Where the liquidity comes from, and who is on the other side

Every prediction market shows you a number between zero and one and invites you to read it as a probability. Whether you can depends entirely on where the number came from, and there are only a few places it can come from. This page is about which one you are looking at, how to tell, and what each one does to the price you actually get. What a contract is in law is a separate question, handled in event contract or swap.

How it works

Three mechanisms, and they answer the question "who takes the other side" differently enough that the same 60 cents means three different things.

An order book: the counterparty is another user

A central limit order book holds two queues per outcome — bids from people willing to pay, asks from people willing to sell — and the venue's job is to match them. Polymarket's documentation puts the consequence plainly: prices "aren't set by Polymarket — they emerge from supply and demand as users trade with each other." The same shape is what Limitless documents for its order-book markets, with a separate Yes book and No book, and what Kalshi runs.

Two mechanical details change how you read the screen.

The displayed price is usually not a price you can trade at. Polymarket displays the midpoint of the bid and the ask, and falls back to the last traded price when the spread is wider than 10 cents. If the best bid is 34 cents and the best ask is 40 cents, the screen says 37 and you will pay 40 or receive 34. The number in the middle is where nobody is standing.

In a binary market, a bid on one side is an ask on the other. Kalshi's order book endpoint returns only bids, for both Yes and No, because a Yes bid at X is a No ask at 1 − X. To get the Yes spread you take the best Yes bid and subtract the best No bid from a dollar. It is the same book either way, but if you read only one side of it you will conclude there are no sellers when in fact every No bid is one.

The property that matters: an order book can be empty. If nobody is quoting, there is no price, and the venue does not invent one.

An automated market maker: the counterparty is a curve

An AMM replaces the queue with a funded pricing function. You always get a quote, because the curve always has one, and the quote moves against you as you buy — that is not a defect, it is how the mechanism stays solvent. Myriad documents the constraint the curve enforces: outcome prices always sum to $1, so a price of $0.60 is the curve's implied 60%. Manifold runs the same shape in play money, with limit orders resting on top of the curve.

Three things follow, and they are what an order book does not have.

1. The price is always there, and it is always worse as you get bigger. Slippage is your own footprint. On a book, a large order eats successive levels; on a curve, a large order moves the curve, and the average price you pay is strictly worse than the price you saw.
2. Somebody funded it. The curve's depth comes from liquidity providers, and their money is the reason a quote exists in a market with no other traders. In return they collect a share of the fee, and they hold the position the curve took against you.
3. Volume can be low and a price still shown. This is the single most misleading combination in the sector. On Myriad on 19 September 2026, all 70 open markets returned by the public API were AMM markets, and of the 39 denominated in a stablecoin, the median traded volume was about $11,500 and the median liquidity figure about $1,000. Every one of them displayed a confident two-decimal probability.

A pool: nobody quotes anything

In a parimutuel pool there is no counterparty to your trade at all, because there is no trade. Everyone entering an outcome pays into that outcome's pool, the pools close, the event happens, and the winning pool divides the total. PancakeSwap's Prediction product — a crypto price game, not a venue in this catalogue, cited here because its documentation states the arithmetic outright — runs 5-minute rounds and publishes the formula as a ratio: the payout ratio for the UP pool is the total value of both pools divided by the value of the UP pool, and 3% of each round's prize pool goes to the treasury.

Read that formula again, because it contains the whole difference. Your payout depends on what other people do after you commit. A price you were quoted on a book or a curve is fixed at the moment you take it; an implied price in a pool is provisional until the pool closes. If the side you chose gets crowded after you enter, your return shrinks, and nothing about your own entry changed. There is no price to check and nothing to arbitrage, because there is nothing tradeable until settlement.

Hybrids, and the shape that is neither

Futuur is the hybrid worth knowing. Its API documentation describes ordinary order types — a market order executes against the best available price in the book, a limit order rests in it — and its own card records the rest: where the book is empty, an LS-LMSR market maker quotes both sides, so a price exists for any outcome at any time. What you see in the API is a two-sided ladder in every real-money market we sampled on 19 September 2026, including levels of 0.03 shares and one cent of notional. That is not the shape a queue of human orders makes.

Limitless is hybrid in a different sense: it runs order-book markets and AMM markets side by side as distinct market types, each with its own fee schedule, rather than falling back from one to the other inside a single market. Which one you are in is a property of the market, not of the moment.

And ForecastEx is the shape that is neither. You submit a bid for the Yes or the No position; when opposing bids sum to $1.00 the exchange pairs them, and each side's contract is with the exchange rather than with the other bidder. There is no resting inventory and no curve — the counterparty is the clearing house, funded by the person whose opposite bid completed yours.

Why a thin market's price is not a probability

This is the part the screen hides. A price is what someone will transact at. A probability is a belief about the world. They converge only when it is worth somebody's while to push them together, and four things stand between.

The capital is tied up until resolution. Buying at 50 cents a contract that settles in twelve months costs you fifty cents and a year of whatever that fifty cents would otherwise have earned. At a one-year Treasury yield of a few per cent, a contract that is genuinely a 50% proposition is worth a couple of cents less than 50 to hold, and more than that on longer dates. This is why fair value on a long-dated contract sits below the honest probability rather than at it, and why a venue that pays interest on posted collateral — ForecastEx does, by rulebook — is not making a marketing point but removing a real term from the equation.

The fee lands on both legs. Entry and exit are two taxable events in almost every schedule here, and if you plan to close before resolution you pay twice. Polymarket's taker fee is C × feeRate × p × (1 − p), which for 100 crypto-category shares at 50 cents is $1.75 on a $50 trade — 3.5% of the stake, and 7% if you round-trip it at the same price. That is a wide band to be right inside before you make anything.

The question can resolve against how you read it. Resolution risk is not a tail. It is the source named in the market's rules, the timetable it is read on, and what happens if the source does not say what the question assumed. It belongs in the price, and on a thin market nobody has put it there.

Nobody is coming to fix it. In a deep market a mispriced contract is somebody's job. In a market with a thousand dollars of depth, correcting a five-cent error earns fifty dollars before fees and ties up capital for months, so nobody does it, and the error stays. The absence of that person is the actual reason a thin market's number drifts — not that its traders are worse, but that being right about it pays nothing.

None of this makes the number useless. It makes it a price with a known list of things inside it, which is a different object from a forecast, and the honest way to use it is to name which of the four you think is moving it.

What it costs

Published schedules, as they stood on 19 September 2026, with units. All of these are the venue's own documentation, and a fee is per fill unless it says otherwise.

• Polymarket. Takers only; makers are never charged. The formula is fee = C × feeRate × p × (1 − p), with C the number of shares and p the price, so the dollar fee is symmetric around 50 cents and largest there. Rates by category: 0.07 crypto, 0.05 sports, economics, culture and weather, 0.04 finance, politics, tech and mentions, and zero on geopolitical markets. On 100 crypto shares the published table runs from $0.07 at a penny to $1.75 at 50 cents.
• Limitless. AMM markets: a flat 0.40% on all trades. Order-book markets: takers only, with a buy fee of 0.40% to 3.00% and a sell fee of 0.42% to 1.50%, both varying with the price you trade at. The buy fee sits at its 3.00% maximum anywhere from a penny to 50 cents, and falls to 0.42% at 99 cents. Resting a limit order that is not immediately matched costs nothing.
• Myriad. AMM fees are set per market, 0% to 2% on buys. Reading the fee field on every open market through the public API on 19 September 2026: of the 39 open markets denominated in a stablecoin, 32 charged 1% on a buy and all 39 charged nothing on a sell; the 31 markets denominated in points charged nothing on either side.
• A parimutuel pool. The fee is a cut of the pool rather than of your stake — 3% of each round's total in the PancakeSwap example — and it is taken once, at settlement, because there is only one event to charge.

Two patterns are worth carrying away. A price-dependent fee is not a percentage of your position, and cheap-looking contracts are where it bites hardest: on Limitless a buy at 5 cents carries the same 3.00% as a buy at 50. And maker-free is not free — you avoid the fee by resting an order, which means accepting the risk that it fills only when someone else knows something you do not.

What "deep" means, in numbers

"Deep liquidity" is a claim, and four numbers with dates on them settle it. Ask for these, or compute them yourself from the venue's public order-book endpoint.

1. Traded volume over a stated window — 24 hours, or since listing. Lifetime volume on a market that has been open for eight months tells you nothing about today.
2. Open interest, which is money currently at risk rather than money that has changed hands. Volume can be one bot trading with itself; open interest cannot.
3. The spread, in cents, between the best bid and the best ask right now.
4. Depth at a stated distance from the midpoint — how many dollars are resting within two cents either way. This is the one nobody quotes and the only one that tells you what your own order will do.

What the range looks like in practice, both measured on 19 September 2026:

• Polymarket's "Will the U.S. invade Iran before 2027?" market: best bid 15 cents, best ask 16 cents, a 1-cent spread, with about $34,100 resting within 2 cents of the midpoint on the bid side and $60,900 on the ask side, across 15 bid levels and 77 ask levels.
• Futuur's market on the 2026 Brazilian presidential election: best bid 50 cents, best ask 56 cents, a 6-cent spread, nothing at all within 2 cents of the midpoint, and about $30 of bids and $48 of asks within 5 cents.

Both are long-running questions with real public interest, priced near the middle, on venues that work. The depth differs by roughly three orders of magnitude, and no adjective on either site distinguishes them.

For a sense of the top of the range: across the 300 highest-volume open order-book markets on Polymarket that day, 243 carried a two-sided quote, the median spread was about 0.2 cents, and only 5 had a spread of 5 cents or more. That is the population the phrase "prediction markets are efficient" was built on, and it is a few hundred markets, not the listing.

What you can do about it

What the venue charges on top of all this is published, and published in four incompatible units — what a trade actually costs converts them.

On a venue that settles on a public chain you can go one step further and look at which addresses made up that thin book. It is less than it sounds: what wallet tracking shows covers what an address does and does not tell you about the person behind it.

1. Identify the mechanism before you size the position. The venue's own documentation says which it is, usually in one page — Polymarket and Kalshi document an order book, Myriad an AMM, Limitless both as separate market types, Futuur a book with a maker behind it. If you cannot find that page, that is information too.
2. Pull the order book and compute depth at 2 cents yourself. Kalshi and Futuur both state in their documentation that the book endpoint needs no authentication, and Polymarket's returned a full book to an unauthenticated request on 19 September 2026; the numbers in the section above were computed that way, in one request each. See the venues listing and the API cards for the endpoints.
3. On a curve, quote your own size, not the displayed price. Ask the interface for the average price on the order you actually intend to place. The gap between that and the headline number is what the market costs you to enter, before any fee.
4. Price the round trip, not the entry. Multiply the fee by two if you intend to close early, add the spread, and only then ask whether your edge survives. On a 50-cent crypto contract at Polymarket's published rate that is roughly seven cents of the dollar before you are right about anything.
5. In a pool, decide with the pool closed, not open. The implied odds you see early are a function of who has entered so far. If the product lets you enter late, late is when the number means something.
6. Treat a price with no depth behind it as one person's opinion, correctly reported. That is what it is. It is still worth reading — it is just not a market's estimate until there is a market.