Polymarket Odds Accuracy 2026: 9,836 Scored
Polymarket odds accuracy measured on 9,836 markets resolved in October 2026: a day before settlement the price misses by 7.7pp, an hour before by 1.7pp.
Affiliate disclosure: contains affiliate links; we may earn a commission at no extra cost to you. 18+ only — see BeGambleAware.org. Last updated 17 October 2026.
Key takeaways
- Polymarket odds accuracy depends almost entirely on when you look. One hour before settlement the price was out by a weighted average of 1.09 percentage points. Twenty-four hours before, the same markets were out by 4.81pp.
- A day out, cheap contracts win more often than they cost. Contracts priced 5¢-40¢ resolved YES 31.5% of the time against an average price of 23.9¢ — a +7.65pp gap that survived an event-clustered bootstrap.
- That is the reverse of the textbook longshot bias, and the reverse of the largest published study of Polymarket. Contracts at 70¢-95¢ ran the other way, resolving 73.1% against a 80.6¢ price.
- The correction arrives late. The gap barely moved between three days out and six hours out, then collapsed in the final six hours. The crowd is not wise early; it is wise last.
- Volume does not buy accuracy. Markets above $250,000 of volume had the worst miscalibration of any band we measured and a 60.4% hit rate, against 68.6% in the $10,000-$25,000 band.
On this page
Every page that asks whether Polymarket odds accuracy is real answers the same way: it quotes the 2024 election, mentions Brier scores, and tells you the crowd is usually right. None of them tells you the thing that decides a trade — when the price becomes right, and how wrong it is before that.
The gap matters because almost nobody trades at settlement. If a 20¢ contract is fairly priced an hour out but worth 29¢ a day out, every number in those articles is describing a market you were not in.
So we scored the whole week. Every figure below comes from 9,836 markets that resolved between 10 and 17 October 2026 with at least $10,000 of volume — $888.9m of turnover, read from Polymarket's public Gamma and CLOB APIs on 17 October 2026, with each market's price taken from its own hourly history at fixed distances before it settled. Where our numbers disagree with a published study we say so, and where they replicate one we say that too.
What Polymarket odds accuracy looks like across 9,836 resolved markets
The sample is a census, not a selection: every order-book market above $10,000 of volume that Polymarket resolved in those seven days. 5,795 were sports (58.9%), 3,306 crypto (33.6%), 401 weather, and the remaining 334 spread across culture, tech, finance, politics and economics. The median market turned over $31,437.
6,154 of them had a tradeable price 24 hours before they settled. The rest were listed and resolved inside a single day — we come back to those below. For the 6,154, here is the same set of markets scored at four distances from settlement.
Brier score: mean squared error of the price against the 0/1 outcome, where 0 is perfect and 0.25 is a coin flip. Brier skill compares it to always quoting the sample's base rate. Mean error is the volume-weighted average distance between a price bucket and the rate it actually resolved at.
Read the bottom row against the one above it. Three quarters of the week's forecasting value arrives in the final six hours. A day out, the exchange is beating a naive base-rate guess by 17% and getting the favourite right two times in three — respectable, and nothing like the oracle the coverage describes.
Our 24-hour Brier of 0.2042 and 67.5% hit rate are worth holding up against the one other public dataset at the same horizon. SimpleFunctions scores 22,658 resolved Polymarket markets at the 24-hour mark and reports a Brier of 0.199 and a 67.1% hit rate. Two independent samples, five months apart in method, landing within half a point of each other. The headline number is not in dispute.
Where the price is wrong, and by how much
A Brier score hides the direction of the error. Calibration does not: bucket every market by its price, then ask how often the bucket actually resolved YES. A perfectly calibrated 20¢ bucket resolves 20% of the time.
The confidence intervals below are bootstrapped over events, not markets, because one football match can list thirty correlated contracts and treating them as thirty independent observations would make everything look significant. An asterisk means the price sits outside the interval the outcomes support.
Prices 24 hours before settlement, 6,154 markets across 2,621 distinct events. Gap is realised rate minus average price, in percentage points. Intervals are 2,000 bootstrap resamples drawn at event level; * marks a bucket whose price falls outside its interval.
Six buckets miss by more than the outcomes can excuse, and they miss in a pattern rather than at random. Everything below 40¢ paid more than it cost. Everything above 70¢ paid less. The middle of the book, between 40¢ and 70¢, is where Polymarket is doing its best work a day out.
The size of the misses at the cheap end is the part worth sitting with. A bucket averaging 7.32¢ that resolves 16.55% of the time is not a rounding error; it is more than double. And the error is not an artefact of one strange market type — we re-ran it with the tolerance on the price lookup tightened from six hours to ninety minutes, with every multi-outcome market dropped, with sports alone, and with only markets above $50,000 of volume. The shape held in all four. Dropping multi-outcome markets made the 5-10¢ gap larger, at +15.93pp.
The bias runs backwards from the textbook
The favourite-longshot bias is the oldest finding in betting-market research: punters overpay for longshots, so cheap contracts win less often than their price implies. It has been measured in horse racing since 1949 and it is what most writing about prediction markets assumes.
Our week runs the other way, and so does the comparison that matters. Aidan Walker's 2026 Princeton thesis — 188,509 resolved binary markets and 1.88m daily price observations from November 2022 to December 2025, the largest public analysis of Polymarket there is — finds prices "systematically compressed toward 0.50", with longshots overpriced and favourites underpriced, and a recalibration slope of 1.112. It even uses the same event-clustered bootstrap we do.
We find the opposite sign on both ends. That is a disagreement to explain rather than wave away, and the likeliest account is that the exchange has changed. Walker's window ends in December 2025; our week is 58.9% sports and 33.6% crypto, dominated by per-match spreads, over/unders and hourly price contracts that barely existed in his dataset. A finding about Polymarket in 2023 is largely a finding about election markets. The clock differs too: we anchor on the moment Polymarket settled each market, he uses daily observations.
What both datasets agree on matters more than where they differ: the price is not the probability, the error has a direction, and that direction is stable enough to be worth knowing before you click.
The correction arrives in the last six hours
Tracking the same two bands across the four horizons shows when the market does its thinking.
Same 6,154 markets at every row. Band populations differ by row because prices move between buckets as settlement approaches.
From three days out to six hours out, the bias is flat. It does not drift, decay or improve. Then, inside the last six hours, both ends snap toward the diagonal and the 70-95¢ band overshoots to the other side of it.
That shape says something about where the information comes from. If the crowd were steadily aggregating opinion, the error would shrink steadily. It does not. It sits still until the money arrives — team news, a line move, the first half of a match — then resolves almost at once. Polymarket's forecasting reputation is earned in the last few hours of a market's life and borrowed for the rest of it.
Can you actually trade the gap?
A +9pp gap on a 7¢ contract is a very large number, so the obvious question is whether buying every cheap contract a day out is a strategy. Start with the arithmetic, then the reasons to distrust it.
Polymarket's taker fee is published as fee = C × feeRate × p × (1 − p), where C is the number of shares and p the price, so as a share of your stake it is feeRate × (1 − p) — largest on the cheapest contracts. Makers are never charged. We read each market's own live rate rather than the published category table, because the two disagree: sports runs three schedules, two of them at 0.03 and the main one at 0.05.
Buy the YES side of every market in the bucket at its 24-hour price, hold to settlement. Fee is each market's own live taker rate applied to the bucket's average price. Selected buckets; the full set follows the same shape.
The fee does not touch it. On a 7¢ contract the fee is 4.58% of stake against a gross return of 126%, and on the losing side of the book a 1.21% fee is the least of a 13.8% loss. If the gap is real, fees are not what stops you taking it.
Execution is a different matter, and it is where we would put the money. Our 15 October measurement of 897 live order books found a median of $197 available at the best ask, with 41.9% holding under $100, and a $1,000 market order filling at the quoted price in only 32.8% of markets. Under 15¢ the books are thinnest of all: a $1,000 buy below 2¢ filled 296% above the quote.
So the honest statement is narrow. At the midpoint, on paper, over one week, cheap contracts were underpriced by enough to clear the fee several times over. Collecting it would mean filling a couple of hundred dollars per market near the mid, across hundreds of markets, without moving any of them — and the bias being there next week, which one week cannot tell you. We did not trade it and we are not telling you to.
Volume does not buy accuracy
The intuition that deeper markets are better forecasts is repeated everywhere. TradeTheOutcome's accuracy report puts a number on it: 61.4% accuracy below $10,000 of volume rising to 84.7% above $1m. Our week says the opposite.
Prices 24 hours before settlement. Miscalibration and resolution are the reliability and resolution terms of the Brier decomposition — lower miscalibration is better, higher resolution is better.
Part of this is fair to the deep markets: a contract only attracts $250,000 of turnover when people disagree, so the biggest books hold the hardest questions, and the resolution column falls to 0.0300 as they do.
But miscalibration controls for difficulty, and it is worst in the largest band at 0.0085 — more than three times the $100k-250k band's 0.0025. The deepest markets are not just answering harder questions; a day out they answer them with more bias than the small ones. "High volume means trustworthy" is a screening rule this week's data does not support.
Polymarket odds accuracy by category
Prices 24 hours before settlement. Brier skill is measured against each category's own base rate, so the categories are not directly comparable on difficulty. Taker rates read live from each market's own fee schedule.
The ordering replicates Walker's thesis — politics at the top, sports at the bottom — on a later sample with no overlap. Two independent datasets agreeing on the shape while disagreeing on the longshot bias makes the ordering worth trusting.
The practical reading is less flattering. The categories Polymarket forecasts best are the ones it now trades least: 37 politics and 53 tech markets resolved in the week against 4,788 sports. Nine in ten resolutions came from the two weakest categories, and crypto carries the highest taker fee at 0.07 — the fee side is in our Polymarket fees guide.
A third of the week was coin flips
3,682 of the 9,836 markets — 37.4% — were listed and settled inside twenty-four hours, so they never had a 24-hour price to score. They are not a rounding error, and they are not like the rest of the exchange.
2,672 of them are crypto, which is 80.8% of every crypto market that resolved that week: hourly and daily contracts on whether a coin is up or down. Their base rate is 50.05% — as close to a coin flip as a sample of 3,682 gets. And one hour before settlement they were still a coin flip, scoring a Brier of 0.2106 and picking the right side 60.0% of the time, against 0.0899 and 86.0% for the markets that had been listed a day or more.
The late convergence that makes Polymarket interesting does not exist in short-dated crypto markets, because no information arrives to converge on: a one-hour BTC up/down contract is a near-random walk carrying the dearest taker rate on the exchange. Those markets are not forecasts. They are a coin flip with a spread.
Check the horizon before you check the price
Every figure here is one week of settled markets. Prices, depth and fee schedules all move — read the live market before you act on any of it.
Visit Polymarket.com →Affiliate link. 18+ only. Polymarket restricts access from the United States, Canada and other listed jurisdictions. Prediction markets carry a substantial risk of loss. Nothing here is financial advice.
What this sample cannot tell you
It is one week — 10 to 17 October 2026, during a dense football and college-football calendar. A bias measured over seven days is a measurement, not a law.
The clock is settlement, not kick-off. On sports, Polymarket can resolve a couple of hours after the final whistle, so for roughly a third of markets the outcome was already known at our one-hour reading, which flatters that row. The 24-hour row is almost always genuinely pre-event and is the one we would defend.
The price is a midpoint, not a fill. A real buyer pays the ask, and on a 7¢ contract with a 1¢ tick the half-spread alone is a material share of stake.
The composition is not the Polymarket of the headlines. 92.5% of resolutions were sports or crypto. If your interest is elections, 37 markets is all this dataset has to say — not much, very accurately.
My verdict
Polymarket prices are a good forecast at the moment almost nobody needs one. An hour before settlement they are out by about a point and get the right side 86% of the time. A day before, they are out by nearly five points and get it right two times in three — and the error has a direction you can state in one sentence: cheap is too cheap, and heavy favourites are too dear.
I would not build a strategy on that from one week of data, and the order books our slippage run measured would cap any such strategy at a couple of hundred dollars a market. I would absolutely use it as a reading rule. A 25¢ contract a day out is not a 25% chance; in this sample it was a 35% chance. An 80¢ favourite is not an 80% chance; it was closer to 73%. Knowing which side of the book you are standing on is worth more than any single price.
What would change my mind is a month of this. If the direction flips in November, a week of Polymarket is noise and the only durable finding here is the convergence curve — which is the one I would bet on being stable anyway. New to the mechanics? Start with how Polymarket works.
Related guides
- Polymarket Slippage 2026: 897 Books Measured — the execution cost that decides whether any of the gaps above are collectable.
- Polymarket Fees 2026: What You Actually Pay — the full taker schedule by category, and why the published table and the live markets disagree.
- How Does Polymarket Work? Explained for 2026 — the YES/NO share mechanics behind every price in this article.
- Polymarket vs Kalshi 2026 — the regulated alternative, and how its market structure differs.
- Polymarket Resolution and UMA Disputes — what happens between a market closing and the outcome we scored it against.
- Polymarket Maker Rewards 2026: $852 a Market — what the exchange pays to have the depth these trades would need.
Polymarket odds accuracy FAQ
Are Polymarket odds accurate?
It depends on when you read them. Across 6,154 markets that resolved between 10 and 17 October 2026, the price one hour before settlement scored a Brier of 0.0899 and picked the right side 86.0% of the time. Twenty-four hours before settlement, the same markets scored 0.2042 and 67.5%. The price is a strong forecast late and a mediocre one early.
What is a good Brier score for a prediction market?
Zero is perfect and 0.25 is what you get by quoting 50% on everything. More useful is the Brier skill score, which compares the market to always quoting the sample's own base rate: our 24-hour figure of 0.169 means the price removed about 17% of the error that guessing the base rate would leave. An hour out that rose to 0.634.
Does Polymarket have a favourite-longshot bias?
In this sample it has the reverse of one. Twenty-four hours out, contracts priced 5¢-40¢ resolved YES 31.5% of the time against an average price of 23.9¢, and contracts at 70¢-95¢ resolved 73.1% against a price of 80.6¢. That is the opposite sign to the classic bias and to Aidan Walker's 2026 Princeton thesis on 188,509 markets, which found prices compressed toward 50%. Our week is 58.9% sports and his data ends in December 2025.
Are high-volume Polymarket markets more accurate?
Not in this week's data. Markets above $250,000 of volume had a Brier of 0.2280 and picked the right side 60.4% of the time, against 0.2020 and 68.6% in the $10,000-$25,000 band. Some of that is because deep markets hold harder questions, but the miscalibration term of the Brier decomposition — which controls for difficulty — was also worst in the largest band, at 0.0085 against 0.0025 in the band below it.
Which Polymarket categories are the most accurate?
Politics and tech, by a distance: Brier skill of 0.808 across 37 politics markets and 0.760 across 53 tech markets, against 0.124 for the 4,788 sports markets that made up most of the week. The ordering matches the Princeton thesis. The catch is volume — 92.5% of the week's resolutions were sports or crypto, the two weakest categories.
Can you profit from mispriced Polymarket contracts?
On paper the 5-10¢ bucket returned 126% of stake gross and 121% after its own taker fee, because the fee on a cheap contract is about 4.6% of stake. In practice the books are too thin to take much of it: our 15 October measurement found a median of $197 available at the best ask and a $1,000 market order filling at the quote in only 32.8% of markets. One week is also not evidence that the bias repeats. This is a measurement, not advice.
Are Polymarket's hourly crypto markets accurate?
They are close to coin flips. 2,672 crypto markets resolved within a day of being listed in our week — 80.8% of all crypto resolutions — with a base rate of 50.05%. One hour before settlement they still scored a Brier of 0.2106 and picked the right side only 60.0% of the time, against 86.0% for markets listed a day or more in advance. Crypto also carries the exchange's highest taker rate at 0.07.
Sources and method
Population: every Polymarket order-book market with at least $10,000 of lifetime volume that resolved between 00:00 UTC on 10 October 2026 and 14:55 UTC on 17 October 2026 — 9,889 markets found, 9,836 scored after dropping 53 that had not settled to a clean YES/NO or carried no price history. Total volume $888.9m. Metadata came from the Gamma API, paged by volume band because the offset parameter stops at 2,000; prices came from the CLOB /prices-history endpoint at hourly resolution, read with explicit start and end timestamps. A market's price at a horizon is the last hourly observation at or before it, within a six-hour tolerance; tightening that to ninety minutes changed no figure here by more than 0.3pp. Intervals are 2,000 bootstrap resamples drawn at event level. Fee rates were read from each market's own live schedule. All figures are one week of settled markets and nothing here is financial advice.
- Polymarket docs: fees — the fee = C × feeRate × p × (1 − p) formula, the published category rates, and "makers are never charged fees".
- Walker, "The Price is (Almost) Right" (Princeton, April 2026) — 188,509 resolved markets, Nov 2022 to Dec 2025, Brier skill 0.398 overall, 0.565 politics and 0.264 sports, and the compression-toward-50% finding this run contradicts.
- SimpleFunctions: prediction-market calibration — 22,658 resolved Polymarket markets at the 24-hour mark, Brier 0.199 and a 67.1% hit rate, which our independent sample replicates.
- TradeTheOutcome: Polymarket accuracy report — the 61.4%-to-84.7%-by-liquidity claim this week's data runs against.
- Casino.org: are prediction markets accurate? — the best-ranking general explainer, and a page with no calibration table in it.
- Gamma API: /markets — resolved outcomes, volume, event grouping and the per-market feeSchedule used above.
- CLOB API: /prices-history — the hourly price series every horizon in this article is read from.