DN forecast-quality research · 22 November 2026 · 2027 methodology guide
Short answer: Prediction markets can be useful complements to polls, but neither should be declared the universal winner. A meaningful comparison needs identical events, time-matched forecasts, probability calibration, coverage and liquidity checks. The 2026 US midterm evidence shows why a blanket “polls have been replaced” claim would be misleading.
What matters: accuracy without a shared denominator is advertising
A market with contracts on hundreds of races cannot be ranked against a polling average available for only a smaller subset without disclosing the denominator. Likewise, a poll estimating vote share cannot simply be compared with a binary market’s event probability. Both must be translated into an explicit common outcome and forecast time. If the comparison lacks those fields, mark its verdict unscored.
Nine tests for a 2027 forecast
| Test | Required evidence | Common mistake |
|---|---|---|
| 1. Same event and outcome | Matching election race, question and resolution date | A generic approval poll is not an election-win probability |
| 2. Same forecast timestamp | Market price and poll average at the same cutoff | Last-minute prices have more information than older polling |
| 3. Same sample of races | Only comparable constituencies with both data types | Including hundreds of unpolled races favors market coverage but not paired accuracy |
| 4. Probability calibration | Brier or log score for forecasts on settled outcomes | A correct winner call can hide severe overconfidence |
| 5. Liquidity and cost | Order book depth, bid/ask and volume for each contract | A thin market’s last trade is not necessarily a useful probability |
| 6. Poll uncertainty and likely voters | Sample method, mode, weighting, nonresponse and margin of error | Point vote share is not automatically a calibrated outcome probability |
| 7. Selection and publication bias | Track unavailable, tied, uncalled and excluded races | Discarding inconvenient cases alters the reported winner |
| 8. Information source independence | Check whether traders simply mirror the polls | Markets can aggregate poll data rather than independently outforecast it |
| 9. Resolution and regulatory constraints | Check settlement definitions, availability and manipulation controls | Contract rules and geography determine meaningful interpretation |
What the 2026 studies actually suggest
A November 2026 Kalshi-authored research working paper reported well-calibrated prices and improvements in market accuracy as election day approached, but this is vendor research and its own sample and horizon definitions matter. An independent Reason Foundation comparison of 371 paired midterm polls reported that surveys and markets had broadly similar outcome-call accuracy across comparable elections; the small aggregate market edge reflected polls that reported ties. An NBER working paper issued in October 2026 examined historical political bias, finding no systematic partisan overpricing across the author’s election sample while noting favorite-longshot bias. These are distinct findings and should not be merged into one claim of proven superiority.
DN Brier Score Comparison Tool
Enter a single binary-event forecast for each source as probabilities from 0 to 100 and the actual outcome (1 occurred; 0 did not). Lower Brier score means less squared forecast error on this one hypothetical observation. Reliable comparisons need many matched cases.
Five ways a comparison can fool you
- Different times: a market quote on election eve compared with a poll taken a month earlier.
- Different metrics: winner accuracy compared with vote-share absolute error.
- Missing ties: polls that say “toss-up” treated as wrong while market prices are scored differently.
- Different race coverage: unpolled seats included only in the market count.
- Biased sample: analyzing only liquid contracts or settled outcomes convenient to a platform.
How DN should run a reproducible 2027 experiment
Preregister contract and polling selection, exact retrieval timestamps, poll-to-probability conversion if required, scored event outcomes, exclusions, and uncertainty estimates. Record raw market bid and ask rather than a screenshot, the polling source and methodology, market liquidity, and fees. Preserve corrections. Score forecast probabilities rather than only naming a winner. If a source does not provide a comparable probability, do not force a misleading numerical comparison.
Primary sources and method limits
- Kalshi Research, November 2026: midterms calibration study (vendor-authored).
- Reason Foundation, 18 November 2026: independent paired-sample midterms comparison.
- NBER Working Paper 35846, October 2026: political bias in prediction markets.
- American Action Forum: evaluating markets alongside polling.
Methodology: DN has evaluated published analysis and built a comparison worksheet. DN has not itself assembled or re-run a full midterm dataset for this edition. The 2027 theme refers to the research framework’s future use, not 2027 election outcomes.
Disclosure: Market operators and polling organizations have not paid for inclusion. This is educational forecast evaluation, not gambling, investment or trading advice.
Frequently asked questions
Are prediction markets always more accurate than polling?
No. The answer depends on event selection, forecast horizon, probability calibration and method. Recent 2026 comparisons provide mixed but useful evidence.
What is the Brier score?
For a binary event, it is the mean squared difference between the forecast probability and the observed result, where lower is better.
Can I compare a poll percentage to a binary market price?
Not directly. Vote share and outcome probability are different quantities; use a documented conversion model or compare outcome forecasts instead.
What did the 2026 midterms show?
A November 2026 independent paired comparison found polls and markets broadly competitive, with differences depending on tied and covered races.
Does market liquidity matter?
Yes. Thin spreads, low depth and isolated trades can make quotes noisy or hard to execute at the displayed price.
Should I use the latest market odds as a trading strategy?
This article is a forecast-auditing guide, not a recommendation to speculate or trade event contracts.
Is a prediction market’s price a verified probability?
It is an exchange-traded price under contract rules; it may contain costs, liquidity premiums and other distortions.
Has DN independently re-run the 2026 midterm dataset?
No. This article cites published studies and provides a reusable evaluation protocol, but does not claim original election results or a definitive leaderboard.




