan, Nicole, and Rubens Baiocchi. 2026. “Calibration in Prediction Markets: Theory and Evidence.” Working paper.
• Kalshi. 2026a. “Kalshi for Small Business.” Kalshi. Marketing page. Cited only for the position sizes of the cases it chooses to showcase. 2026. https://kalshi.com/hedging.
• ———. 2026b. “How a New York Bar Owner Hedged a Knicks Promotion.” Kalshi News. Published by the exchange on which the trade took place. June 1, 2026. https://news.kalshi.com/p/kalshi-jeffrey-bar-hedge-knicks-nba-finals-small-business.
• KalshiEX LLC. 2026. “Product Certification: Will an Alternative Wage for Californian Goat Herders Be Authorized?” Filing submitted to the U.S. Commodity Futures Trading Commission via the CFTC Portal. Rulebook designation CAGOATWAGE. https://www.cftc.gov/filings/ptc/ptc07312613093.pdf.
• Kyle, Albert S. 1985. “Continuous Auctions and Insider Trading.” Econometrica 53 (6): 1315–35.
• Manski, Charles F. 2006. “Interpreting the Predictions of Prediction Markets.” Economics Letters 91 (3): 425–29.
• Moneywise. 2026. “NYC Bar Owner Used Kalshi to Hedge a Free-Tab Promotion.” Republished by Yahoo Finance; the reporting is attributed onward to Fortune and sourced via a Kalshi spokesperson. 2026. https://finance.yahoo.com/markets/options/articles/nyc-bar-owner-used-kalshi-114500509.html.
• Ng, Hin et al. 2026. “Price Discovery and Trading in Modern Prediction Markets.” Working paper.
• Norton Rose Fulbright. 2026. “The EU’s Approach to Prediction Markets and Event Contracts.” 2026. https://www.nortonrosefulbright.com/en/knowledge/publications/290d594a/the-eus-approach-to-prediction-markets-and-event-contracts.
• NPR. 2026. “French Police Probe Suspected Weather Device Tampering After Odd Polymarket Bet.” April 23, 2026. https://www.npr.org/2026/04/23/nx-s1-5797876/polymarket-paris-weather-bet.
• Palumbo, Nicholas. 2026. “A Microstructure Perspective on Prediction Markets.” Working paper.
• PYMNTS. 2026. “CFTC Steps in as New York Sues to Shut down Kalshi.” July 31, 2026. https://www.pymnts.com/legal/2026/cftc-steps-in-as-new-york-sues-to-shut-down-kalshi/.
• Quintenz, Brian. 2026. “Wall Street Has Used These Financial Contracts for Decades. Now a Goat Herder Does Too.” Fortune. Opinion column. The author is a former CFTC Commissioner (2017–2021) and an adviser and board member of Kalshi; the disclosure appears in the piece. Sole public source for the $50,000 premium figure. August 26, 2026. https://fortune.com/2026/08/26/cftc-commissioner-brian-quintenz-prediction-markets-kalshi/.
• Reichenbach, Felix, and Martin Walther. 2025. “Exploring Decentralized Prediction Markets: Accuracy, Skill, and Bias on Polymarket.” Working paper.
• Reuters. 2026. “Kalshi Adds India to Restricted List.” 2026. https://sg.finance.yahoo.com/news/kalshi-adds-india-restricted-list-235910717.html.
• Semafor. 2026. “Prediction Markets Are Rebuilding Insurance and It’s Getting Weird.” Opinion/analysis column. Source of the insurable interest argument and of the first account of the Osasuna transaction, several details of which were subsequently disputed — see the Next Event Horizon account listed below. June 9, 2026. https://www.semafor.com/article/06/09/2026/prediction-markets-are-rebuilding-insurance.
• Stershic, Amanda, and Kritee Gujral. 2020. “Arbitrage in Political Prediction Markets.” The Journal of Prediction Markets 14 (1): 69–99.
• Surowiecki, James. 2004. The Wisdom of Crowds. Doubleday.
• Tech Startups. 2026. “Polymarket Paid Influencers to Fake Winning Bets in 140 Million-View Marketing Campaign, WSJ Investigation Finds.” Reporting on a Wall Street Journal investigation. July 3, 2026. https://techstartups.com/2026/07/03/polymarket-paid-influencers-to-fake-winning-bets-in-140-million-view-marketing-campaign-wsj-investigation-finds/.
• Thaler, Richard H., and William T. Ziemba. 1988. “Parimutuel Betting Markets: Racetracks and Lotteries.” Journal of Economic Perspectives 2 (2): 161–74.
• The Block. 2025. “Polymarket Says Governance Attack by UMA Whale to Hijack a Bet’s Resolution Is ‘Unprecedented’.” March 26, 2025. https://www.theblock.co/post/348171/polymarket-says-governance-attack-by-uma-whale-to-hijack-a-bets-resolution-is-unprecedented.
• ———. 2026. “Coalition of 44 State Attorneys General Says CFTC Lacks Authority over Sports Prediction Markets.” July 28, 2026. https://www.theblock.co/post/410013/state-attorneys-general-cftc-letter-prediction-markets.
• The Guardian. 2026. “UK Prediction Market Boom Raises Regulatory Questions.” July 26, 2026. https://www.theguardian.com/world/2026/jul/26/uk-prediction-market-boom-regulation-polymarket-gambling.
• UMA Protocol. 2026. “DVM 2.0.” UMA Documentation. 2026. https://docs.uma.xyz/using-uma/voting-walkthrough/dvm-2.0.
• Wolfers, Justin, and Eric Zitzewitz. 2006. “Interpreting Prediction Market Prices as Probabilities.” NBER Working Paper 12200. National Bureau of Economic Research.