What a prediction-market order book does that a crypto book does not

A prediction-market order book looks exactly like a crypto order book. Bids on one side, asks on the other, aggregated size at each price level. The data structure is the same, the API shape is the same, and every habit you carried over from a perpetual futures book will feel like it applies.

Five things are different, and each of them changes what a given observation means. If you already know how to read a level-2 book — the general version is here — this article is the delta.

The price axis is bounded, and it is a probability

A binary contract pays 1 if the event resolves yes and 0 if it does not. The price therefore lives in [0, 1], and it is the implied probability. There is no such thing as a breakout.

This single fact rewrites the arithmetic of every quantity you compute from the book.

Spread is not a percentage of the mid, it is a number of probability points. A one-cent spread is one point of probability, always, everywhere on the axis. But as a fraction of what you are paying, it explodes at the low end: one cent is roughly 1% of a contract at 0.99 and 20% of a contract at 0.05. Quoting spread in basis points of the mid — correct practice on a crypto book — produces a number that is technically right and practically useless here, because it hides that the cheap contracts are the expensive ones to trade.

Depth has a ceiling that direction does not. On a crypto book, a very large bid can absorb an arbitrary amount of selling. On a contract at 0.97, the entire remaining upside is three cents; size resting there is not conviction about direction, it is someone writing insurance and collecting the last of the premium. Reading it as bullish depth is reading a maximum payoff as a signal.

Percentile and z-score treatments have to respect the boundary. Price series that are pinned near 0 or 1 have their variance mechanically compressed. Any statistic you compute across the whole price range will be dominated by the middle of the book without you noticing.

There are two books and they are the same book

Buying YES at 0.62 and selling NO at 0.38 are the same trade. A resting bid on YES is, mechanically, a resting offer on NO at the complement. This is not an approximation — it is the definition of the instrument.

Our own book sampler makes the point without needing any interpretation. In more than 42 hours of continuous sampling ending 17 August 2026 it stored more than 11,000 snapshots across 28 Polymarket markets — 56 tokens, exactly two per market. Of those snapshots, roughly 4,300 had a mid below 0.05 and roughly 4,300 had a mid at or above 0.95; the two counts stay within a handful of each other from one read of the table to the next. Those are not two populations of cheap and expensive contracts. They are largely the same markets counted from both ends.

Several consequences follow, and none of them have an analogue in crypto:

• —Displayed depth is double-counted if you naively sum both books. The same intention appears twice, once on each side, mirrored around 1. Any imbalance measure that treats them as independent sources is measuring the same order against itself.
• —A full set of outcome shares costs 1 and redeems for 1. In multi-outcome markets this bounds the prices in a way that has no equivalent in a spot or perpetual book: the sum of outcome prices is anchored, and when it drifts, the correction is a mechanical operation rather than a directional view.
• —A quote can vanish from one book and reappear on the other without anyone changing their mind about anything. If you are sampling one side, you will read this as a cancel.

Sum the books once, in one orientation, before you compute anything.

The tick grid changes size, and it still does not save the cheap end

The habit worth dropping next is the assumption that the grid is uniform. It is not — and it is not coarse in the place you would expect.

The sampler is a live rolling table that gains a few rows every minute, so a split of it is a photograph rather than a constant. Here are the 11,421 snapshots it held on 17 August 2026 at 10:31 UTC, by where the mid sits — the counts drift within minutes, the medians did not move between reads:

Through the middle of the range the grid is effectively a whole cent: 94% of best bids in that band land exactly on one, and the median spread is exactly one cent. In the tails almost none do — 1.1% below 0.05, 1.6% at or above 0.95 — and the median spread there is a tenth of a cent, the finest step the grid allows. The tenth is the floor everywhere, in every band, including the middle; what changes across the range is whether quotes bother with the finer steps at all.

So the extremes are quoted ten times more finely than the middle, and it changes almost nothing about what trading them costs. Look at the last column instead.

A spread of a tenth of a cent is 0.10% of a contract at 0.9985 and 66.67% of a contract at 0.0015. Same spread, same participants, and — since the two tails of that table are largely the same markets seen from opposite sides — frequently the same orders. The entire difference lives in the denominator. Crossing the spread on the cheap side costs two thirds of what you are paying; crossing it on the expensive side costs a tenth of a percent.

That is the structural fact about the cheap end, and the finer grid is what makes it visible rather than what fixes it. There is somewhere to put a competitive quote down there. It is just that a competitive quote, expressed as a share of the price, is still enormous. The book at the low extreme is not thin because nobody is interested; it is priced in steps that are tiny in probability points and huge relative to the contract.

It is also, for what it is worth, exactly where the pricing is least reliable. On our own measurements of five-minute BTC, ETH, SOL and XRP markets — price read at T+180 seconds of a 300-second window, compared with the observed resolution — the lowest band resolves less often than its price implies: an average price of 0.3586 against 30.99% realised across 1,068 slots, a gap of −4.87 points. The 0.40–0.50 band is negative too, at −2.94 across 3,624 slots. Those two bands and the six others, including the ones that flatter us, are published in full on the public register, where these were read on 17 August 2026.

That shape — long-shot prices bid above their realised frequency — has been observed in betting markets for decades. A contract that costs six cents and pays a dollar sells itself. And as the next section shows, the taker fee is a larger share of a six-cent entry than of a sixty-cent one, so the cheap end is penalised twice.

The cost that dominates everything

On a crypto venue, the spread is usually the binding cost and the fee is a rounding error. Here it is the other way round — and the fee is not a flat percentage.

Polymarket's published schedule charges takers shares × feeRate × price × (1 − price), and charges makers nothing (fee schedule). On crypto markets that schedule puts the rate at 0.07, which peaks at $1.75 per 100 shares on a 50-cent contract: 3.5% of notional, the figure we publish next to our calibration work on the public register.

Two things follow from the shape of that formula rather than from the headline number. Divide the fee by the notional and the price cancels out, leaving feeRate × (1 − price) — the cost as a share of what you put in is largest on the cheapest contracts and falls towards nothing at the top of the range. And whatever that share works out to, it is charged on the whole notional, not on your edge.

This inverts the standard book-reading conclusion. On a perpetual book, the question "can I get filled at the touch" is the important one. On a binary contract, you can nearly always get filled, and the important question is whether the edge you think you have survives a cost charged on the entire position. A three-point view against a cost of the same order is not a trade, however confident you are. Why most prediction-market traders lose money is largely this arithmetic, applied at scale.

The practical implication for how you read the book: resting liquidity is worth far more here than it is on a crypto venue, because the alternative is so much more expensive and because the maker side of that formula is zero. Posting instead of crossing is not a refinement, it is most of the available improvement.

The end is scheduled

A perpetual future has no expiry, which is why it needs funding to stay anchored. A prediction-market contract has the opposite property: a known terminal value, reached at a known-ish moment, determined by a rule written in text.

Two things follow for the book.

First, liquidity drains into resolution. As the outcome becomes clearer, the price approaches a boundary, market makers have less premium to earn and more resolution risk to carry, and the book thins — precisely during the window when information is arriving fastest. The worst execution conditions coincide with the most informative period.

Second, the book cannot price the rule. What you are trading is a resolution rule applied to a named source at a stated time, and the ambiguity lives entirely in the wording: what counts as an official announcement, which source is authoritative, what happens on a partial outcome, what happens if the deadline passes undetermined. None of that is visible in depth, spread or flow. Traders who "were right" and lost read the title and not the rules, and no amount of order-book analysis substitutes for reading them.

A checklist

Before drawing a conclusion from a prediction-market book:

• —Did you sum the YES and NO books once, in one orientation?
• —Is the spread expressed relative to the contract price, not to the mid in basis points?
• —Is the size you are reading near a boundary, where the maximum payoff — not conviction — explains it?
• —Did you check the tick size where you are actually trading, rather than assuming the one you saw at 0.50?
• —Does the edge survive the taker fee — 3.5% of notional at 0.50, and a larger share of notional the cheaper the contract?
• —Have you read the resolution rule, the source and the deadline?
• —Are you sampling the delta stream in sequence, or polling snapshots of a structure that changes between polls?

The book here is a measurement of cost and capacity, exactly as it is in crypto. What changes is that the payoff is bounded, the second book is the first one mirrored, the grid changes size across the range without making the cheap end any cheaper to trade, the fee is charged on notional rather than on edge, and the whole thing has an end date written into it.

Related reading: the tools a prediction-market trader actually needs for where book data sits in a stack, seven ways copying a wallet goes wrong for what the tape does not tell you about the person on the other side, and five ways a large position gets built for the shapes flow leaves behind. The general mechanics of the instrument are in how prediction markets work and reading an order book.