How Polymarket markets settle, cost and price

What this site is, and what it is not

This is a data analysis site. It does not place, route, or execute orders. It holds no funds, keys, or accounts, and it never asks to connect a wallet. It publishes no buy or sell recommendation, no directional call, and no single composite score that stands in for one. Every figure here describes publicly observable data, and every figure that could be wrong is labelled with how it could be wrong.

It is also not affiliated with Polymarket. Where this page describes the venue's behaviour, it describes what the venue publishes and what we have measured against it — the measurements are dated, and the dates are given.

How are Polymarket's trading fees calculated?

Polymarket's published taker fee is not a flat percentage of what you spend. It is:

fee, in USDC  =  rate × p × (1 − p) × shares

where p is the price of that particular fill and shares is how many shares that fill was for. The venue's own documentation gives the worked case: 100 shares at 30¢ with a rate of 0.05 costs $1.05 — that is 0.05 × 0.30 × 0.70 × 100.

Three consequences follow, and each one surprises somebody:

The fee peaks in the middle and vanishes at the edges. p × (1 − p) is largest at 50¢ and approaches zero as a price approaches 0¢ or 100¢. Spending a fixed dollar amount buys usd ÷ p shares, so the fee on that amount works out to rate × (1 − p) × usd: a $1,000 order at 96¢ pays about a twelfth of what the same $1,000 pays at 50¢.

It is symmetric across the two outcomes. p × (1 − p) is unchanged when you swap p for 1 − p, so the fee is the same figure whether you think of the trade as buying YES at 30¢ or as taking the other side at 70¢.

It is assessed per fill, at each fill's own price. An order that crosses several price levels is charged level by level, not once on its blended average. The two differ as soon as an order is big enough to move through the book.

The fee comes out of the proceeds, not out of your cash

On a buy, the trade's proceeds are the shares — so the fee is taken in shares, and you pay exactly the cash you entered. What shrinks is the payout.

Worked through, with a 5% rate:

That 140.71 is the "To win" figure on the venue's own ticket, and 71.07¢ is the average price it displays.

Charging the fee as extra cash on top instead describes a trade the venue does not offer: $101.50 for 142.86 shares reads as 71.05¢, which quietly understates the cost, because it prices a position in which you keep every share you bought.

This site made exactly that error until 2026-08-28, and a second one alongside it: the curve was written rate × min(p, 1−p) × shares, which tracks the real curve closely near the extremes but is twice too large at 50¢. Compounded, the two produced a $2.14 cash charge, $102.14 for 142.86 shares, and a displayed breakeven of 71.5¢ against a true 71.07¢ — overstating the cost of every position. Both are fixed. Both are recorded here because a fee model that has been wrong twice should say so out loud.

One caveat on the worked example above: at 70¢ it cannot distinguish the two curves. min(p, 1−p) equals 1 − p for every price above 50¢, so the share-denominated fee is identical on this side of the book. The curves separate only below 50¢ — at 30¢ the old form charged 0.05 × 0.30 where the real one charges 0.05 × 0.70.

On a sell, the shares leave regardless and the fee comes out of the cash you receive.

The rate is per market, and sometimes it is not published at all

Rates observed on live markets are 0.04, 0.05 and 0.07 — it is a per-market field, not one platform-wide number, and some markets publish an explicit zero. Some schedules mark themselves taker-only; some publish a maker rebate alongside the taker rate.

So there are four distinct states, and this site keeps them apart rather than collapsing them into a number:

1. A published zero. The market states it charges no taker fee. Every all-in figure can then be stated exactly.
2. A published rate on the curve above. Used as published.
3. A published rate on a curve we do not model. The schedule carries an exponent, and the venue's general form is rate × p × (1 − p)^exponent. Only exponent 1 is modelled here. Applying the exponent-1 number to a different curve would be a fabricated figure, so the fee reads as unknown instead.
4. Nothing published. Unknown.

An unknown fee is never rendered as zero. In states 3 and 4 every figure that would have included a fee is shown pre-fee and carries an explicit "excludes fees" label. Zero is a claim; unknown is the absence of one, and the two must not look alike.

If you are reading Polymarket's API yourself: the rate is in feeSchedule.rate. The sibling takerBaseFee and makerBaseFee fields are not the rate — measured across 855 live markets they were pinned at 1000 on every single one while the real rate varied. Reading 1000 as basis points would put a 10% taker fee on every market on the platform.

Why is the price I pay higher than the price on the page?

Because the number on the page is a midpoint, and a midpoint is not a price anyone is offering.

A market's book is a stack of resting orders at different prices. Buying does not transact at the midpoint; it walks that stack from the best price outward, taking whatever is resting at each level until the order is filled. Your average fill is the total cash divided by the total shares, and it is worse than the top of the book by an amount that grows with your size. That gap — the distance between the best available price and your average — is the cost of the size, and it is separate from the fee.

Three things follow:

• A price with no size behind it is not a price. "42¢" on a screen and "$2,000 of it at 42¢" are different claims, and only the second one is a trade.
• The visible book is finite. When an order is larger than what is resting, the remainder does not fill at a worse price; it does not fill at all until someone posts. An unfilled remainder is unknown, not expensive.
• NO is priced by complement, and there is only one book. A binary market's two tokens redeem against each other, so buying NO at 18¢ is the same transaction as selling YES at 82¢. There is one ladder, read from either end.

What is an effective breakeven?

The effective breakeven is your all-in cost per share, read as a probability: the cash you put in, divided by the shares you actually end up holding.

For the $100 buy above it is 71.07¢, which reads as: this trade needs the true probability of the outcome to be above 71.07% to be worth taking. On a sell the direction flips — the true probability has to stay below the number.

It is deliberately one figure rather than two. The average fill price and the breakeven are the same quantity in different clothes: the breakeven is the page price plus the spread you crossed plus the fee. Showing both invites the reader to treat them as two independent facts and compare them, which is a comparison with no content.

Note what it does and does not contain. It contains the price you crossed to and the entry fee. It does not contain the cost of getting back out — that is a separate round-trip figure, because a position held to settlement never pays it. And when the fee is unknown (states 3 and 4 above) the breakeven falls back to the pre-fee average and is labelled as excluding fees, rather than quietly reporting a smaller hurdle than the real one.

It is arithmetic, not a view. It states what has to be true for a trade not to lose. It says nothing about whether that is likely.

How do I go short, and what is my stake?

There is no separate short instrument. You go short an outcome by buying its NO side, and NO costs 1 − p.

This changes how the result should be measured. If YES trades at 82¢, NO costs 18¢ — so 18¢ is the capital at risk, and any move has to be measured against 18¢, not against a dollar and not in raw cents.

The identical three-cent move is a four-fold different result depending on which side you were holding, because the stakes differ by more than four times. Quoting moves in cents hides that entirely, which is why every return figure on this site is stated against the capital actually committed.

The illustration above is pure arithmetic on the mid: it excludes the spread you would cross and the fee, both of which are covered above. And note the symmetry point from the fee section — the fee is the same on either side, so it is not what makes one side better.

Why don't the outcomes of one event add up to 100%?

Sometimes they should and sometimes there is no reason they would, and the first thing to establish is which case you are looking at.

Polymarket calls a group in which exactly one outcome can win a negative risk event — a field of candidates, one winner. In such a group the YES prices across every leg should sum to about 1.00, and a deviation from 1.00 is meaningful. In an event that is merely a list of related questions, the legs are independent and nothing requires them to sum to anything. Adding those up and finding 1.43 is not a mispricing; it is a category error.

Even inside a genuine mutually exclusive group, there are four ordinary reasons the number you get by adding up what is on screen is not the group's real sum. All four are common enough that a hand-computed sum should be assumed wrong until they are ruled out.

1. Some legs have no quote, and their "price" is an artifact. A leg quoted 0.0000 bid / 1.0000 ask has nobody on either side. Its midpoint is exactly 0.50, and 0.50 is not a weak price — it is arithmetic performed on an empty book. Measured on the Next Prime Minister of Ethiopia group: 25 of its 33 legs quoted 0.0000 / 1.0000, and taking their midpoints at face value put the group's sum at 13.51 against the ~1.00 the group must actually sum to.

2. Some legs are placeholders that the site itself does not draw. Polymarket pre-creates a lettered slot for every seat a field might grow into, and the public API serves those slots like any other market — open, un-archived, order book enabled, accepting orders. Only an active flag says the venue's own page does not show them. Measured on the Oklahoma Democratic Senate primary on 2026-08-25: four named candidates carrying real quotes and $6.5k–$80k of volume, alongside 25 "Will Person C / D / …" slots with zero volume and a 0-bid / 1-ask book. Counting those 25 turns a four-way race into a 29-way one.

3. The list you are adding up may be truncated. Large groups are long: on 2026-08-18, 253 open groups had more than 40 legs, the largest with 128. Any interface that shows the top N legs is showing a subset, and a sum over a subset is not a smaller version of the real sum — it is a different number, and it reads as a mispricing that is not there.

4. There is no single price to add. Every leg has a bid and an ask. Summing midpoints across a dozen legs produces a total that no one could transact at, because you cannot buy every leg at its midpoint simultaneously. A deviation you could actually act on has to be computed by walking each leg's book at a real size — which is a different and much smaller number than the on-screen deviation.

What this site publishes, and when it refuses to

A group price sum is published here only when all of the following hold: the event is a mutually exclusive group; there is more than one leg; every leg carries a price; and the listing is not truncated. If any one of those fails, no sum is published and the reason is stated in its place.

That is a deliberately narrow rule, and it was narrowed after getting it wrong: a truncated group's top-40 subset sum was once published as though it were complete, and the missing tail read as a mispricing of several percentage points that did not exist. A sum that is confidently wrong is worse than a blank with a reason next to it.

What actually decides the payout?

The settlement text does — not the title.

Every Polymarket market carries a paragraph of prose stating the conditions under which it pays out. The title is a headline for that paragraph, and the two routinely disagree: the headline asks a simple question, and the text qualifies it with a source, a deadline, a tie-break, and a definition of the terms it uses. The payout follows the text.

That paragraph is the reason a market page exists here at all. It is the one piece of genuinely unique writing a market has; everything else is identity and numbers that read nearly the same across thousands of markets. A market that carries no settlement text gets no page on this site rather than a page with nothing in it.

Nothing on this site paraphrases it. Where a summary is shown, it is a verbatim sentence lifted out of the text, not a restatement of it. A summary that is subtly wrong is worse than no summary at all, because the reader has no way to tell which of the two they were handed.

Which sentence states the YES condition?

Polymarket's settlement text is templated closely enough that one sentence usually carries the whole condition, in the form "This market will resolve to 'Yes' if …". Where that sentence can be identified, it is quoted verbatim and shown above the full text.

The identification is deliberately strict. A candidate sentence has to contain both a resolution verb and the Yes outcome, so a stray "yes" elsewhere in the prose cannot match; and it has to be at least 40 characters, so a fragment like "Resolves Yes." cannot replace a useful disclosure with a useless headline.

Measured over 800 real markets carrying settlement text, 205 produced such a sentence — about 26%. That is the expected result, not a broken pattern. The misses are overwhelmingly markets whose text never states a per-leg YES condition because there is not one to state:

• one shared description covers every leg of a group ("…will resolve to the player who wins the 2026 …"), so no sentence is about this leg;
• the outcome pair is not Yes / No at all ("…will resolve to 'Odd' if the total combined rounds …").

Widening the pattern to catch those would mean asserting a YES condition the settlement text does not contain. For those markets the full text is shown with a note saying no single sentence could be identified — which is the correct answer for them.

When does a market actually settle?

Two dates are published per market, and they are not interchangeable.

• The settlement deadline is the one settlement actually runs on. This is the date shown here whenever it exists.
• The listed end date is administrative metadata, and it is frequently inaccurate.

Where only the administrative date exists, it is shown with a note saying so, and it is never used for anything time-sensitive — no countdown, no time-decay figure, no "closes in" arithmetic. This site has a standing rule against trusting it, and the rule exists because it has been wrong often enough to break anything built on top of it.

One thing neither date tells you: a deadline is when settlement is due to run, not a guarantee that the real-world answer is known by then. Markets whose underlying fact is still contested at the deadline are exactly the ones where the settlement text — the source it names, and its tie-break clause — does the work.

Where does the settlement answer come from?

Some markets name a primary source — the specific publication, feed, or official release the outcome will be read from. Where a market names one, it is shown verbatim.

Many markets name nothing. In that case the settlement text itself is the only description of how the answer is determined, and this site shows the field as absent rather than filling it in with a plausible guess. An inferred source would look exactly like a stated one, and the reader would have no way to tell them apart.

If a market matters to you and it names no source, that absence is itself information: read the full text and decide whether it is specific enough to be adjudicated the way you expect.

What if the settlement text changes after I have read it?

It can change, and a change to that paragraph is a change to the terms of the trade.

This site stores a digest of each market's settlement text so that a change is detectable. The text is whitespace-normalised before the digest is taken, because the upstream feed reflows the same prose between passes — and a digest that moved on reflow would announce "the rules changed" for markets whose terms were untouched, which is precisely how a change alert becomes something people learn to ignore.

Two limits, stated rather than papered over:

• We hold no per-field history, so we cannot tell you when the text last changed. No timestamp is shown for it, because the only timestamps available describe when any part of the market record changed — a volume tick moves them — and presenting one as a rules timestamp would be a fabrication.
• We publish no list of "edge cases" or "clauses most likely to surprise you." Deciding which clauses of a specific market's text matter is a judgement, and nothing here makes that judgement. Selecting sentences by keyword would produce a list that looks curated and is not.

What is available is the digest itself. Whether a reader is told "this changed since you last looked" depends on something being able to remember the version you last saw — and on this site, today, nothing does. That comparison ships in the PolySmarter browser extension, which keeps your baseline locally; the web pages do not yet keep one. When they do, the baseline will live in your own browser and never reach us.

Which price does this site show for a market?

The best evidence available, in a fixed order of preference, with the order itself being the point:

1. A tight two-sided midpoint — both sides quoted, less than 10 percentage points apart. A real, current quote.
2. The last traded price — an actual print. Stale but real, and better than the midpoint of a book nobody is quoting.
3. A wide two-sided midpoint — weak, but it is still a quote.
4. Nothing at all — for a book spanning 90 or more percentage points, which carries no information whatsoever. Its midpoint is an artifact, not a weak signal, and showing "50¢" for a market with no market would be worse than showing nothing.

Absence is shown as absence. A market with no published bid and a market bid at zero are different facts and are never merged into one.

What does "not covered" mean? It looks like an error

It is not an error. It is a normal cold state, and it is kept strictly apart from four others:

Merging any two of these would be a lie of a specific and expensive kind: "no data available" leaves the reader unable to tell whether to wait, to upgrade, to look elsewhere, or to report a bug.

One more distinction runs through every figure on the site: the vintage of the data and the time the page was assembled are recorded separately. They are routinely far apart, and stamping a stored result with the time you asked for it is the single most tempting misrepresentation available in a system like this one.

Glossary

Mutually exclusive group (negative risk)

A set of markets in one event where exactly one outcome can win. Their YES prices should sum to about 1.00, which makes the sum meaningful; in an event that is merely a list of related questions, it is not. See why the legs rarely add up on screen.

Settlement deadline

The date settlement actually runs on, as distinct from the administrative end date published beside it, which is frequently inaccurate. See when a market actually settles.

Primary source

The publication, feed, or official release a market names as where its outcome will be read from. Often absent, and shown as absent. See where the settlement answer comes from.

Effective breakeven

Cash in, divided by the shares actually held — your all-in cost per share, read as a probability. It answers "what would have to be true for this not to lose", and nothing else. See the full explanation.

Book Lean

Which side the resting order book structurally leans to, taking the size at each price level into account rather than only the best bid and best ask. It describes the current state of the book; it is not a probability forecast, and it is not a signal to act on. We call it Book Lean rather than microprice because nobody should have to learn an academic term before reading a chart.

Executable Relative Value

Whether a price deviation inside one event could actually be traded at a real size. A deviation measured on midpoints assumes every leg fills at its midpoint; walking each leg's book at a real size turns that into a larger execution cost and, usually, a much smaller deviation. It is the difference between a gap you can see and a gap you could act on. See reason 4 above.

Not covered

We hold no capture of that kind for this market — nothing was computed, and nothing failed. Distinct from an insufficient sample and from an error. See the five states.