Prediction Markets

Explanation

A prediction market posts a binary question about a future event, such as whether a specific outcome will occur, and users buy yes or no contracts against each other. The market price, the fraction of the most recent trades that went to yes, is the implied probability, and it updates continuously as money flows in and news breaks. The mechanism's claim to accuracy rests on a few properties working together. Incentive alignment means a participant only bets if they genuinely believe they have an edge, so uninformed bettors lose money systematically and tend to exit while informed bettors make money and stay, and over time the market selects for real signal rather than loud opinion. Information aggregation means the market draws in whoever actually knows the most about a given question rather than sampling a random population the way a poll does. Continuous updating means a market behaves like a live feed rather than a single snapshot, often moving before mainstream coverage has converged on a narrative, with the speed of updating proportional to the size of the underlying liquidity pool.¹

The sharpest distinction is that outcome probability is not the same thing as vote share. A market showing a candidate at seventy percent means that if the same election were run a thousand times under current conditions, that candidate would win roughly seven hundred of them, and the candidate could win by a single point and the seventy percent estimate would still have been correct. A poll measures what share of the population intends to vote for each option; a market measures the probability the outcome goes one way or the other, which is the quantity people actually want to know.¹

Why it matters

Prediction markets became a mainstream forecasting tool after outperforming most conventional polling on a high-stakes election where pollsters called the race too close to call while market-implied odds gave a clear lead.¹ That result matters beyond any one election, because there is a large class of questions, elections, regulatory decisions, policy outcomes, geopolitical events, where a public probability estimate has real economic value but no good mechanism for producing one existed at scale before markets like this. Polls, pundits, and news coverage are substitutes for the same underlying need, and on the available evidence they are worse substitutes.

Inside information as a feature

Prediction markets explicitly rely on some participants having better information than others, which is the mechanism that makes them work rather than a flaw in the design: without participants holding a genuine edge, a market would simply reflect prior odds and the aggregation function would do nothing. The harder problem is curation, distinguishing legitimate edge such as domain expertise or better synthesis of public information from illegitimate edge such as trading on material non-public information, which requires an active editorial standard rather than a purely mechanical one.¹

The arson problem

A structural vulnerability follows directly from the same logic: if the event being predicted can itself be influenced, a large enough position creates a financial incentive to cause the event, the clearest example being wildfire-acreage markets nicknamed arson markets, where a large enough long position on acreage burned has a positive expected value from actually starting a fire.¹ Keeping such markets small and prioritizing informational value in what gets listed is a design constraint rather than a full solution, since the problem scales with market size and bettor wealth, and editorial curation does not obviously stay a workable enforcement mechanism as the number of markets grows into the tens of thousands.

Liquidity as the prerequisite

Prediction market odds are only as accurate as the underlying liquidity pool is deep: a market with a handful of participants on a niche question produces odds no better than an informal poll, while a market with billions of dollars in volume on a major election produces odds that outperform essentially every other forecasting mechanism.¹ Giving predictions away for free, with no access fees and no trading fees, is the audience-building step that starts the resulting flywheel, where accurate odds draw more viewers, more viewers build more credibility, more credibility brings more institutional legitimacy, and more legitimacy brings more liquidity, which in turn produces more accurate odds.

Crypto rails as a structural enabler

Leading prediction markets are commonly built on stablecoin settlement and permissionless blockchain rails rather than traditional payment infrastructure, which is not incidental. Traditional rails require identity verification on every user, restrict participation by geography, and add settlement latency that makes real-time odds updating difficult, while crypto rails let a participant anywhere bet on a locally relevant event without the platform needing a banking relationship in every jurisdiction. This connects directly to the broader thesis that the same infrastructure, stablecoin settlement plus permissionless access, that prediction markets rely on is also the infrastructure autonomous AI agents need in order to transact on their own.¹

Hyperliquid's outcome markets: a payoff primitive

Jeff Yan reframes prediction markets not as a standalone product category but as one configuration of a more general convex-payoff primitive. Spot trading and perpetual futures are both linear instruments; an outcome-markets primitive exists for nonlinear or convex beliefs that cannot be expressed through any combination of the two, payoffs that are binary, bounded, or option-like. These are fully collateralized contracts where both sides post capital and settlement lands somewhere between the two outcomes: binary outcome markets are prediction markets in the classic sense, continuous ones behave more like options, and a third variant aggregates opinion where there is no single objective outcome to resolve against. The design philosophy is to build one small, native primitive that all of these reduce to, then let builders deploy specific markets on top of it. The same known failure modes carry over directly to a permissionless version of this primitive and could be amplified by it: the arson problem, where an influenceable event creates a perverse incentive, and the deeper problem of how a market's settlement is determined to be objectively true in the first place.²

Yan was building decentralized prediction markets in 2018, years before founding Hyperliquid, around the same time as an early rival product.³ In his own account, the idea and the infrastructure choices were largely right even then, offchain matching with onchain settlement, but the market was not ready: it was not obvious anyone wanted onchain financial products yet, and after an early speculative run-up and crash within the same year, it became difficult to get anyone to try the product at all, a discouraging experience on a first attempt.² The outcome-markets primitive is described as the same instinct coming full circle once ecosystem, liquidity, and genuine demand finally exist together. The broader path runs from decentralized prediction markets, to automated trading built from watching gaps between decentralized-finance products and reality, to perpetual futures, to spot trading, and finally back to outcome markets, on the view that prediction markets and perpetual futures are both instruments for expressing a probabilistic belief about a continuously evolving number, and require structurally identical infrastructure: permissionless settlement, a neutral order book, and trust established through curation rather than gatekeeping.

CZ: timing, the regulatory tailwind, and price discovering truth

CZ describes himself as a strong proponent of prediction markets, calling them huge potential and hot, and notes his own venture fund invests across several of them. His account adds two things to the underlying mechanism. Timing was the missing ingredient in earlier attempts: the idea itself was very obvious well before it worked, and it needed the right conditions to actually ignite, a pattern he extends to other ideas that return for a second wind with a small but decisive tweak once the environment has changed. And prediction markets do price discovery and truth discovery simultaneously, using price to discover truth, which he describes as the reverse of how ordinary markets work, where information typically arrives first and price moves in response to it; here the price itself is the information being produced.⁴

A separate, later account adds regulatory texture. An early prediction-market platform based in the United Kingdom shut down under regulatory pressure around 2013, an early instance of the same come-and-go pattern: the right idea arriving in the wrong regulatory environment. By 2026, the primary US derivatives regulator has moved from sympathetic tone to active support, filing briefs defending prediction-market firms against state-level challenges and releasing new rulemaking specifically for the category, a meaningful tailwind given the federal-versus-state tension is concentrated on sports-outcome markets, historically regulated state by state as gambling. The strongest single claim offered is that prediction markets are far more accurate than any other forecasting method, because anyone with genuinely good predictions has an incentive to be in the market, the same filtering mechanism, uninformed bettors lose and exit while informed bettors profit and stay, extended to weather forecasting and risk pricing more broadly, on the premise that real money on the line is what makes information good.⁵

An institutional-venture framing

A venture investor's fund announcement names prediction markets as a leading example of a broader thesis that tokenization enables entirely new markets by creating global liquidity pools without needing region-by-region infrastructure. The forward-looking claim is that today's prediction markets, largely sports betting and politics, could extend into new event-risk hedging, insurance, and business-outcome markets whose resolutions feed directly into automated, conditionally triggered capital flows, tying prediction markets to the same rails autonomous economic agents are expected to use rather than leaving them a standalone speculative category.⁶

The academic backbone

A scholarly account frames the market price itself as a real-time probability sensor, an always-on instrument that reads a crowd's collective forecast off live trading rather than off a periodic survey. The mechanism is considerably older than any recent election: sixteenth-century betting on papal succession in Rome is an early documented instance, and formal validation came from academic experimental-economics work in the 1980s showing that markets aggregate dispersed private information into accurate prices, with a well-known university-run market later demonstrating election forecasting that routinely beat polling. Corporate and scientific applications extend the same mechanism beyond politics, including internal markets used to forecast product ship dates or demand, and markets on whether a given scientific finding will replicate. A second, distinct failure mode from the arson problem is adverse-selection unraveling: if some traders may already know an event's outcome, rational counterparties widen their spreads or withdraw entirely, liquidity thins, and the market can collapse before producing a usable price at all, a structural problem curation cannot fix. And a market price is conditionally, not unconditionally, calibrated: a well-known market underestimated both a major referendum outcome and an upset election result in 2016, a direct check against treating the mechanism as an oracle rather than as the best available estimate given current participants and liquidity.⁷

The signal is 24/7

A fintech investor reads prediction markets as infrastructure rather than as a category of gambling: whatever one calls the activity, it runs continuously, and autonomous agents need rails that are open continuously too, so the same infrastructure being built to support round-the-clock betting is also building the rails that let this kind of continuous activity happen at all. The expectation is that both incumbents and new entrants will ship agent-facing features on top of exactly this kind of always-open rail, an early signal of financial infrastructure catching up to the pace at which AI systems actually operate.⁸

Tensions

How well accuracy holds up on thinner, less liquid markets covering obscure questions, where genuinely informed participants may be few, remains an open question beyond the handful of high-liquidity cases that have been directly tested. Regulatory posture in major markets is currently permissive but politically contingent, and a change in administration or regulatory philosophy could reintroduce legal uncertainty. The arson-market scaling problem is real: editorial curation works at current scale and may not work once the number of listed markets grows by orders of magnitude. And a genuine tension exists between the rise of prediction markets and the commercial model of traditional polling, since if markets persistently outperform polls, the polling industry's underlying business case is directly threatened.