How prediction markets work

What are you actually buying at 0.62?

A binary prediction market has exactly two outcomes and two matching instruments: a YES share and a NO share. Each is a claim on the same escrowed collateral, usually a dollar stablecoin. At resolution, one of them is worth $1 and the other is worth $0. Nothing in between, and nothing after — the instrument is extinguished when the market resolves.

That is the whole reason the price lives on a 0–1 scale. A share trading at 0.62 costs 62 cents to acquire and returns a dollar if it wins, so it is a bet at roughly 1.61 to 1 that a market participant was willing to lay. Read the other way, 0.62 is the price at which the marginal buyer and the marginal seller disagree about whether the event is more or less likely than 62%.

The temptation is to call 0.62 “the probability”. It is closer to the probability than almost anything else you can observe, and it is not the probability. It is the price at which someone will trade, which includes the cost of getting in and out, the cost of having capital locked until resolution, and the risk that the rulebook resolves in a way you did not expect. Those wedges are small in a deep market and large in a thin one.

注 Shares are fungible and transferable before resolution. That is the feature that makes a prediction market a market rather than a wager: you can be right early, sell at 0.90, and never find out how the event ends.

Where do the shares come from if nobody is selling?

This is the mechanism people miss, and it explains most of the rest. Shares are not a fixed supply that traders pass between them. Anyone can deposit $1 of collateral and receive one YES and one NO — a complete set. The reverse works too: hand back one YES and one NO together and you get your $1 back, at any time, without waiting for resolution.

Splitting and merging is the arbitrage that holds the market together. If YES is bid at 0.70 and NO is bid at 0.35, someone can mint a set for $1 and sell both legs for $1.05, riskless, until the bids come back down. If YES is offered at 0.55 and NO at 0.40, someone can buy both for $0.95 and merge them into $1. The two bounds are strict, and they are what makes YES + NO sum to one instead of merely tending toward it.

Two consequences follow immediately. First, there is no short interest, no borrow fee and no squeeze on a prediction-market share the way there is on a stock: supply is elastic, and anyone can create more by posting collateral. Second, the maximum you can lose on a position is what you paid for it — a long share cannot go below zero, and the collateral behind the set is already posted.

How does a trade actually happen — order book or AMM?

Two designs dominate. The older one is an automated market maker: a formula holds the inventory and quotes a price that moves as you buy, with no counterparty needed. It is elegant for markets nobody is watching, because it always quotes something, and it is expensive for size, because the price walks up as you fill.

The other is a central limit order book, which is what the large venues use for anything with real volume, Polymarket included. Participants post limit orders — a price and a size they are willing to be filled at — and the venue matches incoming orders against them by price and then by time of arrival. Matching is typically done off-chain for speed, with settlement on-chain, so the ledger of who owns what stays verifiable while the book stays fast.

The distinction matters when you place an order. A limit order joins the queue and pays no spread if it is filled — but it is only filled if the market comes to you, and it is filled first precisely when the market is moving against you. A market order takes what is resting and pays the spread plus whatever depth it consumes. On thin markets, the second cost dwarfs the first, which is why the same trade can be cheap at noon and expensive at three in the morning.

If you are going to trade against a book rather than through one, learn to read it. The order book guide covers depth, spread, imbalance and the difference between displayed and available liquidity.

Who is providing the liquidity?

Four groups, with different motives. Market makers quote both sides continuously and aim to earn the spread while holding as little directional risk as possible; they are the reason a price exists at all in the quiet hours. Arbitrageurs enforce the complete-set bound above, and the bounds between related markets — a “winner” market and the matching “top two” market cannot disagree by much without someone taking the difference.

Then there are informed traders, who arrive when they know something and are the reason market makers widen. And finally the flow that pays for the whole structure: participants trading for opinion, entertainment or conviction. The distribution of outcomes among that last group is not a mystery, and it is worth knowing before you start.

Liquidity is not evenly spread. It concentrates in headline markets and, on short-dated markets, in the final minutes before resolution, when uncertainty collapses and makers can quote tightly. A market with a 0.02 spread and $50,000 within a cent of mid, and a market with a 0.15 spread and $300 resting, are the same instrument with completely different economics. Check the book before you decide the price is wrong.

What happens at expiry, and what can go wrong?

At resolution, the winning shares are redeemed for $1 each out of the escrowed collateral and the losing shares become worthless. There is no counterparty who might not pay: the money was posted when the sets were minted. The interesting risk sits one step earlier, in deciding which side won.

That decision comes from a rulebook plus an oracle. The common design is an optimistic oracle: someone proposes the outcome with a bond, a dispute window opens, and if nobody challenges it within the window the proposal stands. A challenge escalates to a vote of token holders, which takes longer and introduces its own incentives. The mechanism is generally sound, and the failure mode is almost never fraud — it is ambiguity.

• The rulebook, not the title, is the instrument. “Will X happen by June 30?” is decided by the source and the timestamp named in the fine print, in the time zone named there.
• Ambiguity is priced, badly. Markets whose wording admits two readings tend to trade as if the friendlier reading is certain, until the day it isn’t.
• Resolution takes time. Even an uncontested market has a dispute window, and a disputed one can take much longer. Capital is locked for that whole period, which is a real cost on a low-margin trade.
• A market can resolve to a third state — invalid, or 50/50 — in some rulebooks. That is not a bug; it is a term you agreed to.

The practical habit is simple and rare: read the resolution criteria before you read the price. If you cannot state in one sentence what would have to be true for your shares to pay, you do not have a position, you have an opinion with money attached.

Is the price the probability?

It is the market’s best available estimate, and it is systematically off in places you can measure. Costs create a wedge: if taking liquidity costs a few percent of notional, a share worth 0.62 in probability terms cannot trade at 0.62 both ways. Locked capital creates another: a market resolving in nine months has to offer something over the risk-free rate to be worth holding, which pushes prices toward the extremes in ways that have nothing to do with belief.

And behaviour creates a third. The classic finding across betting markets is a favourite–longshot pattern, where low probabilities trade rich and high probabilities trade cheap. Whether that holds in a given venue at a given moment is an empirical question, not a law — which is why we publish our own measurement rather than assert the pattern. The public register shows how five-minute markets on BTC, ETH, SOL and XRP actually resolved by price band, including the bands where the market prices the outcome better than our own signal does.

The general skill here is calibration: reading a probability as a claim that can be checked over many events rather than as a verdict on one. That is its own guide — why a 70% forecast should be wrong 30% of the time.

How is this different from a bookmaker?

A bookmaker is your counterparty. It sets the odds, takes the other side of your bet, and manages its own risk across the book it has accumulated. Its prices carry an overround: convert every quoted odd on a bookmaker’s market to an implied probability and the sum exceeds 100%, and that excess is the margin. Because the bookmaker holds the risk, it can also decline it — limiting or closing accounts that win consistently is a normal part of that business model.

An exchange is not your counterparty; another participant is. The price is discovered by the book rather than posted by the house, the venue’s revenue is an explicit fee rather than a spread it built in, and the complete-set identity holds the implied probabilities at roughly 100% plus the spread instead of 100% plus a margin. No one refuses your order because you have been right too often.

The trade-off is honest rather than one-sided. The exchange gives you a two-way market, transparent costs and the ability to change your mind. It also gives you no guaranteed fill, an oracle you have to trust, a venue you have to trust, and the possibility that when you want out, nobody is bidding.

If you want to compare the two directly, the arithmetic is conversion: decimal or fractional odds into an implied probability on the 0–1 scale, then subtracting the overround to see what the bookmaker was really charging. The odds converter does both.

What does a round trip really cost?

Three components, and traders routinely count only the first. There is the explicit fee, quoted by the venue. There is the spread, paid every time you cross it — and you cross it twice on a round trip if you enter and exit as a taker. And there is depth: the difference between the best price and the average price you actually got, which grows with your size and shrinks with the market’s.

The reason to count all three is that the costs are the same size as the edges people chase. Our own register puts the cost of taking liquidity on Polymarket at roughly 3.5% of notional, and the widest measured gap between price and outcome in our calibration table is of the same order of magnitude. An edge that does not survive its own transaction costs is not an edge; it is a fee you pay with extra steps.

常见问题

Can I lose more than I put in?

Not on a spot prediction-market share. The most a long share can lose is its purchase price, and the collateral behind the complete set is already escrowed. That is different from a leveraged derivative, where the position can be closed against you — see liquidations and cascades.

What happens if I hold shares that lose?

They resolve to zero and are removed. There is nothing to settle and nothing to pay: your loss was realised when you bought, and the collateral goes to the holders of the winning side.

Why do YES and NO sometimes not add up to exactly one?

Because the quoted spread sits between the bounds. The best YES ask plus the best NO ask will be a little above one, and the two best bids a little below, and the gap is the spread the market makers are charging. A persistent violation of the bounds is an arbitrage, and it does not persist.