Prediction markets (Polymarket)

Bounded [0,1] event contracts on an on-chain CLOB, microstructure essentially nobody else has written up cleanly. Thin books, probabilistic pricing, event-driven flow; the maths transplants but the payoff resolves to 0 or 1.

See it move

What is a prediction market, in microstructure terms?

A prediction market is an order-driven venue trading binary outcome contracts: a contract on "will X happen?" that pays 1 if yes, 0 if no. The price floats in $[0,1]$ and reads as the market's implied probability of the event. Polymarket runs these as an on-chain central limit order book, so the canonical microstructure applies, with three twists: bounded payoff, resolution, and thin event-driven flow. The sandbox above is that book, clamped to $[0,1]$.

Intuition first: a Polymarket contract on "Will candidate A win?" trades at 0.62. That price is a probability; the market thinks there is a 62% chance. If A wins, the contract settles at 1 (you make 0.38 per contract held long); if A loses, it settles at 0 (you lose 0.62). The "price" is never going to 105 or 50: it lives in $[0,1]$ and ends at exactly 0 or 1. Everything microstructural follows from those two facts.

The venue is a real order book: Polymarket uses an on-chain CLOB (a central limit order book with price-time priority, resting limit orders, a spread and depth) not an AMM bonding curve. So the microprice, order-flow imbalance, queue position and adverse selection all apply, computed exactly as they would on any book. This is why a microstructure-literate reader has a real edge here over the typical participant. The three twists that make it its own venue: (1) bounded payoff, price in $[0,1]$, so volatility, inventory risk and impact all behave differently near the bounds; (2) resolution, the contract terminates in a binary settlement, so inventory risk is jump-like, not diffusive; (3) event-driven, thin flow, activity clusters violently around news about the event, and the book is thin between bursts.

How does market making differ when the payoff is bounded to [0,1]?

The quoting maths is the same, but the bounds change the dynamics. Near 0.50 the contract behaves like an ordinary asset; near 0 or 1 the price cannot move much in one direction, so volatility is asymmetric and compressed, inventory risk is naturally capped, and the Avellaneda–Stoikov reservation-price skew interacts with the ceiling and floor.

Intuition: a contract at 0.95 can fall to 0 but can only rise 0.05. The payoff is bounded and the risk is asymmetric in a way no equity or crypto price is. A market maker quoting near a bound faces a lopsided book: limited upside on one side, the full distance to 0 (or 1) on the other. Your inventory's worst case is bounded (you know the maximum loss per contract is the price you paid, or 1 minus it) which is unusual and useful, but the distribution of outcomes is highly non-normal and concentrated near resolution.

Inventory risk is bounded but binary: you cannot be run over to infinity (the price cannot exceed 1 or fall below 0), but at resolution your position does not mark to a continuous price; it jumps to 0 or 1. So the inventory you carry into resolution is not diffusive price risk but a bet on the outcome, which a market maker explicitly does not want. Flattening before resolution is the discipline; the sandbox above shows the skew distorting near the edges and the terminal settlement jump.

How do event and news trading transplant to prediction markets?

Naturally, because on a prediction market the contract is the event. Event trading and news trading are the native mechanism: news that shifts the probability of the outcome moves the price directly, and the fastest correct interpretation of that news is the edge. There is no cleaner venue for event-driven trading, because there is no other factor in the price.

Intuition: on an equity, news is one of many things moving the price (sector, macro, flow, noise). On a Polymarket contract about a specific event, the only thing that moves the price is information about that event: the price is a pure probability estimate. So event/news trading is not a strategy you bolt on; it is the entire game. A reader who can interpret event news quickly and correctly has a direct, unobstructed edge.

The canonical event-trading machinery applies: distinguish scheduled vs unscheduled events (a debate at a known time vs a breaking development), trade the repricing, and beware that the slow-repricing edge is gone. The surviving edge is latency-to-react plus correct interpretation, increasingly an AI/NLP problem (read what AI changes for HFT). The microstructure twist: flow is violently clustered around event news (a Hawkes-like burst), and the book is thin between bursts. A market maker must widen or pull quotes when news hits, exactly when adverse selection spikes, and an event trader must reach the thin resting liquidity before it is pulled or repriced.

Why is pricing different: probabilities, not prices?

On a prediction market the price is a probability in $[0,1]$, so the natural quantities differ. There is no concept of "return" in percent terms; the relevant maths is calibration, implied probability, and the relationships between related contracts (mutually exclusive outcomes must sum to at most 1). Mispricings show up as probabilities that do not cohere, not as a cheap or expensive "price".

Intuition: thinking in prices misleads you here. A contract at 0.62 is not "cheap" or "expensive": it is a 62% probability, and the question is whether 62% is the right probability. The edge is being better-calibrated than the market, or spotting that a set of related contracts is internally inconsistent.

And calibration over time: as resolution approaches, a well-functioning market's probabilities should sharpen toward 0 or 1. The microprice and order-flow signals on this venue forecast the next move in the probability, not a price: the microprice imbalance logic applies, but the quantity it predicts is bounded.

What is resolution / settlement risk, and why is it unique?

Resolution risk is the risk specific to prediction markets: the contract is settled by an oracle's ruling on whether the event happened, and that ruling can be ambiguous, disputed, delayed or wrong. Unlike a price that always exists, a binary outcome must be adjudicated, so a market maker bears not just price risk but the risk that the resolution itself goes against expectation or is contested.

The oracle is the settlement layer, and it can fail. A Polymarket contract resolves when an oracle (a decentralised dispute-resolution mechanism, e.g. an optimistic oracle with a challenge period) rules on the outcome. If the event's truth is genuinely ambiguous ("did X 'officially' happen?"), the resolution can be disputed, delayed, or decided in a way the market did not expect. You can be right about the world and still lose if the oracle rules otherwise.

Inventory carried into resolution is a bet, not a spread. A market maker wants to earn the spread and stay flat; but near resolution the book goes thin (everyone is de-risking), so flattening is expensive exactly when it matters. Inventory you cannot shed before resolution settles to 0 or 1, a jump you did not choose to take. Settlement is on-chain, so the crypto counterparty, smart-contract and gas considerations apply on top of the event risk: a smart-contract bug or oracle exploit is a settlement risk with no equity-market equivalent.

The honest summary: prediction markets are an underexploited venue. The microstructure is genuinely less competitive than crypto perps or equities, so a microstructure-literate trader has a real edge, but that opening is paid for in resolution/settlement risk, thin books, and event/oracle ambiguity. It is the apply venue with the most distinctive risk profile in the whole atlas.

Worked example

A simplified prediction-market round on a synthetic binary contract ("Will event X occur by date D?", trading at 0.62, a 62% implied probability, payoff in $\{0,1\}$, thin book, resolution at D) illustrative, as of a 2026 worked snapshot. You quote bid 0.60 / ask 0.64, a 4-percentage-point spread, wide because the book is thin and adverse selection around event news is severe. A taker hits your bid: you buy 100 contracts at 0.60, inventory $q = +100$, capital at risk $= 100\times 0.60 = 60$.

Balanced case. A buyer lifts your ask: you sell 100 at 0.64, back to $q = 0$. Round-trip gross is $(0.64 - 0.60)\times 100 = +4.00$ on 60 of capital, the captured spread, with no inventory carried into resolution. This is the clean ideal: earn the spread, stay flat, never bet the outcome.

Bounded-payoff case. The contract drifts to 0.90 (the event now looks likely). Your inventory's upside is capped (it can only reach 1, so at most $+0.10$ more from here) while its downside (if the event collapses to 0) is the full $-0.90$. Near a bound the payoff is asymmetric in a way no continuous price is; the A–S skew has to respect the ceiling.

Cross-outcome arbitrage case. Suppose "A wins" trades 0.55, "B wins" 0.30, "C wins" 0.20: a complete, mutually-exclusive set summing to $1.05$. Selling each in proportion locks in the 0.05 overpricing (less fees and gas), a structural arbitrage with no single-asset equivalent; the mirror case (sum below 1) is the underpriced version. The numbers are illustrative and synthetic; real prediction-market spreads, depth, resolution mechanics, oracle behaviour and fees vary by venue and contract, so check the venue and oracle spec, as of 2026. Educational only, not investment advice; no P&L is promised.