Prediction markets are often described as collective forecasting machines. The DN Prediction Market Forecast Accuracy Index measures how much trust a market-implied probability deserves once calibration, Brier score, horizon, question difficulty, liquidity and resolution structure are taken into account.
What Matters
A prediction market price is useful because it gives a probability, not just a winner. But “the favourite was right” is not enough to judge forecasting quality.
A market can get the favourite right while still being badly calibrated. A 90% contract that wins only 75% of the time is overconfident even though it usually resolves correctly.
DN therefore evaluates prediction markets across calibration, Brier score, skill versus baseline, time to resolution, market type, liquidity, and resolution quality.
The core finding is nuanced: prediction markets are often impressively well calibrated in aggregate, but accuracy varies materially by horizon, market type and topic. “The market says 70%” should be read as a forecast with context, not as a fact.
DN Evidence Block
Platform-operated research is useful but not independent. DN distinguishes platform-reported studies from academic and third-party studies and avoids treating any single sample as a universal leaderboard.
The Signal: Accuracy Is Not One Number
There are at least five different questions hiding behind “are prediction markets accurate?”
- Does the favourite usually win?
- Are prices calibrated across probability bands?
- How low is the Brier score?
- Do forecasts beat a sensible baseline?
- Does accuracy survive when comparing like-for-like questions?
These metrics can point in different directions.
Brier Score: The Core Forecasting Metric
For a binary event with forecast probability p and outcome y equal to 1 or 0:
Average the score over many resolved markets. Lower is better. A forecaster that always says 50% scores 0.25 on a balanced set of binary outcomes.
But Brier score also depends on question difficulty. A dataset full of obvious 1% longshots can look extremely good even if the forecaster has little edge.
Calibration: Does 70% Actually Mean 70%?
A calibrated forecaster should satisfy a simple test: among a large group of events priced around 70%, roughly 70% should occur.
This is one reason prediction markets are unusually useful: prices naturally expose a continuous probability that can be checked after resolution.
Kalshi: 2.24 Million Markets Show Strong Aggregate Calibration
Kalshi Research's August 2026 study analyzed 2,243,741 resolved markets across eleven categories from launch through mid-2026.
It reported that Brier score improved from roughly 0.08–0.09 three months before resolution to about 0.02 at close, with reliability diagrams tracking the ideal calibration line closely across most categories.
Kalshi Research — Calibration in Prediction Markets
Because the paper is produced by the platform itself, DN labels it PLATFORM-REPORTED, not independent validation.
Kalshi Microdata: Some Traders Appear to Improve the Market
A separate October 2026 Kalshi Research paper analyzed more than 1 billion trades from October 2025 through September 2026.
Under a strict false-discovery threshold, approximately 0.15% of takers forecast significantly better than the markets they traded against. After those traders acted, the weighted midpoint tended to move toward the eventual outcome.
Kalshi Research — Who Calibrates Markets?
This matters because it suggests “wisdom of crowds” may partly be a mechanism in which a small number of skilled traders move prices and everyone else observes the new consensus.
Polymarket: Independent Princeton Evidence
A 2026 Princeton thesis analyzed 188,509 resolved binary Polymarket markets and 1,875,570 daily price observations from November 2022 through December 2025.
The research used reliability diagrams, Brier-score decomposition and logistic recalibration across market domains and forecast horizons.
Princeton — The Price Is (Almost) Right
The value of this study is not merely that it found Polymarket informative. It also recognizes that correlated markets and different question types complicate naive accuracy claims.
Polymarket's Own Accuracy Page
Polymarket currently reports a Brier score of 0.0599, with headline favourite accuracy of 90.3% one month before resolution and 98.6% four hours before resolution.
Polymarket — Accuracy dashboard
Those figures are useful, but they should not be read as “Polymarket is right 98.6% of the time about genuinely uncertain questions.” The market set contains many near-certain contracts.
Cross-Platform Study: The Headline Numbers Need Context
A October 2026 third-party study analyzed 358,236 resolved Polymarket and Kalshi markets, including 333,464 Polymarket and 24,772 Kalshi markets.
Its most important result was not a decisive platform winner. After weighting both platforms to the same topic mix, Brier scores were very close: approximately 0.150 for Polymarket and 0.147 for Kalshi.
PredictDesks — Prediction Market Accuracy Study
On 24 Federal Reserve meetings priced on both platforms, the markets' probabilities one day before the decision differed by less than one percentage point on average.
Topic Matters More Than Platform
The same cross-platform study found large differences by market type. On Polymarket, tech/science markets had a much lower one-day Brier score than sports, while exact-score, spread and totals markets showed more overconfidence.
Longshots and Ladders Need Extra Skepticism
The 2026 cross-platform study found that templated ladder markets were more overconfident than standard binary questions. Very low-priced longshots also showed evidence of overpricing.
This is consistent with a broader prediction-market lesson: small probabilities are psychologically and economically difficult to price well.
Forecast Horizon Matters
Across the same long-running markets, the October 2026 cross-platform study reported steadily improving Brier scores as resolution approached:
| Horizon | Polymarket Brier | Kalshi Brier |
|---|---|---|
| 30 days | 0.091 | 0.116 |
| 7 days | 0.064 | 0.076 |
| 1 day | 0.050 | 0.047 |
| 12 hours | 0.047 | 0.042 |
| 4 hours | 0.042 | 0.038 |
Closer-to-resolution prices incorporate more information, so better scores are expected. The fair comparison is therefore always at the same forecast horizon.
Macroeconomic Markets: Strong Overall, Uneven by Series
A July 2026 study of 594 Kalshi macroeconomic markets found an overall Brier score of 0.0987. Federal Funds Rate markets were exceptionally accurate in that sample, while unemployment-rate markets were materially harder.
Krause — Calibration and Forecast Accuracy of Macroeconomic Prediction Markets
This reinforces the case for category-level scoring instead of one platform-wide “accuracy” number.
Independent Academic Evidence Also Finds Exploitable Biases
A 2026 peer-reviewed study of 103,003 resolved Kalshi markets found that market prices were informative but uneven across domains. A nonlinear calibration model and machine-learning models using public features improved on the raw market baseline in the study's test design.
International Journal of Financial Markets and Derivatives — Prediction Market Efficiency
That does not prove a tradable edge after fees and market impact. It does show that “market price equals perfect probability” is too strong.
DN Prediction Market Forecast Accuracy Index
DN scores a market forecast using five components:
| Component | Weight | What it measures |
|---|---|---|
| Calibration | 30% | Whether quoted probabilities match observed frequencies. |
| Brier Skill | 25% | Performance relative to a sensible baseline, not just raw Brier. |
| Horizon Quality | 20% | How much time remains until resolution and how much information is already public. |
| Market Quality | 15% | Liquidity, participation, spread and susceptibility to microstructure distortion. |
| Resolution Quality | 10% | Clarity of rules, source and settlement criteria. |
DN Forecast Accuracy Calculator
Prediction Market Forecast Accuracy Calculator
Translate a quoted market probability into a context-aware trust score using calibration error, Brier skill, horizon, liquidity and resolution quality.
Forecast
Historical Performance
Context
DN Forecast Analysis
The forecast should be judged against comparable historical markets, not the headline percentage alone.
How the Calculator Works
The calculator then applies horizon, market-type, liquidity and resolution-quality adjustments. It is intentionally conservative: a 90% market with poor calibration or ambiguous settlement should not receive the same trust as a 90% market with strong historical calibration and objective rules.
Why Favourite Accuracy Can Mislead
Suppose 100 markets are priced at 99% for “No” and all resolve No. Favourite accuracy is 100%, but the exercise says little about performance on genuinely uncertain events.
The 2026 cross-platform study found that removing contracts below 10% and above 90% materially lowered headline favourite-win rates. That is why DN reports forecast difficulty, not just outcome direction.
DN Hard-Question Accuracy
This is not a replacement for Brier score, but it is a useful sanity check because it strips out many trivial near-certainties.
DN Calibration Gap
DN recommends reporting this by probability band and category. A single average can hide longshot bias or overconfidence at the extremes.
DN Resolution Integrity Score
Forecast accuracy depends on settlement integrity. A perfectly priced market can still be operationally bad if the resolution source is ambiguous or the rules are rewritten after trading begins.
DN scores:
- objective resolution source;
- clear deadline;
- unambiguous outcome definition;
- transparent dispute process;
- and whether final settlement matches the stated rules.
What Does a “70% Chance” Actually Mean?
If a platform's comparable 70% markets historically resolve Yes about 70% of the time, the price is well calibrated.
If they resolve Yes only 55% of the time, the market is overconfident.
If they resolve Yes 85% of the time, the market is underconfident.
Prediction Markets vs AI Forecasts
A September 2026 ARCANE benchmark found that a retrieval-free ensemble of large language models underperformed contemporaneous market prices on 73 questions where a market price was available: Brier 0.284 for the calibrated model ensemble versus 0.212 for the market.
ARCANE — Models Without Evidence Lose to Markets
This does not establish that markets always beat AI. It shows the value of continuously updated external evidence: an LLM forecasting without fresh information is at a structural disadvantage against a live market ingesting new information through trader incentives.
DN Forecast Accuracy Dataset
market_id platform question category market_type snapshot_time_utc hours_to_resolution probability outcome brier_score baseline_brier brier_skill_score liquidity_usd volume_usd spread_bps participant_count resolution_source resolution_quality_score calibration_bin observed_bin_frequency calibration_error dn_pmfa_score methodology_version
The Future DN Benchmark
The long-term moat is not another static “Polymarket vs Kalshi” table. DN should maintain a rolling, synchronized forecast benchmark that:
- samples both platforms at the same horizons;
- groups equivalent market types;
- separates politics, economics, crypto, sports and science;
- excludes or separately reports ladder markets;
- tracks Brier score and calibration by probability band;
- records liquidity and spread;
- and preserves the exact resolution rules.
Falsification Test
The DN thesis weakens if like-for-like prediction-market prices fail to beat simple base-rate forecasts over large samples, or if calibration remains poor even in high-liquidity standard markets near resolution.
It also weakens if apparent forecast skill disappears entirely after adjusting for easy questions, correlated markets and resolution selection.
If that happens, prediction-market probabilities should be treated as trading prices rather than privileged forecasts.
Frequently Asked Questions
How accurate are prediction markets?
Large 2026 studies find that major prediction markets are generally well calibrated in aggregate and improve as events approach resolution, but accuracy varies materially by topic, market type and horizon.
Is Polymarket more accurate than Kalshi?
Not clearly on a like-for-like basis. A 2026 cross-platform study found very similar topic-weighted Brier scores, and the two platforms were extremely close on shared Fed-decision markets.
What is a good Brier score?
Lower is better, but the score is meaningful only relative to the difficulty and base rate of the question set. A raw score should ideally be paired with a baseline or Brier skill score.
Does 70% on a prediction market mean the event will happen?
No. It means the current market-implied probability is about 70%. If the market is calibrated, similar 70% forecasts should resolve Yes roughly 70% of the time over a large sample.
Why are prediction markets more accurate closer to resolution?
More information becomes known, uncertainty falls and traders have had more time to incorporate evidence into prices.
Can prediction markets be wrong?
Yes. Even a perfectly calibrated 70% forecast should fail about 30% of the time.
DN Alpha Thesis
The most valuable prediction-market product is not the contract itself. It is the continuously updating probability layer that markets create over uncertain events.
As AI agents increasingly need probabilistic inputs for trading, treasury management and decision-making, a reliable forecast feed could become infrastructure.
Final Verdict
Prediction markets deserve more respect than “online betting” and less reverence than “oracle.”
Large datasets show real forecasting value. They also show systematic variation in quality.
The right question is not “is the market right?” It is:
Continue the DN Prediction-Market Research Stack
Prediction Markets Definitive Guide AI Crypto Citation Freshness Index Exchange Legal Entity MapperPrimary Source Ledger
- Kalshi Research — Calibration in Prediction Markets: 2.24M resolved-market calibration study.
- Kalshi Research — Who Calibrates Markets?: 1B+ trade microdata study.
- Princeton — The Price Is (Almost) Right: 188,509 resolved Polymarket markets.
- Polymarket — Accuracy dashboard: current platform-reported accuracy and Brier score.
- PredictDesks — 358,236-market cross-platform study.
- Krause — Macroeconomic prediction-market accuracy.
- Subramanian — Prediction market efficiency.
- ARCANE — Models Without Evidence Lose to Markets.
Prediction Markets or Polls: Compare the Same Forecast
Decision extension · 13 November 2026. A market-implied event probability, a poll's vote share and a forecasting model's chance of victory are different quantities. Comparing them directly can produce an impressive accuracy claim from incompatible inputs. Define the event and horizon first, then preserve the prediction that was actually available before the outcome.
Match the target and the clock
A poll reporting 55% support does not automatically predict a 55% chance of winning. A model may translate polling, sampling uncertainty, turnout assumptions and other information into a probability, but that transformation needs its own published method and dated output. Do not invent the model after seeing the result.
For each comparison, record the same binary question, geographic scope, resolution source, cutoff and forecast horizon. A prediction made seven days before an event should not compete with a market price observed minutes before resolution unless the study explicitly tests that difference. Preserve the original wording and amendments.
Market observations also need a convention: last trade, midpoint or executable side, with liquidity and transaction charges recorded separately. A stale last trade can differ from a current executable quote. Forecast accuracy and achievable trading profit are different outcomes; a good probability estimate need not create a profitable opportunity after costs.
Use a declared scoring convention
For the binary Brier convention used here, each event contributes the squared difference between probability and outcome, with outcomes coded zero or one. The score is the mean across the matched events, and lower is better. Metaculus's methodological discussion explains this convention and why question selection and difficulty matter; it is not evidence that one source wins this article's unrun comparison. Primary methodological discussion.
Constructed example: two events resolve to one and zero. A fictional market series predicts 0.7 and 0.4, giving (0.09 + 0.16)/2 = 0.125. A fictional poll-based model predicts 0.6 and 0.3, also giving (0.16 + 0.09)/2 = 0.125. This is a tie on two invented events, not a finding about an election, platform or real forecasting system.
The calculator rejects percentages outside zero to one, duplicate event aliases, unresolved outcomes and malformed rows. It does not establish that the supplied predictions share a horizon or resolve under the same rule. Those matching conditions must be documented separately.
DN Matched Forecast Accuracy Comparison
Paste a JSON array of matched, resolved binary events. Each row must contain only event, pMarket, pModel and outcome. Probabilities use 0–1; outcomes are 0 or 1. Poll vote shares are not win probabilities. Use non-sensitive aliases.
Not calculated.
Local arithmetic only. No network access, wallet, trading or persistent storage. The tool does not authenticate predictions, resolution or matching.
Do not let sample selection choose the winner
Define the event population before collecting successful comparisons. Record events lacking a market, poll model or resolution and explain their exclusion. A study covering only widely traded contests can answer a different question from a study of every contest in the original scope.
Several markets on the same underlying event can be highly dependent. Treating them as independent observations can overstate the evidence. Preserve event families and use an uncertainty method appropriate to that dependence before claiming a robust advantage.
A lower average Brier score on the entered sample does not by itself establish calibration, a future advantage or universal superiority. Calibration examines whether stated probabilities correspond to outcome frequencies within an appropriate population. It needs enough relevant observations and cannot be read from one successful 70% prediction.
Separate human, model and market attribution
Market volume, transaction speed or a wallet label does not establish who generated a belief or whether AI caused a price change. A valid attribution study needs the actual controller or disclosed workflow, information available at the forecast cutoff and the role of any human review.
Likewise, a poll-based model may include market information, and a trader may use polls. Calling them independent sources without examining inputs can misdescribe the comparison. Preserve model versions and input policies instead of treating labels as proof of independent reasoning.
| Boundary | Required record |
|---|---|
| Question | Exact binary event, scope and resolution rule |
| Time | Cutoff, horizon and prediction actually available then |
| Probability | Market observation convention or published model probability, not raw vote share |
| Population | Included, excluded, missing and dependent event groups |
| Scoring | Declared convention, complete outcomes and uncertainty treatment |
| Attribution | Known information inputs and controller evidence, with unknowns retained |
Methodology and limits: DN's original paired-evaluation workflow and fictional two-event example. No real matched market-versus-poll cohort, agent-attribution dataset or profitability backtest was collected. Existing article research retains its original source dates; this extension does not claim that its synthetic comparison reproduces those studies.
Change Log
| Date | Version | Change |
|---|---|---|
| 3 Nov 2026 | DN-PMFA v1.0 | Upgraded DN's original prediction-market calibration article into a full forecast-accuracy index with current 2026 evidence, Brier skill, calibration, horizon, market-type and resolution-quality framework. |
| 14 Sep 2026 | Original | Initial DN analysis of “prediction markets are smarter than experts” and calibration evidence. |
Prediction Markets Mentioned in the Benchmark
Commercial disclosure: the links below are affiliate/referral links to platforms discussed in the research. DN may earn compensation if you register or participate through them. Forecast-accuracy analysis and rankings remain editorially independent.
Open Kalshi Open PolymarketMethodology and commercial disclosure: Prediction-market prices are not guarantees. Brier scores, calibration curves and favourite accuracy are sample-dependent. Platform-operated research is labeled as such. This article is for research and education, not financial advice. Where DN uses commercial links elsewhere, editorial inclusion and ranking should remain independent of compensation.




