How accurate are prediction markets? A study of 358,000+ resolved Polymarket and Kalshi markets

Prediction markets are well calibrated, and they get sharper as the event gets closer. We checked 358,236 resolved markets, 333,464 on Polymarket and 24,772 on Kalshi, against what actually happened. A day before resolution, contracts priced around 70¢ won about 75% of the time in Polymarket's standard markets, and across all markets the favourite won 76.9% of the time. On long-running markets that we could follow for a month, the favourite won 87.7% of the time 30 days out and 92.8% a day out.

But the famous "90%+ accurate" numbers mostly measure easy questions. When the market was genuinely unsure, priced between 10% and 90% a day before, the favourite won only 65.9% of the time. What is being forecast matters far more than where: sports games are about three times as hard to call as political questions. And one fast-growing family of markets is badly calibrated: the templated "ladder" markets (temperature brackets, point spreads and totals, exact scores) that make up 38% of our sample, where cheap outcomes won far more often than their prices said.

This is an independent study by the PredictDesks research desk: neither platform commissioned or reviewed it. The full dataset is free to download at the end.

Key findings

1. In standard markets, prices mean what they say. A day before resolution, the average gap between a price and the share of such contracts that won was 0.9 percentage points on Polymarket and 1.7 on Kalshi. Contracts priced 90¢ or more won 98.0% of the time on Polymarket.
2. Accuracy improves steadily as resolution approaches. On the same set of long-running Polymarket markets, the Brier score improved from 0.091 a month out to 0.064 a week out and 0.050 a day out (lower is better).
3. Easy questions inflate the headline numbers. 66% of the Polymarket markets in our sample resolved "No", most of them long shots that sat at a cent or two for weeks. Strip out everything priced below 10% or above 90% and the favourite's win rate falls from 76.9% to 65.9%.
4. Topic matters more than platform. On Polymarket, tech & science markets scored 0.040 a day out and sports markets 0.196. Sports, the biggest category on both platforms, is also the hardest to forecast.
5. Polymarket and Kalshi are about equally accurate. Weighted to the same mix of topics, the two platforms' Brier scores were 0.150 (Polymarket) and 0.147 (Kalshi). On 24 Federal Reserve meetings priced on both, their odds a day before the decision differed by 0.8 points on average.
6. Ladder markets are overconfident. In temperature brackets, point spreads and totals, and exact-score markets (38% of our sample, and 47% of Polymarket's 2026 markets), outcomes priced under 10¢ won 1.6 times as often as their price implied, and favourites priced 80¢ to 90¢ won only 76% of the time.
7. In standard markets, long shots are overpriced. On Polymarket, every $100 spent on contracts priced below 5¢ a day before resolution returned about $81, before fees.
8. A month out, prices are a little timid. In standard Polymarket markets, prices sat slightly too close to 50% a month out (a recalibration slope of 1.04, where 1.00 is perfect) and were spot on a day out (1.00).
9. Volume is not a reliable sign of accuracy. Above $5,000, bigger markets forecast better in some categories and worse in others; there was no consistent pattern.
10. Upsets are real but rare. In 1.1% of markets, the outcome that won had been priced at 10% or less a day earlier. The list includes the election of Pope Leo XIV and the 2025 Kamchatka megaquake.
11. Over a third of Polymarket's markets are coin flips by design. 37% of the 641,571 closed Polymarket markets with $5,000+ volume were short-term crypto "Up or Down" markets. We left them out: any study that keeps them in will look much worse.

Why another accuracy study?

Published figures for prediction market accuracy range from a Brier score of 0.02 to almost 0.20, and from "67% right" to "98.7% right". Most of that spread comes from what each study counted, not from the markets themselves.

Three things explain most of the disagreement:

1. Easy questions. Most markets resolve "No", and most of those trade at a cent or two for weeks. Each one adds an almost perfect score, so a sample full of them looks brilliant. Polymarket's own page says 78.2% of its markets resolved "No": a forecaster who always said "No" would be "right" 78% of the time.
2. Coin flips by design. Polymarket lists hundreds of thousands of five-minute, hourly and daily crypto "Up or Down" markets that open near 50%. Nobody can forecast them well, so a sample that includes them looks poor. That is likely why one public dashboard reports a Brier score near 0.20.
3. When the price is taken, and on which markets. A price a month out and a price an hour before kick-off are different forecasts. Comparing horizons on different sets of markets also mixes timing with topic: the markets still open a month before resolution are long-dated political and business questions, while a day out the sample fills up with sports games.

We built this study to deal with all three: we report how far each price was from the outcome (not just whether the favourite won), we leave out the coin-flip markets and report them separately, and we compare horizons on the same markets.

The data

We downloaded every closed Polymarket market with at least $5,000 of volume (641,571 markets) and a random sample of Kalshi: 4,000 of Kalshi's 15,377 market series, with up to 80 markets drawn from each. Then we kept the markets that:

• resolved cleanly to Yes or No (no 50-50 or voided results),
• had at least $5,000 of volume (Polymarket, in dollars) or 5,000 contracts traded (Kalshi),
• traded for at least a day, so there is a price to check a day before resolution,
• and were not "Up or Down" crypto markets, multi-leg parlays or other combination bets.

That left 358,236 markets across 117,954 events: 333,464 on Polymarket and 24,772 on Kalshi. For each one we pulled the hourly price history from the platform's public API and recorded the price 30 days, 7 days, 1 day, 12 hours and 4 hours before the market's reference time (the earlier of its scheduled end and the moment trading closed).

The sample reflects how fast these markets have grown. Polymarket closed 752 markets with $5,000+ volume in 2023 and 515,460 so far in 2026, 197,823 of them "Up or Down" crypto markets. Most of the markets in our sample therefore resolved in 2025 and 2026, and sports games make up 57% of it. 38% of the sample (137,394 markets) are templated ladder markets, which we analyse separately below.

How we measured accuracy

A good forecaster needs two things: the favourite should usually win, and the probabilities should mean what they say. We used five measures:

The Brier score is the standard yardstick, but it depends heavily on the questions: a set of near-certain markets scores close to 0 whoever forecasts it. That is why we also report the skill score and always compare like with like.

Finding 1: in standard markets, prices mean what they say

We split the sample in two. Ladder markets slice a single number into many brackets or lines: a city's high temperature in one-degree steps, a game's point spread or total at several lines, or every possible exact score. Both platforms now create them from templates by the thousand, and they behave differently (Finding 6). Everything else, 220,842 markets or 62% of the sample, we call standard markets.

In standard markets, sort every market by its price a day before resolution and the share that resolved "Yes" tracks the price closely. On Polymarket the biggest gap in any 10-point band was 2.9 points; on Kalshi, 4.5 points. In ladder markets the line bends away from the diagonal: cheap outcomes happen too often and favourites too rarely.

Three patterns stand out:

• Polymarket's standard markets were close to perfect, with a recalibration slope of 1.00. If anything, modest underdogs were a touch cheap: contracts at 10¢ to 30¢ won 1 to 3 points more often than their price.
• Kalshi's standard markets were a little bold (slope 0.92): cheap contracts won more often than priced and favourites less often, mostly in sports and "mention" markets. Kalshi's sample is about 13 times smaller than Polymarket's, so treat the exact numbers with care.
• Ladder markets were overconfident on both platforms. That is the subject of Finding 6.

Finding 2: accuracy improves as the event approaches

Comparing horizons on different markets is misleading, so here we follow the same markets through time: 30,014 Polymarket and 3,411 Kalshi markets that were trading at least 30 days before resolution.

These are long-running markets (elections, policy, awards, economic data), which are easier than the average market, so their scores are better than the overall figures. The useful point is the slope: on Polymarket about 29% of the error disappears between a month and a week out, and another 22% between a week and a day out. Our Polymarket favourite rates are 3 to 4 points below the platform's own published figures (90.3% a month out, 94.2% a week out, 96.6% a day out), which come from a different, unpublished sample, but the shape of the improvement is the same.

Finding 3: the easy questions flatter the averages

Most prediction markets ask whether something unlikely will happen by a date, and most of the time it doesn't. In our sample, 66.2% of Polymarket markets and 63.0% of Kalshi markets resolved "No".

Markets priced under 10¢ or over 90¢ a day out are nearly always right: the favourite won 97.4% of them. The real test is the middle of the range.

In those genuinely uncertain markets the Brier score was 0.215, against 0.25 for a coin toss. Prices in this range are still informative (a favourite at 65¢ won about as often as 65% implies), but a single 65% call is far from a sure thing. If you take one lesson from this study: read a price as a probability, not a prediction.

Finding 4: what is being forecast matters more than who forecasts it

The gap between categories dwarfs the gap between platforms. Elections, economic data and culture markets are relatively easy; sports games, weather and "mention" markets are hard.

Why is sports so hard? A game is a genuinely close contest decided in a few hours, and bookmakers have spent decades making their lines efficient, so there is little extra information for a prediction market to add. A question like "Will the Fed cut rates next week?" is different: the answer is usually well telegraphed. Sports markets made up 57% of our Polymarket sample and 56% of our Kalshi sample, which is the main reason both platforms' raw averages look worse than older studies of mostly political markets.

Finding 5: Polymarket vs Kalshi

On raw averages, Polymarket scored 0.150 a day out and Kalshi 0.143 (95% confidence intervals, resampling whole events: 0.149–0.151 and 0.139–0.147). But the two platforms list different things, so that comparison mostly measures their mix of topics.

Weighting each platform's category scores to the same mix (politics, crypto, sports, culture, tech & science, economics, weather and mentions), the gap narrows to almost nothing: a Brier score of 0.150 for Polymarket and 0.147 for Kalshi.

The one clear difference is in sports, where Kalshi's markets in our sample scored better: 0.180 (95% interval 0.175–0.185) against 0.196 (0.194–0.197) for Polymarket. But the two platforms list different sports markets. Kalshi's sample leans on player props and other one-sided questions (only 37% resolved Yes, against 41% on Polymarket), while Polymarket's includes tens of thousands of esports and tennis matches, the hardest sports to call (Finding 11). Like for like, we found no reason to prefer one platform's prices.

The cleanest test is the same question on both platforms. Both have listed every Federal Reserve rate decision since 2023, so we compared the probability each gave to the decision the Fed actually made.

On the 24 meetings priced on both a day out, the average gap was 0.8 points (1.1 points a week out). Polymarket gave the right answer the higher probability 11 times and Kalshi 12 times, and their Brier scores on these meetings were 0.013 and 0.012. Arbitrage keeps the two prices together, which is why the same question rarely gets a different answer (how cross-platform arbitrage works).

Two meetings show the limits of any market. A week before the September 2024 meeting, both platforms gave a half-point cut less than 8%; a day before, both had it near 53%, and the Fed cut by half a point. Before the September 2026 meeting, the first hike since 2023 went from about 55% a week out to about 87% a day out on both. And in July 2026 both platforms gave a hold only about 80%, reflecting real hawkish risk: three officials dissented in favour of a hike.

Finding 6: ladder markets are overconfident

Ladder markets are the fastest-growing part of both platforms in our samples. On Polymarket they went from 35 markets in 2024 to 122,671 in 2026, 47% of that year's markets. They are also where prices are least reliable.

The direction is the same for every type: outcomes priced under 10¢ a day out won 1.6 times as often as their price implied, while favourites priced 80¢ to 90¢ won only 76.1% of the time. Point spreads and totals were the worst (slope 0.72), exact scores were also badly overconfident (0.61), and temperature brackets were the least affected (0.93). Exact-score markets still have a low Brier score, because most scores are priced near zero and lose; the slope shows the problem.

Two things probably explain it:

• Thin books. A ladder splits one question's money across many brackets or lines, so each one is thin, and many quotes come from automated market makers working from a model rather than from traders with a view.
• The midpoint is not a price you can trade. We measure the midpoint between the best bid and the best ask. On a cheap bracket the ask is often well above the midpoint, so part of the gap would disappear for anyone actually buying.

If you trade ladder markets, check the spread before you trust the price, and treat a 5¢ bracket as more likely than it looks.

Finding 7: in standard markets, long shots are overpriced

If prices were perfectly calibrated, buying one contract in every market at a given price would break even at resolution. We treated every market as two contracts (Yes at its price and No at the rest) and checked what each price band actually paid.

In standard Polymarket markets this is the classic favourite-longshot bias, concentrated in the cheapest contracts: every $100 spent below 5¢ returned about $81. People overpay for lottery tickets. Modest underdogs were, if anything, slightly cheap (contracts at 10¢ to 30¢ returned $104 to $107), and from 30¢ up returns stayed within a few dollars of break-even, which is what efficient prices look like.

Kalshi's standard markets did not show the sub-5¢ discount ($96 per $100), and its contracts at 5¢ to 30¢ paid well over their cost, the "bold" pattern from Finding 1. In ladder markets, contracts under 10¢ paid about $159 per $100 at the midpoint, which is Finding 6 seen from the trader's side.

None of this survives careless trading. These returns are at the midpoint, before fees and the spread, which take a bigger share of a cheap contract. A strategy built on them needs positive expected value after costs and sensible position sizing.

Finding 8: a month out, prices are a little timid

A recalibration slope above 1 means prices sit too close to 50%: favourites are underpriced and long shots overpriced. In standard Polymarket markets the slope was 1.04 a month out, 1.01 a week out and 1.00 a day out. The Princeton thesis found a stronger version of the same thing on 188,509 Polymarket markets (overall slope 1.112).

The likely reason is the cost of waiting. Money locked in a near-certain contract for months earns nothing unless the platform pays interest or holding rewards, so traders demand a discount for tying it up, and that discount fades as resolution nears. Kalshi's standard markets were slightly bold at every horizon.

Finding 9: volume is not a reliable sign of accuracy

You might expect markets with more money in them to be more accurate. Above our $5,000 floor, the answer depends on the market.

Brier skill score of Polymarket markets a day before resolution; higher is better.

Bigger Polymarket markets scored better in economics and worse in sports and crypto, and about the same in politics. Two things are probably going on. Big markets attract the hardest questions (a close NFL game draws far more money than a 2¢ long shot), and the information that makes a forecast good arrives whether a market has $10,000 or $10 million in it. Thin markets still have a cost, but it shows up in the spread and slippage you pay to trade rather than in how well the midpoint forecasts. Below $5,000, which we did not study, single traders can move prices a long way.

Finding 10: the biggest upsets

Across both platforms, 3,981 markets (1.1%) were won by an outcome priced at 10% or less a day before. A 5% outcome should happen one time in 20, so across hundreds of thousands of markets thousands of upsets are expected. Some of the biggest, by volume:

A few of these are pure surprises, like an earthquake. Others moved on news that broke in the final day, which is exactly what markets are supposed to do: the price a day before was a fair reading of what was public then. Some prices moved before the news did. Unusual bets on María Corina Machado hours before the 2025 Nobel Peace Prize announcement led the Norwegian Nobel Institute to look into a possible leak, and the market on Nicolás Maduro leaving power became the first criminal insider trading case involving a prediction market.

Finding 11: a market that grew up fast

Raw accuracy on Polymarket has fallen year by year: its Brier skill score a day out went from 0.607 in 2023 to 0.303 in 2026. That does not mean the crowd got worse. The mix changed: sports went from a sliver of the platform to the majority of its markets. Political markets scored about the same every year, while sports markets got harder as the platform added ladder markets (Finding 6) and esports and tennis, the two hardest sports to call (Brier 0.214 and 0.204 a day out, against 0.191 for basketball). Together they went from 7% of Polymarket's sports markets in 2024 to 21% in 2026.

What this means if you trade

• Treat prices as probabilities. In a standard market a 70¢ contract is a 70% chance, and our data says that is usually about right. It still fails three times in ten.
• In standard markets, long shots are a tax. Contracts below 5¢ returned about $81 per $100 on Polymarket before fees. If you buy them, buy them for fun or because you know something specific.
• Be careful with ladder markets. Temperature brackets, spreads, totals and exact scores were the least reliable prices in our data. Check the spread before you trade, and do not read a 90¢ favourite there as a 90% chance.
• Look for edges far from the event. Prices a month out were slightly too timid, and the error shrinks fast in the final week. That is where patient research can pay, if the money you tie up could not earn more elsewhere.
• Do not expect a big market to be smarter than a small one. It will be cheaper to trade, not necessarily more accurate.
• Shop between platforms. Accuracy is about the same, so the better price is simply the better deal. Our Compare page shows Kalshi and Polymarket side by side, and the EV calculator works out whether a price is worth taking after fees.

Limitations

• The reference time is not always the moment the outcome became known. For a "by December 31" market that resolved early, we measure before the market closed, and closure can lag the event by hours. Short horizons (4 and 12 hours) can include in-play sports trading. Lean on the 1-day and 7-day figures.
• Kalshi is sampled. We used a random quarter of Kalshi's series and up to 80 markets from each (30 for some sports and mention series), so very large Kalshi series are under-represented compared with a full census.
• Volume is measured differently. Polymarket reports dollars traded, Kalshi reports contracts (each pays $1). The same 5,000 threshold therefore admits somewhat smaller Kalshi markets.
• Markets within an event are not independent. Every candidate in an election is its own market. Our confidence intervals resample whole events to allow for that.
• Prices are midpoints, before fees. Polymarket's price history is its midpoint; for Kalshi we used the bid-ask midpoint, or the last trade when the spread was wider than 20¢.
• Survivorship. Markets that were cancelled, voided or resolved 50-50 are left out.
• Ladder markets are identified from their wording and series. A few templated markets may be counted as standard and vice versa.

Methodology in full

• Sources: Polymarket's Gamma API (market list, outcomes, tags) and CLOB price history; Kalshi's public trade API (series, settled and archived markets, hourly candlesticks). Data pulled on 25 October 2026.
• Reference time: the earlier of the market's scheduled end date and its actual close (Polymarket); the earlier of the close time and the expected expiration time (Kalshi).
• Prices: the last hourly price at or before each horizon, if it was no more than 6 hours old (12 hours for the 30-day horizon). We record the price of the first outcome listed (Yes in Yes/No markets).
• Categories: Polymarket's own tags, then the question wording for older markets without tags; Kalshi's series categories. Crypto price markets, temperature markets and "mention" markets (whether a word is said) are kept as separate categories.
• Excluded: "Up or Down" crypto markets (237,442 on Polymarket), Kalshi multi-leg parlays, markets that traded for less than a day, and markets that did not resolve cleanly to Yes or No.
• Ladder markets: temperature-bracket markets, point-spread and total (over/under) markets, and exact-score markets, identified from the question wording (Polymarket) and series (Kalshi).
• Returns by price: each market counted as two contracts (Yes at its price, No at one minus it); the return is total payout divided by total cost for one contract of each.
• Confidence intervals: 1,000 bootstrap resamples of whole events.

Download the data

The dataset is free to use with credit to PredictDesks (licence: CC BY 4.0).

• Market-level data (CSV): one row per market with 100,000+ volume (dollars on Polymarket, contracts on Kalshi), with its category, result and prices at each horizon.
• Summary tables (CSV): every aggregate in this article, by platform, category, horizon, volume and year.

To cite: PredictDesks Research, "How accurate are prediction markets? A study of resolved Polymarket and Kalshi markets", October 2026.

Prediction market accuracy: frequently asked questions

How accurate is Polymarket?
A day before resolution, its favourite won 76.8% of all the markets in our sample, and in its standard markets prices were off by 0.9 points on average. On long-running markets the favourite won 87.7% of the time a month out and 92.8% a day out.

How accurate is Kalshi?
In our sample of 24,772 Kalshi markets, the favourite won 78.4% of the time a day out, and 93.3% on long-running markets. Weighted to the same topics, its Brier score (0.147) was close to Polymarket's (0.150).

Is Kalshi or Polymarket more accurate?
Neither, by a meaningful margin. Their raw averages differ because they list different things. On the same topics, and on the same questions such as Fed decisions, their prices land within a point or two of each other.

Are prediction markets more accurate than polls?
Usually, though they are not independent of polls: traders read the polls and add other information. The best long-run evidence comes from the Iowa Electronic Markets, which beat the polls 74% of the time across five presidential elections. More on the history: a short history of prediction markets.

What is a good Brier score?
Zero is perfect and 0.25 is what you get by always saying 50%. Across all markets a day out, Polymarket scored 0.150; on long-running markets 0.050. The number depends heavily on the questions, so compare scores only on similar markets. More: Brier score explained.

How often does the favourite lose?
A day before resolution, about 23% of the time across all markets, and 34% of the time when the market was priced between 10% and 90%.

Do long shots lose money?
In standard markets, yes: people overpay for small chances of a big win, and selling them ties up a lot of money for a small return. Contracts under 5¢ returned about $81 per $100 on Polymarket in our data. Ladder markets were the exception: their cheap brackets won more often than their midpoint price, though wide spreads eat into that.

Keep reading

• How accurate are prediction markets? The track record
• Brier score explained
• Kalshi vs Polymarket: which is better?
• Prediction market strategies
• How to calculate expected value
• The Kelly criterion for prediction markets
• Insider trading on Kalshi and Polymarket
• Polymarket whales