How accurate is Polymarket? The tape says: depends on the book.

Polymarket called most of the major US political resolutions in 2024. People love quoting that. Then they go and treat every Polymarket price like a probability handed down from God. That’s the mistake.

The right question isn’t “is Polymarket accurate?” The right question is “accurate at what, in which conditions?” A 91-cent print on a $1.1B book and a 70-cent print on a $463 book are not the same animal. One is the considered view of every whale in the room. The other is whatever the last person paid to move it.

So here’s the markets-desk version. Seven sections. No equity-research preamble.

The headline number is the trap

“Polymarket called X of Y elections” is a meaningless statistic without conditional structure. A market priced at 70 cents that resolves YES isn’t “right” in any useful sense. It priced a 30% chance of NO. If you only judge it on the resolved-YES side, you’re grading on a curve nobody asked for.

The real question is calibration. Are the 70-cent markets right roughly 70% of the time? Are the 90-cent markets right roughly 90% of the time? That’s the only honest accuracy frame for a probabilistic market.

Public researchers have run this on Polymarket and Manifold across the 2024 cycle and earlier. Calibration plots, Brier scores, the whole spine. The methodology is well-established and we link out to the longer treatment in the methodology. The short version: where calibration has been measured on deep-book political markets, prediction markets calibrate well at the headline level. Tails (the 95c+ and 5c- buckets) are noisier, because the dollars on the wrong side of those markets are tiny, and a single trade moves the print.

And another thing — calibration on the deepest 50 markets tells you nothing about market 51. Most Polymarket markets, by count, are small. Most accuracy claims, by citation, are about the big ones. Watch the swap.

The Polymarket-marketing version of accuracy and the markets-desk version of accuracy are different statistics. The first asks “how often did the favourite win?” The second asks “were the prices well-calibrated across the whole distribution?” The first is a coin-flip benchmark dressed up as evidence. The second is the real test. The headline you keep seeing is the first.

Liquidity is the discriminator

The single biggest predictor of whether a Polymarket print is information or noise is book depth. Not category. Not topic. Depth.

Thin books drift. Under $50K of total volume, the price is whatever one motivated buyer paid last week. The bid-ask is wide, the resting size is small, and a $500 print can move the mid five cents. We covered the $463 weekly we wrote up on Day 2 for exactly this reason. Total volume $463. That’s a single dinner cheque. Not a market.

Mid books — call it $50K to $5M of cumulative volume — trade reasonably. Two-way action, some whales on the leaderboard, recognizable price discovery. Most of Polymarket’s political “long tail” sits here. The prices are usable, the size matters, but a single $50K trade can still move the print a few cents.

Whale books — the $10M+ category — are where Polymarket earns its accuracy reputation. The 2028 Dem nominee book has cleared $1.1B in cumulative volume. The rich-book situation on House 2026 sits at $5.4M, which is mid by Polymarket standards and credibly tradable. On a book of $50M or more, you’re looking at a two-way market with real depth on both sides. The price is doing work.

And another thing — book depth and time-to-resolution interact. The $1.1B on the 2028 Dem nominee book is deep, but the contract resolves in 2.5 years. Most of that depth is far-from-the-money, which is to say, noise. Near-resolution markets with the same dollar volume read tighter than far-from-resolution markets with twice the volume. Discount accordingly.

One more wrinkle. Cumulative volume isn’t the same as current depth. A market can have $20M in lifetime turnover and a $50K resting top-of-book today. The print is real only to the size that’s actually offered. If the book in front of you is thin, the print behind it is decoration.

When Polymarket is wrong, in three flavours

Markets don’t miss randomly. They miss in patterns. Three you’ll see repeatedly.

Flavour 1: thin book plus outlier whale. One trader with conviction can hold a thin book at the wrong price for days. The Wisconsin Supreme Court 2025 print, the various Trump-cabinet-pick markets from late 2024, plenty of resolved examples. If the volume profile is one $200K print and a long flat line, you’re reading one person’s view, not a market’s. Fade or ignore.

Flavour 2: news asymmetry. The market priced what was known on Tuesday. The answer turned on what was announced Wednesday. Markets aggregate information that exists. They don’t conjure information that doesn’t. When the resolution turns on a sealed indictment, an unreleased jobs number, a private polling result — the price was always going to miss. This isn’t an indictment of the market. It’s an indictment of asking the market the wrong thing.

Flavour 3: resolution dispute. The contract is written on a specific phrasing. Reality fits the spirit but not the letter. UMA gets called in to adjudicate, the dispute round lights up, and the price oscillates 20 cents on the procedural read instead of the underlying outcome. The tell: a market that prints 40c-60c-40c-60c over a few hours after the event has already happened. That’s not uncertainty about the world. That’s uncertainty about the contract. UMA decisions are the tell.

Note what all three failure modes have in common. None of them are “the market got the probability wrong.” They are “the market got the wrong inputs,” “the market was asked an unanswerable question,” and “the market priced the contract not the outcome.” If you’re scoring Polymarket as a forecaster, you have to control for all three before you call a miss a miss.

Polymarket vs polls, the short version

Long version is Marcus’s longer version of this. He’ll walk you through Brier scores and calibration plots and the cycle-by-cycle history. Worth your time if you’re comparing the two seriously.

The tape-reading take is shorter. Polls cost real money to run and trail by days. Markets cost almost nothing to read and update in seconds. For news cycles — the “what just happened” questions — the market reads first. For “who actually shows up on election day,” the poll reads better, because turnout is structurally about a population the market isn’t sampling.

Concrete example. On Polymarket, Gavin Newsom has been pricing around 25c on the 2028 Dem nominee book, with Kamala Harris near 8c. In Echelon’s Apr 17-20 Democratic primary poll, Harris led at 22% with Newsom at 21%. The market thinks Newsom is roughly three times Harris. The voters think it’s a tie. Both signals are real. The market is reading donor flow, recent press, and Newsom’s campaign-ish posture. The poll is reading name recognition and how the room actually feels right now. Neither is “right” in isolation.

The honest answer to “are prediction markets accurate?” is “more accurate than the cheap version of any single poll, less accurate than the consensus of good polls on the questions polls answer well, and more accurate than both on the questions polls don’t answer at all.” Anyone selling you a simpler version is selling.

What I watch when I want to know if a book is right

Five things, in order, before quoting any Polymarket price. A sixth if you’re writing the price up for someone else.

Total cumulative volume. Under $50K, you don’t have a market, you have an opinion. Under $500K, you have a thin market — treat the price as a soft signal. $5M and above, you’re reading something real.

24-hour volume. Cumulative is the past. 24h tells you who’s active right now. A book with $10M cumulative and $200 over the last day is a stale print, not a current price.

Book depth on bid AND ask. Both sides. A one-sided book — deep bids, no offers, or vice versa — means the price you see can’t actually be hit in size. Look at the next 5c level. If there’s nothing behind the top of book, the price is fragile.

Recent trade size distribution. A handful of $100K prints tells you whales are positioning. A wall of $50 prints tells you retail is. Both move the price. Only one is information.

Time to resolution. A 90c near-resolution print is real conviction. A 90c print on a market resolving in 18 months has months of carry, news risk, and re-pricing left. Same number, different signal.

Add a sixth if you’re writing about the market for a public audience. Check the cross-venue print. The same question on Kalshi or Manifold should be close to the Polymarket print. If it’s 5+ cents off and the books are both deep, one of them is wrong and the arb is the story. If one venue is thin and the other is deep, the deep one is right and the thin one is decoration.

What to watch next

The deep books I’m tracking through summer: the 2028 Dem nominee book on Polymarket (whether Newsom can hold 25c through Q3 without a real catalyst), the House 2026 control book at the current rich 84c (whether it bleeds toward the base-rate 70c or holds), and Kalshi’s growing 2028 nominee market as a cross-venue read. If those three move in the same direction in the same week, that’s the signal. If they diverge, that’s the story. We’ll cover both in the recurring archive.

Read the book before you read the price. The print is a number. The book is the truth.