Market Accuracy
Studies of prediction market accuracy show how well prices forecast real events and how Polymarket and Kalshi compare with other forecasts.
Prediction market prices are commonly read as estimates of how likely an event is, from an election victory to an interest-rate cut. When you use a price to follow an event or compare it with your own expectations, the market's track record helps you judge how much confidence to place in it. That record includes how often favorites win and whether the prices accurately express the uncertainty before the outcome is known.
The timing of a forecast matters to that judgment. For a sports match, a price recorded before kickoff and one recorded during play reflect different information. A strong record late in a match says little about how useful the prices were beforehand. Comparing forecasts made at the same stage across similar events gives a clearer picture of how reliably a market anticipates results.
How accurate are prediction markets?
Prediction markets have often produced probabilities that match how frequently events happen, especially close to resolution. This agreement is called calibration. A study of 1,787 Intrade markets found reasonably calibrated prices near expiration and larger biases more than 100 days ahead. A 2026 analysis of Kalshi and Polymarket also reported close agreement between short-horizon sports prices and outcome frequencies.
A 70-cent Yes contract that returns $1 if an event happens is usually read as a 70% probability. Across many comparable events priced that way, calibration means roughly seven in ten happen and three do not. The three losses are part of what a 70% forecast predicts.
Seven of these ten hypothetical events resolve Yes and three resolve No. The 70% Yes rate matches the probability implied by a 70-cent contract that returns $1 on Yes, an example of calibration.
How is prediction market accuracy measured?
Prediction market accuracy is usually measured through hit rate, calibration, and a probability score such as the Brier score. They answer different questions, so an accuracy percentage and a calibration chart can give different impressions of the same record.
| Measure | What it asks | What to pay attention to |
|---|---|---|
| Hit rate | How often was the selected outcome correct? | Counting the favored Yes or No side of each contract differs from picking one winner in a multi-outcome event. |
| Calibration | Did outcomes occur as often as the probabilities implied? | The chart needs enough comparable forecasts in each probability range. |
| Brier score | How far were the forecast probabilities from the results? | Lower scores are better, but easier questions also produce lower scores. |
For a binary contract, the favorite is whichever side, Yes or No, has the higher price. A correct call on a 51% favorite counts just as much toward the hit rate as a correct call on a 99% favorite. In a tournament, counting each team's Yes/No contract separately can produce many correct No calls even if the highest-priced team fails to become champion.
Team C wins this hypothetical tournament, although Team A had the highest Yes price at 30 cents. Since every team's Yes price was below 50 cents, counting the favored No on each contract gives seven correct calls out of eight.
To calculate a binary Brier score, write the Yes probability as a decimal and the result as 1 for Yes or 0 for No. Square their difference, then average across contracts. A 5% Yes forecast that resolves No scores 0.0025. On this scale, zero is a perfect record and one is the worst possible record. Forecasting 50% on every binary question scores 0.25, whatever happens. Some reports score both the Yes and No probabilities and add both squared errors, doubling these values.
The result being scored comes from the contract's exact proposition. A team reaching a final and winning that final are different questions, even though both concern the same tournament. The resolution rules and named source determine what counts as Yes.
Why can prediction market accuracy rates rise without better forecasts?
Adding more easy questions can raise a hit rate and lower a Brier score even when the pricing is equally good throughout. Consider a hypothetical record of 100 sports contracts, with prices saved seven days before each event:
- Eighty contracts concern unlikely championship winners and cost 5 cents each. Four resolve Yes, exactly 5% of the group. The favored No outcome wins the other 76.
- Twenty contracts concern close matchups and cost 55 cents each. Eleven resolve Yes, exactly 55% of that group. Yes is the favorite in all twenty.
The favorites win 87 times, giving an 87% hit rate. The average Brier score is 0.0875.
The second record adds another eighty 5-cent championship contracts, again with four Yes results and 76 No results.
| Hypothetical record | Favorites that win | Brier score |
|---|---|---|
| 80 unlikely outcomes and 20 close matchups | 87 of 100, or 87% | 0.0875 |
| 160 unlikely outcomes and the same 20 close matchups | 163 of 180, or about 90.6% | About 0.0697 |
The added longshots raise the hit rate and lower the Brier score, while every probability group still matches its observed Yes rate. Listing more distant championship possibilities can improve a venue's overall figures without changing how it prices close games.
The second hypothetical record doubles the 5-cent longshots from 80 to 160 and keeps the same 20 close matchups at 55 cents. Both groups' Yes rates still match their prices, while the hit rate rises from 87% to about 90.6% and the Brier score falls from 0.0875 to about 0.0697.
The close matchups alone have a 55% hit rate and a Brier score of 0.2475. A 55% forecast leaves almost as much room for either result as a coin flip, so its score is close to the coin flip's 0.25.
A benchmark helps reveal whether the forecasts add information. In the original record, 15 of 100 contracts resolve Yes. A constant 15% forecast would score 0.1275, so distinguishing the two groups improves on that reference. Here, the 15% comes from results already known. For a forecast of new events, the base rate would need to come from earlier evidence.
A forecaster can be well calibrated by assigning every question the same base rate, while still failing to distinguish likely outcomes from unlikely ones.
Across 100 hypothetical contracts, market prices match the longshots' 5% Yes rate and the close matchups' 55% rate, producing a Brier score of 0.0875. Forecasting 15% for every contract matches the overall Yes rate but misses the difference between groups, giving a higher, worse score of 0.1275.
How accurate is Polymarket?
Polymarket's sports prices were close to observed outcome frequencies in Nam Anh Le's 2026 study, while its political prices tended to understate favorites' chances.
Direct comparisons also give different results by subject. In the December 2025 version of Reichenbach and Walther's study, Polymarket probabilities had slightly smaller errors than bookmaker probabilities across 492 tennis matches. In Good Judgment's comparison of 25 central bank meetings, its professional forecasters beat Polymarket. Good Judgment sells forecasting services and conducted that comparison itself.
The Polymarket accuracy dashboard reports how often the leading outcome matched the result, using resolved markets and snapshots from one month to four hours before resolution. It also shows Brier scores, which measure probability errors, and a chart comparing quoted probabilities with observed outcome frequencies. Those are Polymarket's own results across its catalog, which contains a much broader mix of questions than either comparison above.
How often is Polymarket wrong?
In CoinDesk's 2025 report on Alex McCullough's Dune dashboard, the reported accuracy was about 90% one month before the event and 94% four hours before it. Those figures imply miss rates of roughly 10% and 6% in that filtered record. The analysis removed markets priced above 90% or below 10% after outcomes were already known but before settlement.
Within a calibrated record, 90% favorites should lose about once in ten, while 60% favorites should lose about four times in ten. A larger share of 90% favorites therefore lowers the overall miss rate, even if the quality of pricing stays the same. Evidence of a pricing error comes from favorites repeatedly losing more or less often than their own prices imply.
Among these hypothetical records of 100 Yes favorites, the miss rate falls from 31% to 19% when the number priced at 90 cents rises from 30 to 70. Prices remain equally well calibrated in both records, with 90-cent favorites losing 10% of the time and 60-cent favorites losing 40%.
How accurate is Kalshi?
Kalshi Research's August 2026 paper reported that the higher-priced Yes or No side was correct in 88.3% of markets observed three months before resolution and 97.2% at close. The study covered more than 2.2 million resolved markets from 2021 through mid-2026, counting each binary contract separately.
Kalshi also measured probability errors with Brier scores, excluding contracts that combine several outcomes. Those errors generally shrank toward resolution, although newly listed, harder-to-predict contracts raised the score one day out. Prices at close can include the result of the event once it is known.
Independent research found an advantage for some economic forecasts. In the 2026 Federal Reserve staff paper by Diercks, Katz, and Wright, Kalshi's mean headline inflation forecast had a root mean squared error of 0.080 percentage points on release days, compared with 0.100 for the Bloomberg consensus. This measure gives large misses more weight, and the lower error for Kalshi was statistically significant. Core inflation and unemployment forecasts showed no significant difference. The market's median and most likely policy-rate forecasts also matched the realized rate one day before every Federal Reserve meeting studied in the 2022–2025 sample.
On headline inflation release days, Kalshi's mean forecast was more accurate than the Bloomberg consensus in the 2026 Federal Reserve staff study. Its root mean squared error was 0.080 percentage points, compared with Bloomberg's 0.100.
Kalshi's prices still showed systematic errors in Bürgi, Deng, and Whelan's study of data through April 2025. Unlikely outcomes were priced too high and likely outcomes too low, a pattern called the favorite longshot bias. Low-priced contracts won too rarely to cover their cost after fees, while high-priced contracts produced small positive returns in that sample. A high platform-wide hit rate can coexist with these errors in particular probability ranges.
Is Kalshi more accurate than Polymarket?
Kalshi scored higher than Polymarket in Clinton and Huang's December 2025 study when they compared election-eve prices. Polymarket scored higher when the authors averaged over a longer period. The venues' separate accuracy reports also cover different catalogs, so their headline percentages compare different forecasting tasks.
The often-quoted ranking of PredictIt at 93%, Kalshi at 78%, and Polymarket at 67% comes from that study's election-eve comparison. Those percentages counted contracts whose logarithmic probability score beat a 50% forecast on the day before the 2024 election. Extending the period back to September included many unlikely events on Polymarket that resolved before Election Day, changing the result. The authors warned that the platforms listed different kinds of questions. Their June 2026 revision reports results pooled across four venues and emphasizes the effect of the mix of easy and close questions.
For an individual comparison, matching starts with the written proposition. A contract about Bitcoin exceeding a threshold at any time during a month differs from one about its price at the month's end. The deadline and named reference price also determine whether the contracts ask the same question.
Bitcoin peaks at $110,900 and finishes at $104,400 in this hypothetical month. A contract requiring it to exceed $110,000 at any time resolves Yes, while one requiring it to finish above that threshold resolves No.
Price differences between related markets can reflect different thresholds as well as disagreement about the same event. When two contracts ask the same question, the outcome determines which probability receives the better score. Repeating that comparison across matched questions shows whether either venue has smaller errors over time.
A comparison also needs the same forecast cutoff and price definition. A last trade can differ from the midpoint between buying and selling quotes, and a forecast recorded after news arrives has information an earlier forecast lacked. Brier.fyi publishes its method for matching questions across platforms and scoring their prices. Counting every trade gives heavily traded markets more influence than assigning equal weight to each question.
For these two hypothetical questions that resolve Yes, giving each question equal weight produces a Brier score of 0.185. Counting every trade lowers the score to 0.045 because the more accurate 90-cent forecast has 900 trades, compared with 100 for the 40-cent forecast.
How accurate are prediction markets in politics?
Political markets have often favored the eventual winner, with a stronger record on clear favorites than close contests. The June 2026 revision of Clinton and Huang's study examined nearly 2,200 political contracts across four exchanges in the final nine weeks of the 2024 US campaign. Among contracts whose Yes price was below 20 cents or above 80 cents on election eve, the favored Yes or No side was correct 96% of the time. For contracts priced between 40 and 60 cents, that share was 55%.
In Clinton and Huang's study of the 2024 US election, the favored Yes or No side won 96% of the time when election-eve Yes prices were below 20 cents or above 80 cents. For prices from 40 to 60 cents, favorites won 55% of the time, close to the 50% expected from a coin flip.
Markets assigned very different chances to the winners of these widely followed contests:
| Event | Forecast before the outcome | Result |
|---|---|---|
| 2016 Brexit referendum | Betfair put Remain at 88% shortly before polls closed, according to contemporaneous reporting. | Leave won. |
| 2016 US presidential election | ElectionBettingOdds.com put Trump's chance around 18% at the start of Election Day, as Reason reported when he won. | Trump won. |
| 2024 US presidential election | Rajiv Sethi recorded election-eve probabilities of 55% on Kalshi and 59% on Polymarket for Trump. | Trump won. |
The 88% Remain price left about a 12% chance for Leave, roughly one in eight. Trump's 18% chance in 2016 represented roughly one in five. Such outcomes should occur occasionally in a calibrated record, although both forecasts receive poor scores for those particular results. The 2024 prices favored the eventual winner without treating the contest as close to certain.
Across a larger record, Le's analysis of Kalshi and Polymarket showed political favorites winning more often than their prices implied.
The level of the race matters too. Undark's account of Sethi and colleagues' 2025 study reports that Polymarket fared particularly poorly on congressional races against forecasts from The Economist, Silver Bulletin, and FiveThirtyEight. The researchers partly attributed the difference to lower trading activity. Presidential, state-winner, and electoral-vote contracts also share many underlying votes, so their results provide less independent evidence than the number of contracts suggests.
Why do studies disagree about prediction markets and polls?
In comparisons of prediction markets and polls, using raw polls or forecasts built from them can change which source scores better. Berg, Nelson, and Rietz found Iowa Electronic Markets vote-share prices closer to final results than raw polls. Erikson and Wlezien projected polls forward to Election Day and found that the poll-based forecasts beat the markets for the same elections.
A poll showing 52% support and a contract showing a 70% chance of winning can agree with each other. The poll's error is measured against the final vote share, while the contract's probability is scored against whether the candidate won.
Candidate A's possible final vote shares spread around the poll's 52% in this hypothetical race. The shaded area above 50% holds 70% of the probability, so that small lead is consistent with a 70% chance of winning.
How do prediction markets compare with expert forecasts?
Markets have outperformed professional economists on headline inflation forecasts and experts predicting Oscar winners. Polymarket fell behind professional forecasting teams on central-bank decisions and specialist models on disease outbreaks.
| Comparison | What the study found |
|---|---|
| Kalshi and professional economists | In the 2026 Federal Reserve staff study, Kalshi's headline inflation forecasts beat the Bloomberg consensus on release days. Core inflation and unemployment showed no statistically significant difference. |
| Oscar markets and entertainment experts | Pathak, Rothschild, and Dudík's 2015 study compared forecasts of Academy Awards winners. Market prices were accurate, and statistical adjustments made them the best forecasts in the comparison. Experts were generally less accurate and less timely. |
| Internal company markets and in-house forecasts | In Cowgill and Zitzewitz's working paper, markets at Google, Ford, and an unnamed materials company improved on expert forecasts. The reductions reached 25% in average squared forecast error, with different benchmarks across firms. |
| Polymarket and professional superforecasters | Good Judgment's 2026 comparison covered 25 central bank meetings. Its forecasters had a lower, better Brier score of 0.102 against Polymarket's 0.126, averaging daily scores within each meeting and then across meetings. Good Judgment sells forecasting services and conducted the comparison itself. |
| Polymarket and disease forecasting models | In Dudley and Magdaleno's 2026 preprint, the CDC FluSight ensemble beat Polymarket on cumulative influenza hospitalizations. Simple statistical baselines also beat the market on monthly measles cases. |
Professional forecasting teams can combine several judgments, update them as information arrives, and learn from previous errors. Traders can use those experts' published work too, so agreement between a market and an expert may partly reflect shared information.
A shared report lies across a trader's laptop and a professional forecaster's notebook. Reading the same research can lead them to similar forecasts.
Are prediction markets more accurate than sportsbooks?
Polymarket had a small forecasting advantage over bookmaker probabilities in one tennis comparison, concentrated in strong favorites and longshots. For close matches priced between 40% and 60%, accuracy was nearly identical. Those findings come from the December 2025 version of Reichenbach and Walther's working paper, which matched 492 ATP and WTA matches through August 2025.
The authors compared probabilities at each match's actual start and removed the bookmakers' margin. Their error measure was the average absolute distance between the probability and the final result. Without the margin adjustment, the two sides of a sportsbook quote usually add up to more than 100%, overstating their combined probability.
In this hypothetical tennis match, the bookmaker's quote implies a 55% chance for Player A and 50% for Player B, totaling 105%. Dividing each probability by that total is one way to remove the bookmaker's margin, bringing the probabilities to 52.4% and 47.6% and their sum to 100%.
Do prediction markets get more accurate closer to the event?
Prices have generally become more accurate as resolution approaches in studies of Kalshi and Intrade. Kalshi's 2026 paper reported Brier scores falling from about 0.087 three months ahead to below 0.02 at close, where lower scores mean smaller errors. The size of that improvement includes changes in which markets were available at each horizon.
Page and Clemen's study of 1,787 Intrade markets showed prices increasingly compressed toward 50% farther from expiration. High-probability outcomes were priced too low and low-probability outcomes too high, with the strongest biases more than 100 days ahead.
Page and Clemen showed how the cost of tying up money could produce this pattern. A trader who spots a small error in a distant contract may have to wait months to collect the return, making other uses of the money more attractive. In their model, that cost pulls prices toward 50%. The authors also noted that later prices may be better calibrated because traders have more information about the outcome.
A sports forecast recorded four hours before resolution may already include part or all of the match. Trading can continue during play, and settlement can follow the final whistle. If the winner was known when the price was recorded, a correct call says little about how well the market forecast the match before kickoff.
A hypothetical football match runs from 3 to 5 p.m., and its contract could resolve at 6, 8, or 11 p.m. Taking a price four hours before those resolution times gives snapshots at 2 p.m. before kickoff, 4 p.m. during play, or 7 p.m. after the result is known.
How well calibrated are Polymarket and Kalshi prices?
Sports prices have often been close to calibrated at short horizons, while political favorites have tended to win more often than their prices implied. These are among the findings of Nam Anh Le's 2026 preprint, using Kalshi and Polymarket data through the end of 2025. Calibration means that events priced at a given probability happen about that often across repeated forecasts.
For a hypothetical group of 80-cent contracts, an 80% Yes rate would match the quoted probability. A 95% Yes rate would mean the prices had understated the chances, while a 65% rate would mean they had overstated them. These gaps appear on a calibration curve, where the number of forecasts in each group helps distinguish a recurring error from chance variation.
The dots mark an 80-cent price, and the ticks mark three possible Yes rates for the hypothetical contracts. The price is calibrated at 80%, understates the chance by 15 percentage points at 95%, and overstates it by 15 points at 65%.
Calibration in political, sports, weather, and crypto markets
The findings differ by both venue and time remaining until resolution:
| Subject and venue | Time before resolution | What the study found |
|---|---|---|
| Politics on Kalshi and Polymarket | Most horizons examined | Prices were too close to 50%. Favorites won more often, and longshots less often, than prices implied. |
| Sports on Kalshi | Within 48 hours, compared with more than a month | Prices were close to calibrated in the short window. Farther out, they understated favorites' chances. |
| Sports on Polymarket | Across the usable time ranges | Prices were close to calibrated overall. |
| Weather on Kalshi | Within 48 hours | Prices were too extreme, overstating the difference between likely and unlikely outcomes. |
| Crypto on both venues | Across the study's available contracts | Prices were broadly close to calibrated, with some underpricing of favorites. Different contract types entered at different horizons. |
In these hypothetical markets, a 20-cent price is too close to 50% when 10% of contracts resolve Yes, and too extreme when 30% do. The dashed line marks prices that match how often events happen.
Even within crypto, forecasting Bitcoin's price five minutes from now is a different task from assessing whether an exchange will list a particular token. The first concerns a price movement over a short interval, while the second concerns a company's decision. A category-wide accuracy figure can combine both and give little detail about either.
Does liquidity make Polymarket and Kalshi more accurate?
Low liquidity can leave a price sensitive to a few traders' judgments, and some studies have linked low activity to weaker forecasts. A small order can move the price sharply when few other orders are available. Higher trading volume, which counts completed transactions, has not consistently meant better accuracy across the markets studied.
Kalshi's own research reported better calibration with higher volume and more participants near resolution. The 2024 election comparison reported by Undark linked Polymarket's weaker congressional forecasts partly to low trading volume. In a broader sample of political contracts, Clinton and Huang's December 2025 study found no relationship between trading activity and accuracy after accounting for the exchange.
Dan Schwarz's analysis of Kalshi and Polymarket markets from 2024 to early 2026 excluded sports, crypto, entertainment, and similar categories. Among the remaining markets on both venues, higher volume went with better accuracy for contracts lasting at least 90 days. For shorter-lived contracts assessed seven or thirty days before resolution, the relationship was not statistically significant.
In a 2008 working paper on TradeSports, Paul C. Tetlock reported that greater liquidity did not reduce, and sometimes increased, discrepancies between prices and outcomes on short-horizon sports and financial contracts.
The open interest shown alongside trading volume measures money committed to outstanding Yes/No contract pairs. It can remain high without much recent trading. Neither figure measures the quality of the participants' information. The order book shows the prices and quantities available to trade now, and a busy trading history can coexist with few orders near the quoted price.
In these hypothetical screener rows, the league-title event has $2.6 million committed to outstanding contracts despite just $6,000 traded in the last day. The Bitcoin event has $1.4 million in daily volume and $180,000 in open interest.
Repeated trading by a few accounts can add volume without adding independent judgments. Deliberate price pushing and artificial trading activity can also distort both the price and the record of trading.
Why do prediction markets work?
Prediction markets give people a financial reason to bring information into a public price. Someone who thinks a contract is underpriced can buy it, while someone who thinks it is overpriced can trade the other side. As offers are taken and revised, the price can incorporate what those traders know.
The information can come from public material such as an injury report or an earnings release. Traders differ in how quickly they read it and how well they connect it to the contract. Accuracy can improve through the actions of a relatively small informed group.
In a 2026 interview about their Polymarket working paper, Gómez-Cram and colleagues reported that about 3% of accounts showed statistically significant skill under their test. They compared actual profits with simulations that randomized trade direction. The skilled accounts reacted quickly to public news and traded against pricing inconsistencies and recurring biases. An account could make money without clearing that statistical threshold, since some profitable results could also arise through luck.
Three colored squares out of 100 mark the roughly 3% of Polymarket accounts that showed statistically significant skill in Gómez-Cram and colleagues' study, as reported in their 2026 interview. The grey squares mark accounts that did not meet the test's threshold.
A simple average of forecasts gives each submitted opinion equal weight. A market price responds to trades and available offers, so an informed trader who trades more can have greater influence. In the Polymarket study, the authors argued that less-skilled participants gave informed traders a reason to participate by trading with them at unfavorable prices.
Correcting a price still has costs. In his account of trading on Augur after the 2020 US election, Vitalik Buterin described contracts paying $1 if Trump lost that stayed near 85 cents for more than a month after the result was declared. Buying them tied up money until settlement. He identified that capital cost, technical difficulty, and concerns about settlement among the obstacles to correcting the price.
Does a high prediction-market win rate mean a profit?
A high win rate can leave a trader breaking even or losing money because likely outcomes cost more to buy. Profit depends on the prices paid and received, together with trading costs.
Suppose you buy one Yes contract in each of 100 different markets at 70 cents. Seventy resolve Yes and return $1 each, while thirty return zero. You spent $70 and received $70 before fees. Across the group, outcomes occurred exactly as often as the prices implied, yet you broke even. Fees would make the result negative.
The same arithmetic applies to buying No whenever most contracts resolve No. In a hypothetical set where 90 of 100 No positions win, paying 90 cents for each costs $90 and returns $90 before fees. The high success rate is already reflected in the purchase price.
Among 100 hypothetical Yes positions bought at 70 cents, the 70 winners gain $21 and the 30 losers lose $21. Buying 100 No positions at 90 cents produces 90 winners gaining $9 and ten losers losing $9, so both records break even before fees.
A trader's profit record also leaves out part of the forecast. Buying at 40 cents records the price someone accepted, without recording how likely they thought the event was. Position sizes, exits, and costs affect profits too. The transactions behind a Polymarket leaderboard total show how much a trader paid, how much they received, and which positions produced the result. A screenshot of wins leaves those details out.
Recording your probability before trading makes it possible to compare your judgment with the market after the result. The trading record then shows whether any forecasting advantage survived the actual purchase and sale prices and costs.
Brier Score
How probability scores are calculated and why different records produce different numbers.
Calibration
How to compare forecast probabilities with the frequency of real outcomes.
Prices and Probabilities
What a contract price implies and where recurring pricing biases appear.
Markets vs. Polls
How election markets, opinion polls, and forecasting models answer different questions.
Market Resolution
Polymarket and Kalshi settlement rules explain why payment can wait after an event ends and why similar contracts can resolve differently.
Wisdom of Crowds
The wisdom of crowds explains how combining forecasts can improve accuracy and why shared mistakes can survive even in a large group.