§ 03 REPORTS / DEFI / 2026-08-03

Prediction Markets from First Principles

FILED UNDER DEFI § 2026-08-03

A mechanism built to extract honest numbers from people who have no other reason to give you one — and what it became once it was finally allowed to be a business.

tl;dr

Contents

  1. Markets that trade beliefs
  2. A short history of betting on everything
  3. The lore
  4. The builders
  5. How the books actually work
  6. Oracles, resolution, and the risk stack
  7. Where the volume actually is
  8. Take rates and the business of belief
  9. The regulatory arbitrage
  10. Futarchy, or markets that decide
  11. Where this goes
  12. Sources and methodology

Markets that trade beliefs

Start with the problem, not the product.

In "The Use of Knowledge in Society" (American Economic Review 35(4), September 1945, pp. 519–530), Hayek made an observation that has aged better than most of twentieth-century economics: the knowledge a society needs in order to act "never exists in concentrated or integrated form, but solely as the dispersed bits of incomplete and frequently contradictory knowledge which all the separate individuals possess." No planner can assemble it, because most of it is local, tacit, and perishable. What the price system does — Hayek's phrase — is function as "a system of telecommunications." A single number carries the compressed judgment of everyone willing to act on it.

A prediction market is that argument run backwards. Instead of using prices to allocate goods, you construct an asset that allocates nothing. It has no cash flows, no productive use, no terminal value beyond a binary event. Its only purpose is to exist so that people will price it, and in pricing it, publish what they believe.

The mechanism is almost embarrassingly simple. A contract pays $1 if event E occurs and $0 otherwise. Under risk neutrality and no frictions, its price is the market's aggregate probability of E. Wolfers and Zitzewitz ("Prediction Markets," Journal of Economic Perspectives 18(2), 2004, pp. 107–126) formalized this and catalogued the contract designs that follow from it: winner-take-all contracts reveal probabilities, index contracts (vote share, revenue) reveal means, spread contracts reveal medians. The standard objection — that price equals probability only under conditions nobody actually inhabits — is addressed in their companion NBER paper, which shows the price-as-mean-belief approximation stays tight across plausible degrees of risk aversion.

Compare it to the instrument it competes with. A poll pays nothing for accuracy, weights the person who has studied the question for six months identically to the person forming an opinion during the call, and has no error-correction loop — if a poll is wrong, nothing happens to the pollster until the event resolves. A market inverts all three. It pays for being right. It weights participants by conviction, because conviction is expressed as capital at risk. And anyone who thinks the price is wrong is paid to fix it, in proportion to how wrong they think it is.

That is the whole idea. Everything else here is an argument about implementation.

There is a second tributary, and it matters more than it usually gets credit for. The elicitation literature asks what payoff function makes an honest forecaster better off than a strategic one. A proper scoring rule is one where a forecaster maximizes expected score only by reporting true subjective probabilities — the Brier score (Glenn Brier, 1950, built for weather forecasting) and the logarithmic score being canonical, with the general theory laid out in Savage's 1971 work on eliciting personal probabilities. Scoring rules give you honesty from one forecaster; markets give you aggregation across many. For decades these were separate literatures.

Robin Hanson welded them together. His market scoring rule observes that a scoring rule shared sequentially — each participant pays off the previous reporter and takes over the report — is both things at once: an automated market maker when the market is thick, a plain scoring rule when it is thin. If nobody shows up, the last reporter still gets scored honestly. If everybody shows up, it behaves like a market. The implementation is the Logarithmic Market Scoring Rule, developed alongside "Combinatorial Information Market Design" (Information Systems Frontiers 5, 2003), which worked out how to let traders bet on joint distributions without turning the market maker into a money pump.

Crypto-native readers should sit with that. LMSR is an automated market maker with a bounded loss parameter and a closed-form pricing function, designed for permissionless liquidity provision against an unknown flow of informed traders. It predates Uniswap by roughly fifteen years. The AMM was invented to solve an information problem, not a trading problem — and it arrived in DeFi stripped of the epistemics that motivated it.

So why does this keep getting rebuilt? Because the mechanism is cheap, the claim is falsifiable, and the claim keeps surviving falsification. Every generation rediscovers that a market is the least expensive machine ever devised for extracting an honest number from people who have no other reason to give you one. And then, with remarkable consistency, someone shuts it down.

A short history of betting on everything

Betting on elections was not invented in 1988. Rhode and Strumpf ("Historical Presidential Betting Markets," JEP 18(2), 2004, pp. 127–141) document large, organized markets for U.S. presidential elections operating in New York from 1868 to 1940 — at the 1916 peak, roughly $165 million in 2002 dollars wagered in a single cycle, more than twice total campaign spending that year, with betting activity at times dominating transactions on the Wall Street stock exchanges. In the pre-scientific-polling era these markets "did a remarkable job forecasting elections," and remained broadly efficient despite documented manipulation attempts. They died of two causes: Gallup's 1936 polling breakthrough destroyed the demand for odds-as-news, and New York criminalized organized election betting.

Accuracy first. Suppression second. Rebirth somewhere else. The pattern is already fully formed by 1940.

Iowa, 1988. The Iowa Political Stock Market launched in June 1988 with real money on Bush versus Dukakis. Forsythe, Nelson, Neumann and Wright's founding analysis (AER 82(5), 1992, pp. 1142–1161) reported that the market "worked extremely well, dominating opinion polls in forecasting the outcome," and introduced the marginal trader hypothesis: most traders exhibit judgment biases, but prices are set by a smaller pool of unbiased, active traders. You do not need a wise crowd. You need a wise margin.

The IEM survived because it agreed to be small. CFTC no-action letters 92-04(a) and 93-66 conditioned relief on non-profit academic operation, no commissions, and investment caps in the $5–$500 range — the first formal U.S. accommodation of a prediction market, and the template for every one that followed: tolerated because it was explicitly not a business.

The accuracy held up for decades. Berg, Nelson and Rietz (International Journal of Forecasting 24(2), 2008, pp. 285–300) compared IEM vote-share prices against 964 national polls across the five presidential elections from 1988 to 2004: the market was closer to the eventual outcome 74% of the time, and dominated polls at horizons beyond 100 days. That last clause is the one that matters. A market that merely echoed polls could not beat them a hundred days out.

The corporate side branch. Inside companies, Cowgill and Zitzewitz (Review of Economic Studies 82(4), 2015) studied internal markets at Google, Ford and an anonymized third firm, building on Plott and Chen's HP work: despite thinness, weak incentives, restricted entry and traders with ulterior motives, the markets improved on internal expert forecasts at all three firms, by as much as a 25% reduction in mean squared error. Google's markets also showed optimism bias concentrated among new hires and on days the stock rose, declining with experience. The market wasn't just forecasting. It was reading the org.

DARPA, 2003. The purest kill in the record. Per Hanson's archive, FutureMAP ran from a May 2001 call for proposals to roughly six teams holding ~$100K awards by summer 2003. The Policy Analysis Market was built by Net Exchange, founded by Caltech's John Ledyard, with Hanson and David Porter on market design. It would have traded Middle East geopolitical indicators — economic health, civil stability, military activity — plus conditional contracts estimating policy impacts.

On July 28, 2003, Senators Wyden and Dorgan denounced it as a "grotesque" federal terrorism-futures betting parlor. It was terminated on July 29 — roughly twenty-four hours later, two months before trading was to open with its first thousand registrants. Poindexter resigned; the $8 million in requested further funding was never appropriated. Killed by optics before it generated a single data point, on a characterization that misdescribed a design centered on macro indicators.

Intrade, 2003–2013. Dublin-based Intrade became the de facto public probability feed of the 2000s. Then: a 2005 cease-and-desist against parent Trade Exchange Network, and on November 26, 2012 — three weeks after the election — CFTC charges for violating the off-exchange options ban and filing false certifications about restricting U.S. retail access. The court eventually found 5,503 unlawful binary option contracts traded by U.S. customers between September 2007 and June 2012, with a $3 million penalty in 2018. Intrade banned U.S. customers the day the suit was filed and suspended all trading on March 10, 2013 citing "financial irregularities." Investigation found a cash shortfall of roughly $700,000 against member balances, and Fortune documented nearly $1.5 million in suspect transfers from customer-trust funds to accounts of CEO John Delaney, who had died on Everest in May 2011. Overdetermined: a jurisdictional strike that severed its largest market, plus a custodial failure. Note precisely what those two things are. A chokepoint and a trusted intermediary. The architecture of the on-chain generation is a response to that specific pair.

Round Peak Kill Rebirth
NY election betting, 1868–1940 ~$165M (2002 $) in 1916 Anti-gambling law; polling displaces odds-as-news IEM, 1988
IEM, 1988– Beat 964 polls 74% of the time Never killed — capped at $500 per trader (still running)
PAM, 2001–2003 Never traded ~24 hours of political optics Design work migrated to academia and industry
Intrade, 2003–2013 Global election probability of record CFTC suit + custodial shortfall PredictIt (2014), then offshore and on-chain
Modern round Polymarket fined $1.4M, Jan 2022; PredictIt relief rescinded, litigated in Clarke — Kalshi, D.D.C., Sept. 2024 — rebirth inside the perimeter

The mechanism behind the loop is structural, not accidental. Prediction markets are institutionally homeless. They are simultaneously gambling under state law, options under the CEA, and forecasting speech adjacent to the First Amendment — and each accuracy demonstration raises their public salience faster than it grows their political constituency. What changed after 2013 is that the rebirth venue became credibly non-killable: first offshore, then on-chain, and then — for the first time — onshore and licensed.

The lore

Mechanisms are argued. Stories are what actually transmit.

Vitalik's Augur position. The canonical account is his February 18, 2021 essay, "Prediction Markets: Tales from the Election". He bought roughly $2,000 of NTRUMP — "Trump loses" tokens — on Augur in September 2020, then scaled to a total of $308,249, netting over $56,803. The detail that made it lore: he made most of those bets after Trump had already lost. NTRUMP was still trading around $0.85 weeks after the networks called it.

He offers four hypotheses for why the mispricing survived: fear of smart-contract failure (dismissed — ITRUMP, the tradable "invalid" token, stayed consistently under $0.02), capital costs (~$1M locked for ~2 months for ~35% annualized), interface friction, and what he called intellectual underconfidence. The sharpest is the asymmetry: near price extremes "the game is very lopsided in favor of those who are trying to push the probability away from the extreme value." Trump backers needed only about $60,000 to absorb his $308,249. That explains more mispricing in this asset class than any behavioral story. Selling a 3% tail requires locking 97 cents to earn 3.

The same essay is the first real oracle stress test in the record. Augur v2's Trump market carried real money while the losing side litigated the outcome nationally, and resolution took "two rounds of disputes, but the no side nevertheless cleanly won." The oracle held; the failure was economic, not epistemic. That diagnosis is, almost line for line, the product spec Polymarket later filled.

Augur's earlier contribution was less flattering. Within weeks of its July 2018 launch, users created death markets on Trump, Bezos, Betty White and John McCain. The Forecast Foundation claimed no ability to censor or terminate them; the only mitigation was REP reporters resolving unethical markets as invalid, and in practice they drew trivial volume. But the trilemma was now on the table — unstoppable markets, moral curation, regulatory legibility, pick two — and every subsequent platform is a different answer to it.

Per DL News (Jan. 28, 2026, single-sourced to on-chain analysis), Vitalik made roughly $70,000 on Polymarket in 2025 on about $440,000 of principal, around 16% annualized, betting against long shots — "I look for markets that are in 'crazy mode' and then bet that 'crazy things won't happen.'" Same trade as 2020, industrialized. On November 9, 2024 he generalized it into "From prediction markets to info finance": start from a fact you want to know, then deliberately design a market to elicit it. Prediction markets are one instance of a category, not the category.

Do Kwon's two bets. In March 2022 the pseudonymous trader Sensei Algod, who had been publicly calling Terra's Anchor/UST design a Ponzi, offered $1 million that LUNA would be lower in a year; Kwon accepted within minutes, both sides depositing to an escrow wallet controlled by Cobie. A day later GCR struck a parallel $10 million bet. Twenty-two million dollars, one Ethereum address, one person's judgment. UST depegged two months into the twelve; per Cointelegraph, Cobie paid the winners out early, hedged a freak recovery by buying LUNC, and lost the hedge in the FTX bankruptcy. By settlement day Kwon was a fugitive under an Interpol red notice. Read as a mechanism, it is a prediction market with every component replaced by a person: no continuous price, no exit before expiry, no way for a third party to express the same view, settlement contingent on one individual's discretion and custody. Twitter bets want to be prediction markets.

Théo. A French national trading through accounts including Fredi9999, Theo4, PrincessCaro and Michie built a position of roughly $80 million across Polymarket's Trump-wins, swing-state and popular-vote markets in October 2024 (CBS/60 Minutes). Profit estimates moved as attribution improved: Bloomberg initially near $48M across four accounts; a Chainalysis analysis reported by The Block identifying nine accounts and ~$78.7M; Bloomberg's revised analysis putting the haul at ~$85 million across roughly eleven accounts (The Free Press).

The thesis is the part worth keeping. He described the position as a bet against the validity of U.S. polling data: believing conventional polls suffered shy-Trump non-response, he commissioned private polls using the neighbor question — who do you think the people around you will vote for — which returned materially more Trump support than the direct-intent question (CBS, Entrepreneur). His buying pushed Polymarket's Trump odds well above polling averages, which produced the obvious objection: market signal, or one man's bankroll? Both, and they are not separable. He was the price — and he was also the only participant who paid to generate private information rather than reprocess public information. The mechanism working exactly as specified, and from the outside indistinguishable from the mechanism being broken.

The Romney whale. In the 2012 Intrade market, Rothschild and Sethi analyzed the full two-year transaction history — roughly 6,300 unique accounts and 12.9 million contracts. Their "Trader B" accounted for about one-third of all money bet on Romney in the final two weeks (a window in which roughly 3.5 million contracts traded) and about a quarter over the whole cycle, for a loss of close to seven million dollars. Steady price-supporting bids, indifference to news: most consistent with deliberate manipulation, plausibly to sustain fundraising, morale and turnout, though hedging of correlated exposure cannot be excluded. The standing bid wall kept displayed prices eerily stable on election night even as results arrived.

The market's Obama-favorable signal was broadly correct anyway, and arbitrageurs — Intrade traded Romney-rich against Betfair — were paid to lean against him. This is Hanson and Oprea (Economica 76(302), 2009) in the wild: a manipulator with a price target is, to informed traders, a noise trader whose losses subsidize information production. Manipulation is not free money for the manipulator. It is a bounty posted for anyone who knows better.

Put 2012 and 2024 side by side and you have the whole ambiguity. One whale fought the truth and paid seven million dollars for the privilege. The other carried private truth and was paid eighty-five. Both moved the price. Only the resolution tells you which was which — which is precisely why the resolution layer is where this asset class lives or dies.

Not all lore is whales. Domer, an ex-poker pro, is Polymarket's #1 all-time trader by volume and profit — 5,000+ markets, over $400M wagered for roughly $3M of profit in a year: craft, not conviction. Whereas Balaji Srinivasan's March 2023 bet of $1M against $1 that BTC would exceed $1M within 90 days, closed early with $1.5M paid out at ~97% below target, was never a probability. It was an advertisement. A public wager and a forecast are different objects that look alike.

And then there is the lore nobody wanted, which is about resolution rather than positioning: the Barron Trump / $DJT market of June 2024, the Ukraine minerals market of March 2025, the Zelensky suit market at nine figures of volume, the $10.5M Venezuela invasion market settled No in January 2026. Four markets that turned not on what happened but on who got to say what happened, and three failure modes hiding in the cluster: token-weighted oracles voting wealth instead of truth, Schelling-point herding on subjective wording, and unpriced information asymmetry. None of them is a matching-engine problem. Resolution is the product, and it gets a section of its own below.

The builders

There are two ways to build a prediction market, and for a decade they did not talk to each other.

One lineage starts from the premise that the hard problem is permission — that the reason event markets never scaled was that no regulator would let them, and therefore the answer is a system nobody has to license. This is the crypto-native line: Augur, Gnosis, Polymarket, and the long tail behind them. The other lineage starts from the premise that the hard problem is legitimacy — that event contracts are derivatives, derivatives have a regulator, and the answer is to go get designated. This is the line that runs from the Iowa Electronic Markets through PredictIt to Kalshi, and then, very suddenly, through every brokerage and sportsbook in America.

Both lineages were right about their own constraint and wrong about the other's. What is interesting in 2026 is not which one won. It is that they converged, and that you can read the convergence in the cap tables.

The crypto-native line

Augur was the maximalist case, and it is worth taking seriously precisely because it failed on its own terms rather than someone else's. Founded in 2014 by Jack Peterson, Joey Krug, and Jeremy Gardner under the Estonian Forecast Foundation OÜ, its REP crowdsale in August 2015 was one of the first ICOs on Ethereum; mainnet arrived in July 2018 after three years. Every layer was decentralized: on-chain order book, ETH-denominated trading, market creation open to anyone, and a native oracle where REP holders reported outcomes and could escalate disputes all the way to a chain fork.

Nobody used it. Augur went from 265 daily active users in early July 2018 to 37 by August 8 — one month after launch. Gas costs, seven-day resolution, ETH volatility contaminating every position's real return, a desktop app that felt like a compiler. Augur v2, deployed July 2020, fixed the mechanical complaints — DAI settlement, 24-hour resolution, fee-free order creation — and it did not matter, because the team's attention had already moved (Krug to Pantera, later Founders Fund). There is no shutdown announcement; the Forecast Foundation simply stopped shipping. The lesson is not "decentralization doesn't work." It is narrower: Augur decentralized everything, including the parts where decentralization buys the user nothing. Nobody wants a permissionless order book. They want a book with orders in it.

Gnosis is the counter-case: it lost as a venue and won as a standards body. It started in 2015 as a prediction-market company, ran one of the larger 2017 ICOs, then pivoted into Safe, CoW Protocol and Gnosis Chain. What survived was the Conditional Tokens Framework — the ERC-1155 outcome-token standard that Polymarket and Seer still build on. Its consumer venue, Omen, was built by DXdao on the CTF in February 2020 with an FPMM AMM and Reality.eth resolution escalating to Kleros, and faded with DXdao. The CTF outlived every frontend built on it, including Gnosis's own.

Then Polymarket, which did the one thing neither predecessor would: it picked which decentralization to keep. Shayne Coplan launched it in June 2020, at 22 — outcome tokens on Gnosis's CTF, USDC collateral, Polygon settlement, and matching in a hybrid central limit order book, off-chain matching with on-chain non-custodial settlement. In the middle of DeFi summer, when every new venue was an AMM, Polymarket built a book. That was the correct call and we will spend most of the next section on why.

The regulatory arc is the whole crypto metabolism compressed into one company. A CFTC fine of $1.4 million in January 2022 for unregistered event-based binary options, US users ordered off. Offshore scale. The 2024 election, where the presidential-winner market alone did more than $3.2B before election day, then a post-election collapse critics called structural. Then the return: in July 2025 Polymarket bought QCX LLC and QC Clearing — a CFTC-licensed contract market and clearinghouse — for $112M, acquiring a license rather than applying for one. QCX already held it before the deal closed; the CFTC's September 2025 action was a no-action letter granting relief on certain recordkeeping and disclosure requirements, not an operating approval. The Amended Order of Designation enabling FCM-intermediated US access came November 25, 2025; the US app launched in early December, with full iOS availability following in May 2026.

And the capital: ICE, the parent of the New York Stock Exchange, committed up to $2B in October 2025 at a ~$9B post-money valuation, becoming global distributor of Polymarket's event data, and completed that commitment with a final $600M tranche in March 2026. Two billion total, not two-point-six. By June 2026 the company was citing annualized revenue above $1B, with US volume going from ~$50M/day in mid-May to more than $200M on June 20 — a company disclosure, not audited financials.

Behind Polymarket sits a long tail that mostly differentiates on where the market lives rather than how it works: Azuro as a B2B liquidity layer whose "Liquidity Tree" lets a singleton pool back thousands of concurrent markets behind dozens of white-label frontends; SX Bet's peer-matched sports book on its own Arbitrum Orbit chain (lifetime figures self-reported); Seer, built by the Kleros team on conditional tokens plus Reality.eth plus Kleros arbitration, the most credibly decentralized live stack and, predictably, one of the smallest. The 2025–26 wave is more interesting for its distribution theses than its mechanisms — Limitless on Base-native short-dated crypto binaries, Myriad embedding markets inside news articles, Drift BET putting event markets inside a perps DEX's cross-margin engine so 30+ tokens serve as yield-bearing collateral while positions are open. XO Market is the sharpest of them, because creator revenue share on user-created markets tests the one thing Polymarket won't do ($6M seed in April 2026 led by 20VC with Picus Capital, Coinbase Ventures and Venture Together, ~$420M cumulative volume by July).

And Hyperliquid, which matters structurally more than its numbers suggest. HIP-4 outcome-contract trading activated on mainnet in early May 2026 with USDH settlement; day-one volume was 6.05M contracts against Kalshi's 546M and Polymarket's 190M. The genuinely permissionless deployment phase — 500,000 HYPE deployer stake, slashable by validator vote — entered testnet at the end of July 2026 with no announced mainnet date. What ran on mainnet from May was validator-approved templates — worth being precise about, because "permissionless prediction markets on the deepest onchain derivatives venue" is a different claim from what shipped.

The regulated line

The regulated lineage is older and, until 2024, almost comically smaller.

The Iowa Electronic Markets have run since 1988 under the CFTC's original 1992 no-action letter, permitting real-money trading for research and teaching with deposits capped at $5–$500. It is still live — it opened 2025 Canadian federal election seat-share contracts in February 2025. Academically it is the proof of concept for the entire field. Commercially it is a rounding error, and that is the point: the no-action structure that made it legal also made it unable to matter.

PredictIt inherited both the structure and the ceiling — operating since 2014 under a no-action letter to Victoria University of Wellington, capped at 5,000 traders and $850 positions. After the CFTC withdrew the letter in August 2022 and ordered a wind-down, PredictIt sued, won a Fifth Circuit injunction, and the Western District of Texas entered final judgment in its favor on July 22, 2025. The amended CFTC Letter 25-20 removed the trader cap and raised per-contract limits to $3,500. A win, and still a market that cannot hold an institution's position.

Kalshi is the line's answer to that ceiling: don't get a letter, get a license. Founded in 2018 by Tarek Mansour and Luana Lopes Lara, it received DCM designation in November 2020 — the first exchange designated specifically to list event contracts — and later stood up its own clearinghouse, Kalshi Klear, registered as the 18th DCO in 2024. Then it litigated, winning summary judgment against the CFTC's "gaming" prohibition in September 2024, surviving the D.C. Circuit's stay denial on October 2, 2024 in time to run election markets through the cycle, and watching the CFTC drop its appeal in May 2025 — which handed the whole DCM industry a precedent. The state-preemption fight that followed is genuinely split, and it belongs to the arbitrage section below.

The valuations went vertical — $185M at $2B in June 2025, $300M at $5B in October, $1B at $11B in December, $1B at $22B in May 2026 led by Coatue — and so did the volume, which the data section takes up in detail.

But Kalshi's actual masterstroke was distribution. It became the back-end. Robinhood launched event contracts in March 2025 routed to KalshiEX and has driven more than 50% of Kalshi's total volume since, the two splitting a 2¢-per-contract fee evenly. That worked so well that Robinhood decided it wanted the whole stack: in a deal that closed January 21, 2026, Robinhood and Susquehanna acquired ~90% of MIAXdx — a DCM, DCO and SEF — and renamed it Rothera. Rothera launched June 4, 2026 and did $2.1B in notional volume in under a month, ranking fourth among US venues. And the economics arrived faster than anyone modeled: in Q2 2026 Robinhood reported 13.6B event contracts traded and $156M of prediction-market revenue — more than its crypto revenue ($100M) and its equities revenue ($129M).

Read that again. A retail brokerage now makes more from event contracts than from equities.

The sportsbooks noticed, and moved: DraftKings, FanDuel, Underdog and Sporttrade all now hold or are seeking DCM exposure. Their reasons are entirely economic, and the arbitrage section takes them up in full.

Venue Legal wrapper Book type Settlement Resolution
Augur (v1/v2) none (permissionless) on-chain order book Ethereum, ETH then DAI REP staking → fork
Omen / Seer none FPMM AMM Gnosis Chain Reality.eth → Kleros
Polymarket (offshore) none (geoblocked US) hybrid CLOB, off-chain match Polygon, USDC.e → pUSD UMA optimistic oracle; Chainlink for price markets
Polymarket US QCX DCM + QC Clearing DCO, FCM-intermediated CLOB fiat, fully collateralized exchange rulebook
Kalshi KalshiEX DCM + Kalshi Klear DCO CLOB, price-time priority self-cleared, fully collateralized internal, pre-designated sources
PredictIt CFTC no-action letter (amended 2025) CLOB fiat, $3,500/contract cap operator
IEM CFTC no-action letter (1992) CLOB fiat, $5–$500 deposits university
Betfair UK Gambling Commission back/lay exchange book fiat operator
Azuro none peer-to-pool EVM chains Chainlink sports feeds
Robinhood / Rothera FCM → KalshiEX; ~90% owner of Rothera DCM/DCO/SEF CLOB fully collateralized exchange rulebook
DraftKings Railbird DCM + DKeX CLOB fully collateralized exchange rulebook
Hyperliquid HIP-4 none on-chain CLOB USDH validator-mediated

Venue characteristics as of August 2026.

Three things the trajectories teach.

Decentralization is a dial, and the winners turned it down. Augur decentralized everything and died; Polymarket centralized curation and matching, kept settlement non-custodial, and won; Seer and Omen stayed principled and stayed tiny.

Licensing is an asset class. Polymarket paid $112M for QCEX. Robinhood and Susquehanna bought MIAXdx. DraftKings bought Railbird. Underdog bought Aristotle. Nobody in this cycle applied for a license they could buy. The binding constraint was never technology or even law. It was time.

The two lineages have merged. Polymarket bought a DCM to come onshore; Kalshi began tokenizing its event contracts on Solana in December 2025 via DFlow and Jupiter, collapsing the boundary from the other side. "Crypto-native versus regulated" described 2020–2024. It does not describe now.

How the books actually work

Underneath the corporate history is a much older argument about how you get a price out of a thin market. It has had three answers.

Scoring rules. Robin Hanson's Logarithmic Market Scoring Rule is the ancestor of every automated market maker in crypto. Define a cost function over outstanding share quantities:

C(q) = b · ln(Σᵢ e^(qᵢ/b))

A trader moving the state from q to q′ pays C(q′) − C(q). The instantaneous price of outcome i is the softmax pᵢ = e^(qᵢ/b) / Σⱼ e^(qⱼ/b) — prices sum to 1 and behave as probabilities by construction (Gnosis LMSR primer). The operator's worst-case loss in the binary case is b·ln 2, generalizing to b·ln n (Pennock). That bounded loss is the design: the subsidy a sponsor pays to obtain a continuously quoted price in a market with no natural counterparties.

Which is exactly the problem here: a market on "will Ukraine sign a minerals deal before April" has no dealer, no two-sided flow, no hedgers. LMSR manufactures a counterparty and prices the subsidy in advance.

Its defects were formalized early, by the people who built it. Othman, Pennock, Reeves and Sandholm: current automated market makers "are unable to adapt to liquidity" — b is fixed at creation, so a $1,000 trade moves the price identically whether the market has done $10k or $10M — and "under normal circumstances, the market maker runs at a deficit" (EC'10; ACM TEAC 2013). Their LS-LMSR lets b grow with volume and lets the operator profit, at the cost of prices summing to slightly more than $1 — a vig. The honest version of a market maker is one that charges.

Constant-product pools, and why LPs bleed. The DeFi generation replaced scoring rules with Uniswap-shaped pools over outcome tokens. The substrate was Gnosis's Conditional Tokens Framework: an ERC-1155 multi-token contract where splitPosition turns collateral into a complete set of outcome tokens and mergePositions turns a complete set back into collateral (developer guide). Omen and early Polymarket ran a FixedProductMarketMaker over those tokens. Every trade in Polymarket's first two years was a swap against such a pool.

Then Polymarket left, and the reason it left is the most important mechanism-design fact in this field.

An outcome token converges deterministically to $0 or $1. A passive constant-product LP is therefore short precisely the option informed traders are long: as news arrives, arbitrageurs walk the pool to the new probability, and the LP sells the winning token too cheap and buys the losing token too dear, the entire way. Paradigm's pm-AMM paper states it without hedging — with existing AMMs, prediction-market LPs "are essentially guaranteed to lose all of their value once the prediction market expires," because outcome-token volatility explodes as expiry approaches at intermediate probabilities (pm-AMM, November 2024). Their proposed fix — a "uniform" AMM whose loss-versus-rebalancing is proportional to portfolio value at every price, plus a dynamic variant that decays liquidity into expiration — is itself an admission that constant-liquidity curves are structurally wrong for any asset with a known terminal date.

Near resolution, only quotes that can be cancelled in milliseconds can price news. Curve inventory cannot. It just donates.

That is why every venue at scale is an order book, and why the AMM designs — Omen, Overtime's Sports AMM, TrueMarkets on Uniswap V3 — cap out at niche volume no matter how elegant the curve.

Peer-to-pool, and parimutuel. There is a third option, which is to stop pretending LPs are traders and let them be the house.

Azuro splits the market-maker role into three actors: data providers push priced odds, LPs fund a singleton pool, and the protocol books bets against that pool. The vig lives in the odds margin, not in an explicit fee — LPs "earn through the spread embedded in sell-side odds pushed by Data Providers," and pool P&L is "the difference between the tokens seeded from the pool into the Conditions and the tokens returned to the pool after those Conditions are resolved" (Azuro docs). Per-event risk is capped by "reinforcement," an allocation equal to the maximum loss the pool can take on a condition, and the LiquidityTree lets every market draw against the full pool while keeping deposit and withdrawal accounting correct. Azuro's own docs are blunt about what this means: ">95% likelihood that LPs will return a positive yield on positions held above one month," with sub-week positions expected negative and "negative yield can extend… in the event of bettor outperformance." That is bookmaker economics, stated correctly. LPs are the house. Houses lose to sharp flow.

Parimutuel pools need no market maker and carry zero inventory risk, but produce no firm price until the pool closes and no exit before resolution. They survive at racetracks and lost everywhere else, because a market that prices once aggregates almost nothing.

Polymarket's book, in detail

Polymarket's exchange is, in its own words, a "hybrid-decentralized exchange model wherein there is an operator that provides matching/ordering/execution services" while "settlement happens on-chain, non-custodially" (CTF Exchange docs). Concretely:

Orders are EIP-712 signed typed structures. A user signs off-chain and submits to the CLOB; the operator matches crossing orders and periodically submits matchOrders transactions to the CTFExchange contract on Polygon, which re-verifies signatures and moves assets atomically. The operator can censor and reorder. It cannot steal.

Each binary market is a pair of ERC-1155 conditional tokens issued by the CTF, and the contract guarantees one YES plus one NO always redeems for one unit of collateral after resolution. YES + NO = $1 is a hard on-chain invariant, which means selling YES is economically identical to buying NO. The exchange exploits this with three settlement paths (Overview.md; the Solidity enum is MatchType { COMPLEMENTARY, MINT, MERGE }):

This is the trick that makes the book work. The order book never needs a natural seller of YES. A 60¢ YES bid and a 40¢ NO bid cross via mint, and the two sides of the market are one fungible liquidity pool rather than two thin ones. It is the CLOB equivalent of the complete-set invariant, and no AMM design gets it for free.

For multi-outcome events — an election with a dozen candidates — the NegRiskAdapter enforces the joint constraint across a set of mutually exclusive binaries: one NO share in any market converts into one YES share in every other market, and converting NO positions across k markets simultaneously additionally returns (k−1) units of collateral. That makes "NO on candidate A" fungible with "YES on everyone else," which is what lets arbitrageurs pin Σ(YES) ≈ 1 across dozens of outcomes with bounded capital. "Augmented" neg-risk adds placeholder outcomes and an explicit "Other" bucket so a late-emerging candidate doesn't break the partition (docs).

The plumbing changed materially in 2026: a hard cutover from CTF Exchange V1 to V2 in April 2026, after which V1 orders were rejected (Dubach 2026), introducing an upgradeable wrapped-collateral token, pUSD, which wraps both USDC and USDC.e through permissionless onramp and offramp (ctf-exchange-v2).

And fees, finally, exist. Polymarket ran effectively 0% for years; in 2026 it introduced dynamic taker fees of the form fee = C × feeRate × p × (1−p), category-dependent, with "makers are never charged fees" (help center) and third-party trackers dating the full rollout to March 30, 2026. The rates themselves are in the take-rate table below.

Notice the shape of that formula. It is proportional to p(1−p), the variance of a binary payout. That is not an accident and we will come back to it.

Kalshi's book

Kalshi is legally two entities and mechanically one conventional exchange: KalshiEX LLC, the CFTC-designated contract market, and Kalshi Klear LLC, the clearinghouse — so Kalshi self-clears fully collateralized contracts inside its own regulated DCO rather than settling on a blockchain. The venue is a price-time-priority CLOB quoted in cents, 1¢ to 99¢.

The fees are parabolic in price for the same reason Polymarket's are: taker fee = round-up(0.07 × C × P × (1−P)), about $0.0175 per contract at 50¢, vanishing toward the extremes (Kalshi fee schedule; cross-checked at Market Math and OddsShopper).

Two venues, different legal universes, converging on the same fee curve. Both are pricing the option value the taker extracts, and that value is p(1−p).

Who actually makes these markets

On Kalshi you can name them. Susquehanna became Kalshi's first dedicated institutional market maker in April 2024 — "the first large-scale institution to onboard to Kalshi as a committed market maker" — and now runs a public prediction-markets desk. The relationship has legal exposure: June 2025 state-court civil complaints in Massachusetts, Ohio and South Carolina, brought by gambling-loss-recovery vehicles, name Susquehanna International Group, LLP as Kalshi's liquidity provider. The same firm later co-acquired Rothera with Robinhood. Follow the market maker and you find the cap table.

On Polymarket the maker set is pseudonymous wallets, and the distribution is genuinely different from equities. Tick-level analysis of a 600-market panel finds "broad maker-wallet diversity with a concentrated tail," with a median Herfindahl index of 0.031 — roughly 32 effective makers per market (Dubach 2026). US equities, by contrast, have a handful of HFT firms doing about half the volume. Polymarket's book is flatter and more amateur, which is both a decentralization win and a liquidity problem.

Why would anyone make a zero-fee book? Because Polymarket pays them directly. The liquidity rewards program samples every book once a minute — 10,080 samples per weekly epoch — and scores each maker's resting orders with a quadratic rule S(v,s) = ((v−s)/v)² · b, where v is the market's maximum qualifying spread and s the order's distance from the size-adjusted midpoint. Two-sidedness is enforced softly in mid-range markets (single-sided scores divided by 3.0) and hard when the midpoint sits outside [0.10, 0.90], where "liquidity must be double-sided to score." Rewards pay pro rata, daily at midnight UTC (docs). The quadratic matters: an order resting at half the maximum spread earns four times less than one at the touch — a steep, deliberate gradient toward tightening. Maker P&L on Polymarket is spread capture plus rewards minus adverse selection: a subsidy economy, not a fee-rebate economy, and since 2026 an explicit maker-taker asymmetry in which takers pay and makers still don't.

Spreads, and the thing that eats makers

Polymarket's book is wide. Median quoted spread on the central price decile is roughly 400 bps of mid — call it 200 bps per side, an order of magnitude wider than the single-digit-basis-point effective half-spreads of liquid US equities — which the same study attributes to "longer-horizon prediction-market positions and substantially smaller market-maker capital" (Dubach 2026). Two further findings from that panel travel. There is a longshot spread premium: relative spreads blow out at extreme prices, precisely where the p(1−p) fee formulas collapse to nothing — the venue stops charging exactly where the maker most needs compensating.

The dispersion is brutal: top-stratum markets did $4.56M to $96.0M of volume in a single 28-day window, random-stratum markets as few as 100 trades. The head is a real exchange. The tail trades on reward subsidies and hope.

As for what kills makers: a binary claim's variance is p(1−p), which is the canonical loss exposure per fill, which is why both venues' fee schedules are proportional to it. Around news, an event-market maker faces the purest available form of toxic flow — the terminal value is about to jump to 0 or 1, so quoting through a resolution event is writing a straddle at stale vol. It is the same force that liquidates AMM LPs. Order-book makers survive it only because they can pull quotes.

Crypto venues carry one additional term the regulated ones don't: makers must price resolution-process risk. Oracle disputes, ambiguous wording, a whale with a stake. The cross-venue literature treats "heterogeneous resolution semantics" as a first-order pricing factor, not a footnote — which is a polite way of saying that on Polymarket, part of your spread is compensation for the possibility that the market resolves wrong.

That is a mechanism problem. It is also the subject of the next section.

Oracles, resolution, and the risk stack

Everything above is machinery for producing a price. None of it matters if the contract settles to the wrong number.

This is where prediction markets differ from every other derivative. An interest-rate swap settles against SOFR; a perp against an index of spot venues. An event contract settles against a claim about the world, and someone has to decide whether the claim is true. There is no feed for "did Ukraine agree to a minerals deal."

UMA's optimistic oracle, precisely

Polymarket does not resolve its own markets. Resolution is delegated to UMA's Optimistic Oracle, which runs on innocent-until-disputed:

End to end, a fully disputed market takes roughly 2–4 days. Trading continues throughout.

Roughly 99% of assertions since 2021 have gone undisputed, which is the design working: the DVM is a rarely-convened court, and the optimistic path is cheap. But the tail is where the money is, and the tail has a structural problem that no amount of tuning fixes.

The DVM rewards agreement with the eventual majority, not independent accuracy. A minority voter is slashed regardless of whether they were factually right. Every contested resolution is therefore a coordination game about what other voters will do, wearing the costume of a truth-finding procedure. That is not a bug that got shipped. It is what stake-weighted Schelling voting is.

What that produces, documented

June 2024 — the first public break. A market on Barron Trump's involvement with the $DJT token resolved in a way Polymarket itself rejected: the exchange said publicly that "we firmly believe that UMA got this resolution wrong" and announced a "near-term solution" to uphold market integrity (The Block); whether Yes-holders were compensated is not confirmed in the coverage. The venue and its oracle disagreed in public, and the venue had no mechanism to win.

March 2025 — the minerals whale. "Ukraine agrees to Trump mineral deal before April?" (~$7M) spiked from ~9% to 100% and resolved YES with no signed agreement in existence. Reporting traced it to a single actor voting ~5M UMA across three wallets — about 25% of votes cast, against typical active participation of 7–8M UMA — who had become a top-five UMA staker before the vote. Polymarket called it an "unprecedented" governance attack and refused refunds on the grounds that "this wasn't a market failure." Nothing in the design prevents a voter from holding a position in the market they are adjudicating.

July 2025 — the suit. "Will Zelenskyy wear a suit before July?" drew $237M+ in volume, with more than 57% of it traded after the market should already have settled. Zelenskyy appeared at the June 24 NATO summit in a black jacket-and-trousers outfit that numerous outlets and menswear commentators called a suit. An initial YES was disputed repeatedly over nine days and flipped to NO, on the grounds that a "consensus of credible reporting" had not been established. A Polymarket power user's verdict — "this isn't decentralized" — came with the observation that centralized resolution would at least be honest about who decides. This was not a whale attack. It was the mechanism working as designed on an ambiguous question with nine figures riding on it.

August 2025 — the structural response. UMA governance passed UMIP-189 on August 6, 2025, migrating Polymarket to the Managed Optimistic Oracle V2: only a whitelist of ~37 vetted addresses — Risk Labs and Polymarket employees plus proven proposers — may propose. Anyone may still dispute. The stated motivation was premature proposals losing bonds and delaying markets up to four days.

Look at what that fixes and what it doesn't. It professionalizes proposing. It leaves DVM voting-power concentration entirely untouched — reporting through late 2025 found roughly nine whales dominating dispute votes with a voting-process revamp delayed.

And then the two that came after the fix. In January 2026 a $10.5M market on a Venezuela invasion resolved No, on the reasoning that a snatch-and-extract raid capturing Maduro was not an operation "intended to establish control" of territory — a definitional question, decided after the fact, on eight figures. In April 2026 a $16.5M market on a streamer's pregnancy cycled through two disputed rounds with YES trading at 97¢ while the written market rules pointed NO. A veteran trader called the oracle "a disinformation engine." The divergence between rules text and trader Schelling expectations is now itself a trade.

The alternatives, and what each one trades away

Augur's fork. v1 escalated disputed reports through rounds of doubled REP staking, terminating in a fork: the REP token itself splits into one universe per outcome, holders migrate, and only the truthful universe is supposed to retain value. It is the most philosophically pure oracle ever designed — truth as the Schelling point of a token's own survival — and it was never the binding failure mode. v1's actual exploit was engineered invalidity: because v1 paid all outcomes equally on an Invalid resolution and traders couldn't price that risk, Binance Research documented a repeatable "design flaw attack" — create a deliberately ambiguous market, wash-trade volume to attract organic flow, post limit orders priced above the Invalid payout, collect. Case studies included a >4,000 ETH price market and ~$2M on a single 2018 midterms contract. Augur v2 made Invalid explicitly tradeable so the risk could be priced. By then the liquidity was gone.

reality.eth plus Kleros. A bond-escalation game — anyone answers with a bond, anyone overrides by doubling it — terminating in Kleros courts with randomly drawn staked jurors and appeal rounds, the stack Seer and Omen run. Juror selection is random and appeal-layered rather than plutocratic, which removes the buy-the-vote attack at lower throughput — but the jurors face the same agree-with-the-majority incentive at smaller scale. Random selection makes capture expensive. It does not make Schelling voting truth-seeking.

Data feeds. Where an objective source exists, use it and skip the humans. Azuro resolves sports from Chainlink node infrastructure, and in September 2025 Polymarket integrated Chainlink Data Streams and Automation to auto-resolve price markets, powering the 5- and 15-minute crypto contracts that had surpassed $3.4B in volume by mid-2026. A quiet but total concession: social-consensus oracles should be used only where no feed exists.

Centralized rulebooks. Kalshi resolves through its own markets team against pre-designated source agencies filed at contract self-certification — leagues and AP/ESPN for sports, CF Benchmarks for crypto, BLS/BEA/Fed for economics, NOAA for weather — with an internal review path and no external arbitration. The CFTC oversees contract terms, not individual calls. This does not eliminate controversy; it relocates it. Kalshi settled the 2025 Oscars viewership market on preliminary Nielsen numbers hours before final figures reversed the outcome, and stood by it; it paid out UFC 327 on the initially announced majority-decision result, later corrected to a majority draw, under a published rule that post-expiration revisions don't count. It is also facing a class action over a "death carveout" that allegedly blocked full payouts on a $54M market tied to Khamenei's removal from office (NY Post).

The trade is symmetric and should be stated as such. Token-voted oracles fail through plutocratic capture and Schelling drift, with no appeal. Centralized rulebooks fail through operational error and fine-print asymmetry, with courts as the appeal layer. Pick your failure mode; you do not get to pick none.

Ambiguity is the attack surface. Every blow-up above started as an underspecified question, not a technical exploit. What counts as "agreeing" to a deal. What garment is a suit. Whether a raid establishes control. Whether death is removal from office. The oracle is downstream of the drafting, and two second-order effects follow. People trade the resolution, not the event — in the Zelenskyy market most volume changed hands after the underlying fact had occurred, capital betting on the oracle's behavior rather than on the world. And rules text and Schelling expectation diverge, with the market pricing the gap: YES at 97¢ against a rules text pointing NO is a market betting that voters will override the words. Every industry mitigation — frozen clarifications, proposer whitelists, tradeable Invalid outcomes, pre-registered sources — shrinks the ambiguity surface without touching the deeper issue, which is that natural language about the future is irreducibly vague and money makes people notice.

The settlement stack underneath

Two risks live below the oracle and are easy to forget.

Chain finality. Polymarket's CLOB settles on Polygon PoS, which historically had probabilistic finality deep enough to matter: a 157-block reorg — roughly five minutes of activity — at block 39,599,624 in February 2023, caused by a bad-merkle-root bug that partitioned block producers, and occurring after a hard fork that had cut sprint length from 64 to 16 blocks specifically to reduce reorgs. For an exchange, that means fills, resolutions and withdrawals inside a reorged window were only ever probabilistically settled. The risk is now much smaller: the Heimdall v2 hard fork on July 10, 2025 cut deterministic finality to ~5 seconds from ~1–2 minutes, with third-party technical coverage reporting reorg depth now capped at two blocks (Polygon's own blog claims only a qualitative reduction). What remains is consensus-bug risk in the finality gadget — materially better, not zero, and the 2023 event was itself a bug in supposedly reorg-reducing code.

Collateral. From launch, Polymarket's collateral was USDC.e — Ethereum USDC locked in the Polygon PoS bridge, issued by the bridge contract rather than Circle. A bridge exploit would have impaired the backing of every open position simultaneously. In 2026 Polymarket migrated to pUSD, backed 1:1 by native USDC in reserve, in partnership with Circle. Read that correctly: it swaps bridge-contract risk for issuer risk — Circle solvency, Circle freeze authority. A deliberate move toward the regulated-custody model, made by a company re-entering the US.

Integrity: the three documented failures

Wash trading. A Columbia Business School study published November 2025 estimated that ~25% of Polymarket's ~$18B lifetime volume — about $4.5B — was wash trading — lifetime volume measured on the study's own on-chain counting basis, closer to the Dune series than to press-reported figures — flagging suspicious patterns in 14% of 1.26M active wallets. The time profile is the tell: nearly 60% of weekly volume in December 2024, under 5% by May 2025, back to ~20% by October 2025; by category, 45% of sports volume against 17% elections, 12% politics and 3% crypto. The December 2024 peak coincides with airdrop-farming expectations, which makes volume partly endogenous to token speculation. Any cross-venue volume comparison here needs a wash-adjusted discount — and volume must never be confused with fee revenue.

Insider trading. Hours before Trump's October 2025 pardon of CZ, an account called "bigwinner01" bought YES on "Will Trump pardon CZ in 2025?" and cleared roughly $56.5K; Coffeezilla linked the wallet to a trader who had previously made a nine-figure sum on Hyperliquid shorting ahead of market-moving Trump announcements. Onchain transparency makes these patterns visible — tracking Polymarket insiders is now a business — but for years no statute clearly prohibited trading a political event market on material non-public information the way securities law does.

That gap is now being tested. On April 23, 2026 the DOJ unsealed an indictment against Gannon Ken Van Dyke, an active-duty Army Special Forces Master Sergeant, alleging he used classified knowledge of the Venezuela operation to earn roughly $409,881 on Polymarket — charged under the Commodity Exchange Act alongside commodities fraud, wire fraud and money laundering, with parallel CFTC civil charges (NYU Compliance & Enforcement). The government's theory is that event contracts are swaps, and that trading them on misappropriated government information is insider trading. Newly filed, not settled law. The direction of travel is unambiguous — and note the symmetry: the same operation whose description decided an eight-figure market's resolution also produced the first criminal information-asymmetry case in the asset class.

Settlement manipulation. The most elegant attack of the cycle. A Stanford study of ~16,000 Polymarket 5-minute BTC up/down contracts (February–April 2026) found Binance order flow in the final 10 seconds before expiry running ~50% above pre-launch baseline, with 821 wallets capturing ~$8.2M as consistent beneficiaries. The Chainlink settlement price aggregates spot exchanges, so nudging thin spot books for ten seconds decides the binary. The 15-minute variant showed much weaker signatures — contract horizon is itself a manipulation-resistance parameter.

This is the exact analogue of the UMA whale vote, one layer down. Whoever can move the oracle's input more cheaply than the contract's payout wins. Replacing a social oracle with a data feed does not eliminate that structure. It changes who has to be bribed.

Documented instances and mitigation state, as of August 2026:

Layer Risk Documented instance Mitigation state
Oracle — governance Stake-weighted vote capture UMA whale, $7M minerals market, ~25% of vote MOOV2 proposer whitelist (UMIP-189); DVM concentration unresolved, ~9 whales dominate votes
Oracle — epistemic Schelling consensus ≠ truth; minority slashing herds voters Zelenskyy suit, $237M+; clavicular market, 2026 Unsolved for subjective questions; objective markets moved to Chainlink
Oracle — input Manipulating the feed the oracle reads Stanford: $8.2M via last-10-second Binance flow Longer horizons; multi-source trimmed benchmarks
Oracle — centralized ops Human error, fine-print asymmetry Kalshi Oscars; UFC 327 correction; Khamenei carveout suit Rulebook precision; litigation as backstop
Contract design Ambiguous wording; engineered invalidity Augur v1 invalid-market scam Tradeable Invalid (v2); frozen clarifications; pre-registered sources
Settlement chain Deep reorgs unwinding "final" fills Polygon 157-block reorg, Feb 2023 Heimdall v2: ~5s deterministic finality
Collateral Bridge exploit; issuer/freeze risk USDC.e bridge dependency pUSD / native USDC with Circle — swaps bridge risk for issuer risk
Market integrity Wash trading inflating volume Columbia: ~25% of $18B; ~60% weekly peak Dec 2024 Incentive removal; surveillance; wash-adjusted metrics
Market integrity Insider trading CZ-pardon "bigwinner01"; Van Dyke CEA indictment, Apr 2026 Onchain forensics; first CEA prosecution pending
Meta Trading the resolution process itself Zelenskyy: 57% of volume post-event Faster, more objective resolution; shorter dispute windows

The 2024–2026 record supports one summary claim: resolution risk, not liquidity and not regulation, is the binding constraint on prediction-market credibility. And the industry's revealed preference is segmentation by question type — data feeds for objective outcomes, whitelisted optimistic oracles for semi-subjective ones, centralized rulebooks with courts as appeal at the regulated venues. Fully permissionless Schelling voting survives only in the residual category where nothing else works.

Which happens to be the category the whole field was originally built to serve.

Where the volume actually is

Start with the number that gets quoted least often, because it makes everything after it legible.

From 2020 through 2022, Polymarket ran an automated market maker on Polygon. Across those three years, decoded on-chain volume was $48.9M (2020), $121.3M (2021), $69.0M (2022) (Dune query 8212838). Cumulative lifetime volume before the order book: roughly $240M. On press-reported figures — $12.9B in July 2026, over $400M in an average day — Polymarket now clears more than that lifetime AMM total in a single day.

That is the shape of the thing. Not a sector that grew steadily and compounded, but one that sat near zero for five years and then went vertical twice — once on an election, and once, much larger, on sports.

Here is the central-limit-order-book era, one-sided (taker-leg only, matching Polymarket's own convention), from Dune query 8212792:

Month Volume (one-sided) Distinct traders Trades Avg trade size
2023 (full year) $35.7M ~1–2k/mo — —
2024-01 $25.5M 4,095 45k ~$567
2024-06 $59.0M 29,429 279k ~$211
2024-09 $295.2M 90,035 1.53M ~$193
2024-10 $1,355.9M 235,277 5.98M ~$227
2024-11 (election) $1,419.2M 304,327 6.55M ~$217
2024-12 $1,155.1M 358,850 9.71M ~$119
2025-01 $711.9M 462,598 5.63M ~$126
2025-03 (trough) $472.5M — — —
2025-08 $522.2M 227,415 4.82M ~$108
2025-10 $1,702.2M 477,845 12.1M ~$141
2025-12 $2,378.0M 519,131 30.2M ~$79
2026-01 $3,528.0M 648,026 52.0M ~$68
2026-03 (peak) $4,984.1M 784,436 115.4M ~$43
2026-06 $4,145.2M 666,374 68.2M ~$61
2026-07 $2,925.8M 544,818 51.9M ~$56

Yearly, on the same basis: 2023 ≈ $0.04B → 2024 ≈ $4.88B → 2025 ≈ $11.27B → 2026 through July ≈ $26.8B.

Two things are worth pulling out of that table. The first is the post-election trough: from $1.42B in November 2024 down to $472.5M in March 2025, a ~67% contraction that held through August 2025. The election was not a floor. It was a spike that decayed for six months, and anyone who marked 2024 as the new baseline was wrong by a factor of three for most of a year.

The second is the collapse in average trade size — from ~$217 in November 2024 to ~$43 in March 2026. Trade count grew ~18x while the trader base grew ~2.6x. That is not more people betting more. That is the same people clicking far more often, on much smaller notional, in markets that resolve in minutes. The product changed underneath the volume line.

A caveat to carry through the rest of this section. Press-reported Polymarket figures are roughly double the Dune one-sided series — The Block puts offshore Polymarket at $10.26B in June 2026 against Dune's $4.15B. Counting both legs closes most of the gap; the residual is plausibly the parlay product and flow outside the queried contracts, but the precise cause is unverified. Treat the Dune series as a verified floor and the reported series as the headline, and do not mix them in the same ratio.

The flippening, and what actually caused it

Kalshi's trajectory, from its own regulatory market reports mirrored into Dune (query 8212811; $1-notional contracts, so contracts ≈ dollars):

Period Kalshi Polymarket (Dune, one-sided) Polymarket (reported)
2023 FY $0.18B $0.04B —
2024 FY $2.01B $4.88B —
2024-11 $1.27B $1.42B —
2025 FY $24.06B $11.27B ~$20B implied
2026-04 $14.81B $4.21B $9B intl + $1.3B US
2026-06 $33.0B (Dune) / $31.5B (The Block) $4.15B $10.26B intl + $3.04B US
2026-07 $41.0B (Dune) / $37.7B (Covers) $2.93B $7.9B intl + $5.0B US

Kalshi grew 12x in 2025 while Polymarket roughly doubled. By July 2026 the sector printed $50.6B combined, an all-time high, with Kalshi at ~74.5% of it — the immediate driver being the World Cup, where Kalshi's tournament-winner market took $832M+, the Spain–Argentina final roughly $1.9B, and Polymarket's tournament-winner contract about $4B (Covers).

For scale: legal US sportsbooks averaged about $14B/month in handle during 2025. Kalshi alone now clears more than twice that in notional — though exchange volume counts secondary churn a sportsbook's handle does not.

The rotation

The category mix tells the real story. Polymarket, by Messari's tagging (Dune query 6538252, two-sided, share of monthly volume):

Month Politics Sports Crypto World affairs Other
2024-03 73.8% 4.3% 5.3% 5.4% 11.2%
2024-10 86.6% 9.9% 1.8% — ~1.7%
2024-11 52.3% 34.0% 4.5% 3.7% 5.5%
2025-02 29.5% 42.9% 15.1% 7.4% 5.1%
2026-01 22.1% 44.4% 22.0% 6.7% 4.8%
2026-07 10.7% 69.3% 15.6% 3.1% 1.3%

Sports passed politics on Polymarket in December 2024 — one month after the election — and never gave the lead back.

Kalshi's rotation was sharper. Pre-2024 it was a financials venue (June 2023: 85% financials/economics). October–November 2024 it was ~90% politics and elections. Sports launched in December 2024 at 96k contracts; by March 2025 sports was already 77.9% of volume. As of July 2026: sports 46.2%, multi-leg "Exotics" 36.5%, crypto 15.4% — and politics 0.14% (Dune query 8212811). Fifty-nine million dollars out of forty-one billion. On the venue that sued the federal government for the right to list election contracts, politics is now a rounding error.

Pew's cumulative July 2024–May 2026 splits confirm the divergence: Kalshi 80% sports / 7% crypto / 4% politics; Polymarket 39% sports / 32% politics / 20% crypto. Polymarket retains a real politics franchise. Kalshi has become a sports and parlay exchange with a crypto desk attached.

Which makes the 2026 midterm ramp the most over-anticipated event in the sector. Polymarket politics ran ~$1.5–2.1B/month (two-sided) in early 2026 at roughly 21–25% of volume, and Kalshi politics is under 1%. The midterms will produce a spike, but on current mix it lands on a base that is 80% sports — a bump, not a regime.

Geography

Polymarket has been US-blocked since the January 2022 CFTC settlement and operated offshore for four years, with acknowledged VPN leakage. Country blocks then snowballed — not from securities regulators, but from gambling regulators. France banned financial transactions in November 2024, formally reclassified prediction markets as illegal gambling in February 2026, and ordered ISP-level blocking on July 16, 2026. Belgium blacklisted it under the Gambling Act. Poland added it to the prohibited-domains registry in January 2025. Portugal's SRIJ moved in January 2026 after more than €103M traded on the Portuguese presidential election; the Netherlands followed in February 2026; Brazil's National Monetary Council banned event derivatives outright in April 2026. Total: 40+ restricted jurisdictions, most of them close-only.

Against that, the US re-entry described earlier — the QCX acquisition, the amended designation, the December 2025 launch, full iOS in May 2026 — and the volume ramp it produced: $1.3B (April) → $1.77B (May) → $3.04B (June) → $5.0B (July). That last figure is 39% of Polymarket's July total, and the only line in the sector growing faster than Kalshi's.

There is no credible country-level volume split for offshore Polymarket. It doesn't publish one, and traffic trackers are not volume. The geographic story here is regulatory, not measured.

Velocity, not stock

One last number, because it disciplines every other number in this section. Kalshi's month-end open interest (Dune query 8212815) was 453M contracts in November 2024 against $1.27B of volume that month — an OI-to-volume ratio of roughly 36%. By July 2026, OI was 792M against $41.0B of volume: roughly 2%. (The Block independently reports Kalshi ending July at ~$788M, within 0.5% of the Dune month-end figure.)

Capital at risk grew well under 2x on end-July open interest — roughly 2.7x at the June 2026 intra-period peak. Reported volume grew about 30x. The gap is entirely velocity — hour-long sports games and fifteen-minute crypto candles churning the same dollars dozens of times. A venue with $800M of open interest doing $40B of monthly volume is not a market where a lot of money is committed to beliefs. It is a market where a modest amount of money is very busy.

Take rates and the business of belief

For five years, the largest prediction market on earth charged nothing.

That is the fact to hold onto. Polymarket's on-chain fee field reads approximately zero from launch through December 2025 (Dune query 8212792; a few small AMM-era experiments in 2023 aside). Then in January 2026 it turned on taker fees for crypto markets, and by March 30, 2026 the schedule covered nearly every category. Five years of land grab, then monetization, in ninety days.

Venue Fee structure Effective take As of
Polymarket Taker only. Fee = shares × rate × p × (1−p). Rates: 0.04 politics/finance/tech, 0.05 sports/economics/culture/weather, 0.07 crypto; geopolitics free. Makers pay nothing and earn 15–25% rebate pools paid daily (Polymarket fee docs via StartPolymarket). July 2026: sports raised 0.03→0.05, crypto trimmed 0.072→0.07. No withdrawal fees. ~0.9–1.3% of one-sided volume (May–Jul 2026 on-chain fees: $29.1M, $40.6M, $37.3M) 2026-08
Kalshi Taker: ceil(0.07 × C × P × (1−P)), max 1.75¢/contract at P=50¢. Maker fees on select series at a 0.0175 coefficient; most series maker-free (pm.wiki). ~1.1% — $263.5M of 2025 fees on $24.06B volume 2025 FY
PredictIt 10% of profits on winning trades + 5% withdrawal fee. Position cap raised $850 → $3,500 after the July 2025 CFTC settlement. n/a (levied on profits, not volume) — highest in the category 2025
Betfair Exchange Commission on net market winnings: 5% base in the UK (2% on some markets), 6–10% Market Base Rate in Australia. The Premium Charge was replaced on January 6, 2025 by the Expert Fee: 0% below £25k rolling gross profit, 20% from £25k–£100k, 40% above £100k. 2–10% of winnings; materially higher for sharps 2026
Azuro Margin embedded in AMM odds by data providers; split between LPs, frontends, providers. ~4.28% of betting volume (DappRadar, mid-2024) 2024
US sportsbooks Vig embedded in the line. Sharp two-way markets 2–4%; props and parlays well into double digits (BettorEdge). 2025 actual: ~$165B handle on ~$16B revenue ≈ 9.7% hold. ~9.7% blended 2025

Two things follow.

First: the exchanges converged on the same fee curve. Both charge proportional to p × (1−p) — maximum at even odds, approaching zero in the tails. The quadratic taxes the flow that looks most like a coin flip and least like information, while leaving longshot hedging essentially free: a 3¢ tail contract on some geopolitical rupture costs almost nothing to trade, a 50/50 NFL moneyline pays the maximum. The fee schedule is an implicit statement about which trades the venue thinks it is providing a service for.

Second: the effective take is roughly 1%, and that is the whole competitive story. Against a ~9.7% blended sportsbook hold, an exchange charging 1% can quote materially better implied odds on the same event and still run a business. Kalshi's 2025 fee revenue of $263.5M on $24.06B of volume works out to 1.095%. Sacra's revenue series — ~$25M annualized in December 2024, ~$735M in December 2025, ~$3.5B annualized in June 2026 — implies about $292M/month on roughly $30B, or ~1.0%. Two independent constructions, the same answer.

Polymarket's revenue has no company disclosure, so it has to be inferred. On-chain taker fees of $29.1M, $40.6M and $37.3M in May, June and July 2026 imply a gross run-rate around $400–450M/year on Dune-visible volume, before maker rebates. (The March–April readings of $181.2M and $284.2M would imply take rates of 3.5–6.7%, which is not plausible against the published schedule and looks like a decode artifact around the V2 migration. Unverified; use the May–July figures.)

So how did a 0%-fee book monetize before January 2026? It didn't. Polymarket ran on venture and strategic capital — most conspicuously ICE's commitment of up to $2B at roughly $9B post-money. The business model during the zero-fee era was market data, distribution, and a claim on a future in which the fee could be switched on. In January 2026 it was.

The comparison that matters most is not exchange-versus-exchange. It is exchange-versus-sportsbook, and it is nine-to-one. Everything in the next section follows from that gap.

The regulatory arbitrage

Two people can take opposite sides of "will the Chiefs win," settle in dollars, and the entire legal and economic treatment of that transaction turns on whether the controlling description is wager or swap.

That is the arbitrage. It is not subtle, and its economics are concrete.

State gaming taxes. Sportsbooks pay gross-gaming-revenue taxes up to 51% in New York, and Illinois layers a 20–40% sliding scale on top of a $0.25–$0.50 per-wager tax. In 2025 the industry paid roughly $3.7B in state taxes on ~$16B of revenue — New York alone $1.3B. A CFTC-designated contract market pays none of it. Just corporate income tax, because the activity is trading.

Federal excise. Wagers carry a 0.25% excise on handle plus a $50 annual occupational tax under IRC §§ 4401 and 4411. Event contracts on a DCM are not "wagers" under that regime. On $165B of handle-equivalent flow, that line alone is ~$400M/year.

Licensing geography. A sportsbook needs a license, a fee, and a compliance stack in each of thirty-odd states. A DCM operates one national market in all fifty — including California and Texas, where sports betting remains illegal and where a large share of the incremental demand lives.

Customer price. ~1% exchange fee against ~9.7% embedded vig.

Bettor-level tax treatment. If event-contract payouts are investment income rather than gambling winnings, traders escape the post-OBBBA cap on gambling-loss deductions (Fortune). IRS guidance has not resolved this.

The clearest evidence that an arbitrage is real is that the people it is run against cross over to run it themselves. DraftKings acquired Railbird Exchange, a designated contract market, in October 2025 for up to $250M and launched DraftKings Predictions in December across 38 states — crucially including states where its own sportsbook is not licensed. FanDuel partnered with CME to launch FanDuel Predicts on December 22, 2025, with sports contracts offered only in states without legal online sportsbooks. Underdog bought Aristotle Exchange. Robinhood, which routes to Kalshi and drove more than half of Kalshi's volume, bought ~90% of MIAXdx with Susquehanna and renamed it Rothera. And Sporttrade did the most legible thing of all: it surrendered its state gambling licences to apply for DCM and DCO status, closing its New Jersey sportsbook on May 25, 2026 and the remaining four states on June 25. A regulated sportsbook voluntarily became a commodities-exchange applicant. That is what a nine-point spread does to corporate strategy.

The offshore arbitrage, and its closing

Polymarket's 2022–2025 posture was a second, distinct arbitrage: non-custodial, USDC-settled, no KYC, offshore, with widely reported VPN leakage. DOJ and CFTC investigations into that leakage were closed with declinations in July 2025.

This arbitrage is closing, and the mechanism is worth noting: it is being closed by gambling regulators abroad and by the operator itself at home. Jurisdiction by jurisdiction the offshore surface shrinks, while the same company spends $112M to buy its way inside the US perimeter. The 40+ restricted jurisdictions and the $5.0B US July are the same story told from both ends: the crypto-native offshore venue is converting into a regulated American exchange with an offshore legacy book attached.

The UK shows what a world without the seam looks like. The Gambling Commission treats prediction-market operators as betting intermediaries requiring a UKGC licence; Betfair and Smarkets run structurally identical order books, legally, under gambling licences. Britain folded the exchange model into gambling regulation twenty years ago. There is nothing to arbitrage, so nobody is arbitraging.

Durable, or closing?

Durable. Election contracts are lawful under an unappealed district ruling — the CFTC voluntarily dismissed its appeal on May 5, 2025. On April 6, 2026 the Third Circuit affirmed 2–1 in KalshiEX v. Flaherty that sports event contracts are swaps under exclusive CFTC jurisdiction, the first appellate holding to that effect. Chairman Michael Selig withdrew the Biden-era proposed ban and is writing rules instead. And ICE, CME, DraftKings, FanDuel and Robinhood all now carry balance-sheet exposure to the DCM model — which changes the political economy of any rollback far more than any court opinion.

Closing at the edges, in four places.

The circuit split. Nevada's Chief Judge Andrew Gordon dissolved Kalshi's injunction on November 24, 2025, holding the contracts fall outside exclusive CFTC jurisdiction; Maryland denied an injunction on similar reasoning; a Massachusetts state judge enjoined Kalshi in January 2026; the Fourth and Ninth Circuits have both sounded skeptical at argument. Third Circuit for, several courts against, 30+ state AGs pushing the other way. This ends at the Supreme Court.

Tribal gaming. The strongest legal theory against preemption, because IGRA is a federal statute rather than state law — so the field-preemption argument that beats New Jersey does not obviously beat a tribe. Tribes lost a preliminary injunction in November 2025 but the Ninth Circuit panel was pointed: "sounds like a bet." Gaming exclusivity anchors revenue-sharing compacts in California and Florida; this is existential for the counterparty, which means it will be litigated to exhaustion.

Taxation replacing prohibition. North Carolina taxes prediction markets at 6% while raising its sportsbook rate to 23%; Kentucky enacted a 14.25% excise on prediction-market transaction fees, both effective January 1, 2027. Note the base: Kentucky taxes fee revenue, not handle or GGR. At a ~1% take, 14.25% of fees is roughly 0.14% of notional against a sportsbook paying tens of percent of a 9.7% margin. States are monetizing what they cannot prohibit, and the differential survives the monetization almost intact.

Product scope. The June 12, 2026 NPRM proposes a three-step public-interest test. Sports and elections broadly survive; player injuries, officiating decisions, discrete in-game actions, and pre-collegiate sports are flagged as likely contrary to the public interest. This is the settlement: concede the most gambling-like products to buy legitimacy for the rest.

The wildcard is conduct. A WSJ investigation found Polymarket paid college-age influencers to produce 1,100+ undisclosed videos featuring fabricated bets; senators demanded a probe, and reporting indicates the CFTC already had an "ongoing and extensive" investigation. Federal accommodation of an asset class is not immunity for behavior within it. The fastest way to lose a regulatory settlement is to embarrass the regulator who granted it.

Futarchy, or markets that decide

There is a branch of this field that never became a business, and it is the branch that would have justified the epistemics. If markets aggregate information better than experts, why stop at forecasting? Why not let them decide?

That is Hanson's "Shall We Vote on Values, But Bet on Beliefs?" (circulated ~2000, later published in the Journal of Political Philosophy). Democracy chooses the objective — a welfare metric, selected by vote. Markets choose the means. A policy is adopted if and only if conditional markets price expected welfare higher under adoption than under rejection.

The load-bearing component is the decision market, and its difference from a prediction market is one sentence long. A prediction market prices P(outcome). A decision market prices E[metric | A] against E[metric | B], and the comparison of those two prices triggers the decision. Trades in the branch that doesn't happen get reverted, so each conditional price is — in theory — an unbiased estimate of the metric under that counterfactual (Hanson, "Futarchy Details"). That one change is where all the difficulty lives. The market's output now causes the world it is pricing.

Buterin's 2014 essay remains the sharpest crypto-native treatment, and it is the critique that stuck. He grants futarchy real virtues — it replaces negligible vote influence with financial incentive, compounds influence toward accurate analysts, depersonalizes governance — then lists six objections: manipulation, which counter-trading neutralizes only given depth thin markets don't have; self-reference, since "markets are 'self-referential' — they consist largely of people buying because they see others buying"; signal-to-noise, because most individual policies move a global welfare metric by less than its background noise; value complexity, because "human values are complex, and it is hard to compress them into one numerical metric," so the metric fight relocates the policy fight; negative-sum participation for informed-but-small traders; and metric gaming, where lobbying migrates to the definition of the measurement. His conclusion — coarse decisions over granular, DAOs over states — has held up for a decade, and his 2024 info-finance essay generalizes rather than retracts it.

The field record is thin. GnosisDAO launched in November 2020 branded on futarchy, said it was "not going all in on futarchy on day 1," and settled permanently into a GNO-weighted Snapshot vote. Futarchy as marketing, voting as mechanism.

MetaDAO is the exception: launched in late 2023 by the pseudonymous Proph3t, the first system where markets actually execute. Per Helius, conditional vaults act as escrow — deposit META or USDC, receive pass-conditional and fail-conditional tokens; losing-branch tokens redeem back to the deposit, winning-branch tokens redeem for the real asset. That is how you simulate reverting a counterfactual on a chain that cannot revert. Each proposal spawns two markets, and an autocrat program checks the TWAP gap between them at finalization and, if satisfied, executes the proposal's Solana instruction (code; v1 required a 5% gap on an OpenBook V2 CLOB, and Solana Compass puts the current conditional-AMM threshold around 3%). TWAP is the anti-manipulation primitive: you must move the price and hold it. Proposals 6, 7 and 8 — including a Pantera Capital offer to buy $50k of META at 100 USDC/META, and a mint proposal whose author committed $250k+ — all failed against market resistance. Small sample. Real evidence.

Paradigm led a $2.2M round in August 2024. Then MetaDAO found product-market fit sideways: its center of gravity moved from governing existing DAOs to an ICO launchpad for "ownership coins", with treasury, IP and mint authority under futarchic control. Per Shoal, the first was MtnDAO on April 9, 2025; by late 2025 the platform had run 8 ICOs raising ~$25.6M, the flagship being Umbra's $154.9M in USDC commitments against a $3M hard cap from 10,518 wallets (The Block). Its own economics are volatile: the Q1 2026 tokenholder report put protocol revenue at $556k, down 78% from Q4 2025's $2.52M, treasury at $12.24M. The quietly important datapoint is the liquidation of Ranger Finance, a launchpad project whose futarchic governance wound it down and returned ~$5.04M to RNGR holders. Market governance executing an orderly shutdown with capital return is the maneuver token-voting DAOs are notoriously incapable of.

Two foundation-run pilots supply the only controlled evidence. Optimism's Season 7 Futarchy Grants Contest, operated by Butter, ran 84 days to June 12, 2025: play-money conditional markets ranked ~22 candidate protocols, with futarchy and the Grants Council each selecting five projects for 100k OP grants against a Superchain TVL growth metric — 430 active forecasters after filtering 4,122 suspected bots and sybils, 5,898 trades, 88.6% first-time governance participants (preliminary findings). Both halves of the result matter. Absolute accuracy was terrible: markets predicted ~$239M of aggregate TVL impact against ~$31M realized, an ~8x overestimate attributed to play money, liquidity anchoring and strategic trading to steer selection. But relative selection beat the committee — futarchy's picks generated ~$32.5M more TVL growth, driven by funding Balancer/Beets, which the Council ranked sixth. Both mechanisms picked Rocket Pool, which delivered $0. Futarchy found the best performer and the worst.

The Uniswap Foundation's parallel program with Butter used real money at smaller scale — a $900k program structured as a $100k first round and two $400k grants in a second. CFM #1 ran a July 2025 window over four lending protocols against a Unichain TVL KPI, needed $20k of Merkl liquidity mining to make the markets tradeable at all, and picked Morpho on ~$19–20k of volume against Euler's $5.6k. Its honesty is its best feature: the raw funded-versus-counterfactual delta was $96.01M against $67.50M of TVL, but a concurrent ETH rally swamped protocol-specific effects and share-based adjustment collapses the estimate to ~$7.13 per grant dollar. It also names the structural attack surface — push the funded branch up, or sabotage the not-funded branch down, because the decision hinges on a delta.

The scorecard, honestly, as of August 2026:

Claim Evidence for Evidence against
Resists manipulation MetaDAO Proposals 6–8 failed against TWAP + counter-traders Only at small scale; Butter names branch-sabotage as unaddressed
Beats committees Optimism: +$32.5M TVL over 84 days n=3 unique picks per arm; higher variance both ways
Produces good forecasts — Optimism: ~$239M predicted vs ~$31M realized
Works as binding governance MetaDAO; Ranger's orderly liquidation and $5.04M return Gnosis retreated to Snapshot; no major DeFi protocol governs by markets
Sustains a business 8 ICOs, ~$25.6M raised; Umbra $154.9M committed vs $3M cap MetaDAO revenue −78% QoQ into Q1 2026; pivoted from governance to launchpad

None of the reasons decision markets are harder than event markets is an engineering problem. Conditionality is never audited: the losing branch is reverted, so its price is never checked against reality, and you can run futarchy for a decade without learning whether your counterfactual prices were any good. The decision then rides on a small delta between two noisy prices, which is why MetaDAO requires a 3–5% TWAP gap — sub-threshold deltas are noise wearing a decision's clothes. Liquidity is thin by default, because event markets borrow it from entertainment demand while decision markets price obscure counterfactuals with capped upside: Butter needed $20k of paid liquidity mining to run one $100k grant decision, which is the business-model problem in miniature. And self-reference is the mechanism, not a bug you can patch — because the output causes the decision, everyone with a stake has an incentive to trade for influence rather than accuracy. An event market pays you only for being right. A decision market also pays you for winning, and when those conflict the second has the more concentrated constituency. Then the metric: Rocket Pool scored $0 on a measure that both futarchy and its control arm misjudged. When the mechanism and its benchmark fail identically, the mechanism isn't the binding constraint. Value complexity, delivered as data.

Through mid-2026 the honest reading is one production system that survived by pivoting from replacing votes to pricing launches, two pilots with encouraging relative results and damning absolute accuracy, and a converging view — anticipated in 2014, restated in 2024 — that markets work best as high-powered inputs to coarse, high-stakes, measurable decisions rather than as a replacement for governance. A smaller claim than Hanson's. Also the first with a track record.

Where this goes

Three things happened at once, and the sector's future depends on which of them was load-bearing.

The mechanisms matured. Twenty years of market design established that fixed-parameter scoring rules bleed liquidity and that constant-product AMMs on binary outcomes hand the LP a near-certain total loss by expiry. Both leading venues now run order books with professional makers, symmetric quadratic fees, and rebates that pay for two-sided quoting — which is why a coin-flip NFL market can be quoted at a 1% round-trip instead of a 9.7% one. The mechanism work is real, it is finished for the liquid case, and it is the least appreciated input to the volume charts above.

The oracle did not mature at the same rate. Resolution remains the load-bearing assumption in every one of these numbers, and it is the piece that has been demonstrably attacked. Note the asymmetry: the more the mix rotates into sports and crypto candles, the less this matters, because a football score and a BTC print are unambiguous. Oracle risk concentrates precisely in the markets prediction markets are supposed to be for — contested, interpretable, socially consequential questions. The category the sector monetizes is the category where resolution is trivial. The category that justifies the sector is the category where resolution is hard.

And the regulation normalized. That is the input everyone underweights, because it does not look like technology. But the single largest determinant of 2026 volume was not a better matching engine. It was a court in October 2024 declining to stay an injunction, and a new chairman in January 2026 withdrawing the rulemaking instead of finishing it.

Put those together and the honest read is this: prediction markets won the legitimacy fight and then spent the winnings on sports. That is not a criticism — it is the only revenue base that could fund the infrastructure. But it does mean the sector's self-description and its P&L have drifted badly apart. A venue running 0.14% politics is not primarily an information aggregator. It is a low-vig exchange for binary sports outcomes that also happens to publish an election price.

What would change that? Not more sports volume. Three things would: open interest growing faster than volume, which would mean capital committed to positions rather than churned through them; long-dated markets on questions with genuine decision value, which requires an oracle nobody can buy and a term structure nobody currently prices; and institutional consumption of the prices themselves, which is what ICE's distribution deal is a bet on. The futarchy record is the closest thing to a controlled test of that last claim, and it shows decision-grade markets need subsidy to exist at all — exactly what you would expect of a public good.

The near-term is easier to call. The midterms produce a spike into a base where politics is a tenth of Polymarket and nothing on Kalshi, so expect the spike to look smaller than 2024 in relative terms and larger in absolute ones. The Supreme Court eventually takes the preemption question, and the entire sector's national addressable market rides on the answer. States keep taxing rather than banning, at rates that leave the differential intact. And the offshore book keeps shrinking as the onshore book grows — the arbitrage that built Polymarket is the one it is now dismantling from the inside.

The structural bet embedded in every valuation in this sector is a single sentence: that swap on a designated contract market continues to beat wager in a casino as the controlling legal description of the same economic act. Not a claim about mechanism design. Not a claim about information aggregation. A claim about which statute applies.

Everything else is downstream of that.

Sources and methodology

This report was compiled from primary sources, Dune data, and cross-verified reporting.

Primary sources include the academic literature (Hayek 1945; Wolfers–Zitzewitz 2004 and the companion NBER paper; Forsythe et al. 1992; Berg–Nelson–Rietz 2008; Rhode–Strumpf 2004; Rothschild–Sethi; Hanson–Oprea 2009; Cowgill–Zitzewitz 2015; Hanson's LMSR and futarchy papers), protocol documentation and source code (Gnosis Conditional Tokens, Polymarket's CTF Exchange and NegRiskAdapter repositories, UMA's DVM 2.0 docs, Azuro's docs, MetaDAO's programs), regulator and court records (CFTC releases, no-action letters and the June 2026 NPRM; D.C. Circuit, Third Circuit and district-court filings), and company disclosures (ICE investor relations, Robinhood's SEC-filed closing documents and earnings coverage, published fee schedules).

Quantitative series come from the Dune queries cited inline: Polymarket AMM-era volume (8212838), CLOB-era volume and fees (8212792), category mix by Messari tagging (6538252), Kalshi volume from mirrored regulatory market reports (8212811), and Kalshi open interest (8212815). Two counting conventions circulate in this sector — one-sided (taker-leg) and two-sided — and they differ by roughly a factor of two. They are labelled at every point of use and are never combined in a single ratio. Where press-reported figures diverge from on-chain figures, both are shown and the on-chain series is treated as a verified floor rather than the headline. Trading volume and fee revenue are reported as separate quantities throughout and are never substituted for one another.

Every factual claim was independently re-verified against a second source where one existed. Where a figure could be traced to only one source, or where the underlying data appeared to be a decode artifact rather than a real reading, it is flagged as such in the text rather than smoothed over. Claims that could not be verified were dropped. All figures and market states are current as of August 2026; where a table reports a point-in-time state, the as-of date is given with the table.