Automated Market Makers in Sports Prediction Markets

Prediction market volume on sports contracts jumped from under $16 billion globally in 2024 to nearly $64 billion in 2025, with sports now accounting for more than 80% of all trading, according to WilmerHale's May 2026 analysis. That growth sits on top of a plumbing question most people never think to ask: how does a market on a college lacrosse semifinal or a live NFL in-game prop get priced when there's no professional trader sitting on the other side of the bet? That's the automated market maker's job. Whether it does that job well or badly is the difference between a platform that works and one that quietly bleeds its liquidity providers dry, and this piece takes a position on which design choices actually hold up under sports-specific stress and which ones just look good in a pitch deck.

Monthly volume on Kalshi and Polymarket combined went from under $5 billion in September 2025 to roughly $24 billion by April 2026, per Pew Research Center. Super Bowl Sunday alone cleared more than $1 billion on a single market. At that pace, a platform that can't price a trade instantly, especially for the long tail of minor-league games and obscure props, fails its users in real time. Traditional order books solve this fine when professional market makers show up to quote both sides. They don't show up for a Belarusian second-division soccer match, and that gap is exactly what AMMs were built to fill. Worth understanding what they actually do before getting into where they break.

What an AMM actually does inside a prediction market

Sports prediction contracts trade between $0 and $1. If the outcome hits, the contract pays a dollar; if it doesn't, it pays nothing. The price at any given moment is just the market's implied probability, dressed up as a tradable asset.

An AMM removes the human on the other side of the trade and replaces it with a formula. Prices move mechanically as orders come in, governed by a cost function or a reserve ratio rather than a matching engine pairing buyers with sellers. Two designs dominate. The Logarithmic Market Scoring Rule, or LMSR, has the market operator act as a permanent counterparty, with price following a logarithmic cost function and the operator's potential loss capped by design. The Constant-Product Market Maker, familiar to anyone who's used Uniswap, prices outcomes off the ratio of reserves sitting in a pool. Worth correcting a common assumption here: this mechanism actually originated in prediction market theory before DeFi popularized it, so calling it "borrowed from crypto" gets the history backwards.

In both setups, there's no order book and no waiting around for a counterparty to show up. Any trade, at any size, executes instantly at whatever price the formula produces. The one lever controlling how violently a trade moves that price, the liquidity parameter (b in LMSR, pool depth in CPMM), gets set by the platform operator, and it shapes nearly every trade-off from here forward. The AMM's core promise is that it's always open and always quoting. What that promise costs, and who ends up paying for it, is the thread the rest of this piece keeps pulling on.

The LMSR's foundational role and why it was the first real solution

Robin Hanson introduced LMSR in papers from 2003 and 2007, aiming at the thin-market problem: how do you get honest probability signals in a market before any traders have shown up to signal anything? His answer was a subsidized market maker that quotes continuously no matter what, and it's still the right starting point for understanding everything that came after.

A continuous double auction only produces a trade when a buyer and seller agree on a price. LMSR skips the negotiation; the platform becomes the permanent counterparty and never turns down a trade. It also comes with a bounded loss guarantee, the operator's maximum possible loss calculable in advance from the liquidity parameter b and the number of outcomes. A platform launching a market on a niche event has zero guarantee that market makers will bother showing up, so LMSR lets the market run from the first trade regardless of whether anyone else joins in.

Polymarket built on LMSR early, precisely for that cold-start advantage. Prices update smoothly and transparently on-chain from trade one straight through to settlement, no gaps, no stale quotes sitting around waiting to be exploited.

But here's the catch LMSR can't dodge: a higher liquidity parameter buys more depth, and it also demands a bigger subsidy the operator has to lock up before the market even opens. That capital sits committed across the full probability range, for every outcome, including outcomes with no realistic chance of ever being close. Most of that money just sits there doing nothing, and the inefficiency gets worse, not better, the bigger the platform scales. This is the part LMSR boosters tend to skip past.

Three structural failure modes that make binary sports markets uniquely hard for AMMs

Binary sports markets break AMMs in ways generic DeFi pools never experience, and the three failure modes compound rather than sit side by side.

Start with settlement loss. In a typical DeFi pool, impermanent loss is a risk that might or might not show up depending on how prices move. In a binary sports market, it always shows up. As the event nears its conclusion, the pool automatically rebalances toward whichever outcome looks more likely, and when the market settles, the losing side's shares drop to zero, leaving the pool holding tokens worth exactly nothing. Unless the fees collected over the market's lifespan outweigh that capital loss, the liquidity provider is left holding the bag, and that happens because the mechanism is built to behave that way, not because of bad luck. Passive liquidity provision here requires an actual view on whether fee income beats a loss baked in from the start.

Then there's jump risk, and this is where sports markets diverge sharply from crypto or equities. A starting quarterback gets ruled out 90 minutes before kickoff. A team scores to go up two possessions with four minutes left. These are instant gaps between what the formula says and what's actually true, rather than gradual price movements gliding along the AMM's curve. Arbitrageurs close that gap fast, trading against the AMM's now-stale price and pulling value straight out of the pool. In-play markets make it worse: score changes happen close to continuously, so the AMM is more or less always a step behind anyone actually watching the game.

The third failure mode has a name: Loss-versus-Rebalancing, or LVR. It measures exactly how much value arbitrageurs extract from an AMM trading at stale prices, isolating the pure cost of adverse selection, what a hedged liquidity provider loses to better-informed traders regardless of which way the price eventually moves. In sports, where information lands in sudden bursts (an injury report, a score, a weather delay), LVR is usually the single biggest cost a liquidity provider faces. Paradigm's research on a prediction-market-specific AMM design, called pm-AMM, flags LVR as the defining risk metric for these outcome-token pools and notes that existing designs deliver inconsistent liquidity because the volatility of an outcome token depends on both the current probability and how much time is left before the event resolves. Two variables, not one; a constant-product formula built for token swaps was never designed to handle that.

How the platform architecture split between AMMs and order books actually maps onto sports markets

Polymarket's own history is the clearest case study, and it argues against treating AMMs as the default. The platform launched on LMSR, then dropped it for a hybrid, decentralized central limit order book (CLOB) in late 2022. The reason was capital inefficiency and the near-certainty of LP loss under the AMM model, the exact failure modes covered above.

The current setup matches orders off-chain through professional market makers, with settlement happening on-chain, and every event carries a unified order book with mirrored orders across both outcomes. In March 2026, Polymarket crossed $10 billion in monthly volume for the first time, driven largely by NCAA Tournament trading, decent proof that CLOB architecture handles serious sports throughput once the volume shows up to support it.

The advantage of a CLOB in a liquid market is real: tighter spreads, room for price improvement, no forced slippage on big orders. But it depends entirely on professional market makers being willing to quote both sides, and that only happens reliably around high-volume, marquee events. Nobody's quoting a third-round golf prop bet on a CLOB. That's where the AMM earns its keep: a minor-league game, a niche prop, an overseas league final, anywhere no institutional market maker has a reason to show up. An AMM still produces a quote where an order book would just sit empty.

The split shows up cleanly in practice. Kalshi and Polymarket lean CLOB-heavy for flagship events where professional order flow is dependable; Kalshi posted $12.35 billion in monthly notional volume in March 2026, with the bulk of that coming from sports event contracts. Platforms oriented around small, user-generated markets, meanwhile, tend toward CPMM-based AMMs, since that long-tail model is exactly the kind no CLOB could ever staff with quoters. The real question for anyone building or picking a sports prediction platform is which segment of the market a given platform is actually trying to serve.

Oracle integration and real-time price feeds as the practical fix for jump risk

Jump risk has a partial fix, and it's fairly intuitive: pipe outside price data straight into the AMM's pricing formula so the curve adjusts the moment something happens in the real world, instead of sitting stale until an arbitrageur corrects it for free.

SportX AMM is one documented example of a sports-focused platform that addresses jump risk through its AMM design, aiming to reduce adverse-selection losses for liquidity providers. The logic is straightforward: real-time sports odds already exist in abundance courtesy of the traditional sportsbook industry, so a platform can feed that data in directly or route it through an external oracle provider like Chainlink. Oracle providers like Chainlink point at the bigger role external data feeds play here, securing the live data stream that lets market makers deploy capital confidently while the game is still being played, then settling contracts after the final whistle.

What does this actually change, and what does it leave untouched? Oracle-fed pricing meaningfully cuts LP exposure to sudden probability shocks. It does not erase the guaranteed loss on the losing side of the pool at settlement; that structural cost from the section above is still there, just smaller. It also introduces a new dependency: a delayed or manipulated feed creates its own category of exploitable mispricing, so the fix trades one risk for a narrower, more contained one. The upside is that sports-specific data infrastructure, live scores, official injury designations, verified play-by-play, is more mature than in almost any other prediction market category. That's a genuine structural edge for building AMMs around sports rather than, say, elections or entertainment awards.

What determines whether a sports AMM actually functions well for participants

The AMM's basic pitch, continuous liquidity with no counterparty required, is real and it works. The open question for any given platform is whether it delivers that pitch at a price participants can actually live with, and most platforms get one of the four levers below wrong.

Liquidity parameter calibration comes first. Set it too low and the market lurches around on modest trade sizes, scaring off anyone informed; set it too high and the operator's subsidy commitment becomes unsustainable. The right number tracks expected volume and volatility, not some fixed default copied from a template. Oracle integration matters just as much: an AMM with no real-time price anchoring starts at a structural disadvantage in sports specifically, and whether a platform sources external odds, official stats, or on-chain signals determines how fast a stale price gets corrected without draining liquidity providers in the process. Fee structure needs to account for the fact that settlement loss on the losing side is a certainty, not a risk; a fee rate copied straight from a perpetual DeFi pool is mispricing the product from day one, full stop. And market scope has to be honest about where AMMs belong. Thin, long-tail, in-play markets are their home turf. Forcing an AMM onto a high-volume marquee game where professional quoters would happily show up just costs participants in slippage and liquidity providers in losses, for no real benefit to anyone; that's the mistake worth naming directly rather than hedging around.

For someone actually evaluating a platform, liquidity depth and spread width are the observable tells. A properly calibrated sports AMM shows spreads that track reasonably close to the external sports betting market, and a large order shouldn't move the price by more than the underlying probability shift genuinely justifies. With platforms like Kalshi carrying an $11 billion valuation, the architecture decisions being made right now define the liquidity rails for a market that could end up many multiples of its current size. AMM design here is a product decision participants feel directly, every time they place a trade.