Methodology
The Precision Funnel™: how the Precision Combined forecast works
Every race is different, so no source has a fixed weight. Everything available goes into the funnel, each source is weighted by the evidence it carries for that race, and every race page shows the weights and why.
The Precision Funnel™
Everything available on a race goes in. What comes out is weighted by evidence.
- Public pollsVetted for size, recency, voter screen, sponsorLive
- Prediction markets, priced by money at riskWeighted by dollars traded; thin markets left outLive
- Local and regional sourcesNewspaper, radio, university polls, same vettingLive
- Attention and momentumNews and public-interest trend, gated to 5-15%Live
- our own member pollWeighted to the voter file; trust earned by track recordComing
Out comes
Precision Combined
One win chance per race, every weight and reason printed
Steps 1 to 4 run today; step 4 carries weight only in races where its gate opens. our own member poll is built into the rule and switches on when it is fielded; until then it carries no weight and race pages say so.
How the Precision Funnel™ builds the Precision Combined number
Everything available on a race goes into the top of the funnel. No source has a fixed percentage: each one’s weight comes from how much evidence it carries for this race, by the published rule on the right.
- Public polls. Every public poll of the race, national, state or local, is vetted before it counts: how many people it reached, how recent it is, whether it screened for likely voters, and whether a party or campaign paid for it. Campaign internal polls are left out.
- Prediction markets, priced by money at risk. For each market we add up the dollars traded (contracts times price). A market with under $10k traded, a spread wider than 10¢, or a 15-point jump on under $25k is left out, because one buyer can set that price.
- Local and regional sources. Newspaper, radio and university polls get the same vetting as step 1. Today they arrive through the public poll feed; we are adding the local sources national feeds miss. In races national pollsters skip, such as state cabinet offices, they may be most of what exists.
- Attention and momentum. News volume and tone (GDELT) and Wikipedia pageviews for the two leading candidates, this week against the three weeks before. Attention is not votes, so it counts only through a gate: two independent sources with enough volume, all moving the same way. Then it carries 5% to 15% of the number; otherwise 0%, shown beside the number. Search interest and social platforms join as they are wired.
- our own member poll. Our panel, weighted to the voter file on party, age, sex, education, income and region. It starts with the same trust as outside polls and earns more only through a scored track record.Not live yet
- The number. Market prices are scaled to add to 100%, a poll lead becomes a win chance with a bell curve 6 points wide, and the sources are averaged with their weights. Every race page prints the weights it got and why.
Full methodology →Why blend? Read the research →How we’ll be scored →
The weighting rule
Weight = trust × evidence, then scaled so the sources used add to 100%.
| Source | Trust | Evidence |
|---|---|---|
| Kalshi, Polymarket | 1 | Dollars traded ÷ (dollars traded + $50k). $50k counts half, $1M about 95%. |
| PredictIt | ½ | Fixed at ½, since PredictIt publishes no volume and caps each contract at $850. |
| Public polls | 1 | Grows with each vetted poll: full credit at 600 respondents, half credit every 21 days of age, less for registered voters or all adults, half for a partisan sponsor. |
| Precision member poll | 1 | Effective sample ÷ 800 (max 1) × share of demographic groups that match the voter file. Left out under 200 respondents. |
So a Senate race with deep markets and many polls lands near an even split, while a race no market lists and only one local paper polled leans almost entirely on the polls that exist. The constants above are the same for every race; changing one is a new method version, dated and logged. Trust values will be refit on our own scored history and the change published. RealClearPolitics is linked for comparison but not used: RCP has no public data feed and its terms don’t allow republishing its averages. Market prices reflect traders’ money, not a poll of voters. PredictIt data is shown under its non-commercial license. The Funnel never uses a number we did not measure or a source we cannot name.
One rule, different races
How the Precision Funnel™ splits the weight
- Kalshi
- Polymarket
- PredictIt
- Public polls
- Precision member poll
Big Senate race, today
Kalshi $2.6M and Polymarket $830k traded, PredictIt listed, eight recent likely-voter polls
31%30%8%31%Market money behind the number: $3.4M
Same race, once our panel is fielded
As above, plus a hypothetical Precision member poll of 1,000 matched to the voter file
24%23%24%24%Market money behind the number: $3.4M
Mid-tier race
Kalshi $150k and Polymarket $48k traded, three recent polls
31%20%10%38%Market money behind the number: $198k
State cabinet race nobody else polls
No markets, one local newspaper poll of 400 registered voters, a hypothetical Precision member poll of 800
30%70%A market that just opened
Kalshi has only $3k traded, so one big order set its price, Polymarket $48k, one recent poll
48%52%Kalshi: left out, $3k traded, under the $10k floor
Market money behind the number: $48k
Illustrations of the weighting rule, computed by the same code that weighs live races. The race profiles are examples, not real races, and we have not fielded a poll yet: the Precision bars show the weight a panel of that size would earn, not any result.
Step 4 in detail
The attention bucket and its gate
Close races are often decided by the last few points, and coverage and public interest can shift before polls catch up. They also measure attention, not votes: a scandal draws coverage too. So this bucket is small, gated and fully published.
- 1Measure. For the race's two leading candidates we compare the last 7 complete days with the 21 days before: did one candidate's share of news articles or of Wikipedia pageviews rise, and did their news tone improve against the other's? A 15%-point swing in share, or a 3-point swing in the tone gap, is a full signal (+1 or −1).
- 2Require volume. News counts in full at 20 articles a day across the two candidates and is left out below half that; pageviews count in full at 2,000 a day, with the same floor.
- 3Gate. The bucket counts only when at least two independent sources (news and Wikipedia today) have the volume, at least two of them moved the same way, and none moved the other way. Otherwise it carries 0% and is shown beside the number.
- 4Weigh. When the gate opens the bucket takes 5%, rising to 15% as the signals get stronger and better supported; every other weight is scaled down to make room. Inside the bucket the race is the Funnel's own number nudged toward the candidate with momentum, so in an even race the bucket can move the number by about 3.5 points at most.
| Signal | Sub-weight | Status |
|---|---|---|
| News volume | 25% | Measured |
| News tone | 20% | Measured |
| Wikipedia pageviews | 25% | Measured |
| Search interest (Google, via TrendsAPI or SerpApi) | 15% | Planned; its share goes to the measured signals |
| Social volume and sentiment (Bluesky, X, TikTok) | 15% | Planned; its share goes to the measured signals |
News data from the GDELT Project (gdeltproject.org); pageviews from the Wikimedia Foundation. The full list of sources we use or evaluated is on the data sources page.
What we add
Our capabilities: reaching the races no one else polls
Local papers have cut their polling and many state and down-ballot races go unpolled. These are the tools we are putting in place to measure them. None has produced a published number yet; each will be disclosed with the race it informs.
Voter-file dataBuilding
We are building on the national file of registered voters, joined to public demographic and consumer data: party registration, age, sex, education, income range, region and up to about 200 other attributes. That lets us recruit the right people, weight every poll to the actual electorate, and report results by county, district, age, sex or party.
A broad, targeted member baseLaunching Dec 2026
Members verify once and earn Precision Tokens for completed polls and for friends who join and verify. Paired with targeted digital, social and traditional media campaigns, that lets us fill the groups a race needs quickly, including the small states and down-ballot races national pollsters skip.
People on the groundPlanned
Local field contributors in the areas we cover, reporting what they see: events, yard signs, turnout efforts and the talk at the local level. Their reports feed the momentum signal, labeled as such, never the poll numbers.
Search and social media scoringBuilding
News volume and tone and Wikipedia pageviews are stored for every candidate now. Search interest, Bluesky, X and TikTok hashtag volume are next, each scored for volume and direction. Like news, they count only through the published attention gate and never for more than 15% of the number.
Where the numbers come from
- Kalshi
- CFTC-regulated U.S. exchange. Last trade price for each candidate or party, read from Kalshi's public market data.
- Polymarket
- The largest prediction market by volume. Last price for each outcome, read from Polymarket's public data feed.
- PredictIt
- Long-running U.S. market with an $850 per-contract cap. Last trade price, shown under PredictIt's non-commercial data license.
- Public polls
- National, state and local polls from VoteHub (CC BY 4.0): the latest poll from each pollster, widening from 30 to 90 days until enough exist, each vetted for size, recency, voter screen and sponsor.
- Precision member poll (not fielded yet)
- Our own panel, weighted to the voter file. Built into the rule with zero weight until the first poll is fielded.
- Attention: news and public interest
- Daily news article counts and tone from GDELT, and Wikipedia pageviews, for each race's two leading candidates. Counts 5% to 15% only when its gate opens (below); the raw series are on every race page.
Market prices refresh about every 30 minutes. Every source is credited and linked on each race page, with the time it was read. Candidates are matched across platforms by party label; when a platform lists a candidate without a party, we match the name against the platforms that do label it.
What this method can get wrong
- Markets can be thin. A race with little trading can move on one large order, so markets under the money floor are left out and the rest count in proportion to the dollars behind them.
- Dollars traded is a lifetime total as each venue reports it, not today's activity, so an old market that has gone quiet still counts for its past volume.
- Polls are vetted on what pollsters publish in the feed: sample, dates, voter screen and sponsor. Interview mode and question wording are reviewed by hand when the pollster publishes them, not yet scored automatically.
- Long shots tend to be overpriced on every platform. Scaling prices to 100% removes the spread but not that bias.
- Polls lag the news and many races have few or none. When a race has no recent polls, the blend uses the markets alone and says so.
- Races with no market and no poll show our editorial rating only, with no combined number, and are labeled that way.
- RealClearPolitics is linked for comparison but not used: it has no public data feed and its terms bar republishing its averages.
We record every forecast daily and grade it after the race is called. See how we will be scored and the research on markets versus polls.
The Precision Panel: our own polling
We have not fielded a poll of our own yet. When the member panel launches, this is how it will work.
Verified, unique adults
Every member confirms a U.S. mobile number and passes identity and uniqueness checks before any reward is paid. Bots, duplicates and VPN farms are screened out.
Quality checks on every response
Attention checks and flags for speeding and straight-lining. Responses that fail are dropped and not paid.
Weighting to the country
Results are weighted to Census benchmarks for age, sex, race, education, region and past vote, and the weighting targets are published with each poll.
Rewarded for honest answers
A token bonus scored with Precision Candor Score™, our implementation of Bayesian Truth Serum (Prelec, 2004). Paying people does not make them honest; verification plus a score that favors honest answers over invented ones is the defense. Details below.
Joining the blend
Panel results enter the Precision Funnel™ with the same trust as outside polls, weighted by their effective sample and how well they match the voter file. More trust comes only from a scored track record, published here.
Precision Candor Score™
How the member bonus is scored
A completed poll pays two ways. The cash fee is for finishing. It is the same for every member who passes the attention checks, and it is not scored. The Precision Token bonus is scored, and it is capped at twice the usual award for that poll. No completed response earns less than the usual award.
To be scored, a member does two things: gives an answer, and guesses how the rest of the panel will split.
The answer score
Was that answer more common than members expected, or less? An answer that beats expectations earns more; one that falls short earns less. The bonus does not simply reward whatever most people said. A minority view can still earn it, if more members held it than the panel predicted.
The prediction score
How close was the guess to the actual split? A perfect guess changes nothing. A guess that is far off reduces the bonus. It never increases it. Two members who gave the same answer can earn different bonuses, because one judged the panel better.
The two scores are added and ranked inside that poll. The highest score receives twice the usual token award. The lowest receives the usual award. Tokens from this bonus are locked on the same terms as every other member token. Polls with too few completed responses are not scored, because the shares would be too noisy to pay on.
Example
100 members are asked whether the border is their top issue. The panel predicted 45% would say yes; 60% did. “Yes” was surprisingly common, so those answers rank at the top and earn up to 2× the poll’s tokens. “No” was rarer than predicted, so those answers keep the usual award. Everyone who finished gets the same cash fee.
Precision Candor Score™ is our implementation of Bayesian Truth Serum, published by Drazen Prelec of MIT in Science in 2004. We did not invent the rule; what we built is how it is applied, calibrated and paid on our panel. A later lab test salted recognition questions with made-up items (Weaver and Prelec, 2013). People paid by the rule claimed to recognize 13% of the fakes, against 24% for controls. When also paid to exaggerate, they claimed 16%, against 55% for people paid to exaggerate without the rule. That test used quiz items, not an election. We will publish our own results once the panel is in the field.
This bonus rewards an honest answer over an invented one. It does not make the panel a mirror of the country. Quotas and weighting do that, and those weights are published beside each result. Whether a forecast was right is scored separately, in public, after the race or event is decided.
The formulas, for the curious
Answer (information) score for the option k the member chose:
I = log( x̄k / ȳk )
x̄k is the share of members who actually chose k. ȳk is the geometric mean of what members predicted that share would be.
Prediction score, a logarithmic proper scoring rule on the member’s forecast y:
P = Σk x̄k · log( yk / x̄k )
Zero when the guess matches the result, negative otherwise.
The response score is I + P. Scores use leave-one-out shares, so a member’s own response is excluded from the shares used to score them. Predictions below 1% are raised to 1% and renormalized, so a blank guess cannot break the logarithm. Responses are ranked within the poll and mapped linearly to a token multiplier from 1× (lowest) to 2× (highest). Cash is never multiplied.
In development
Proprietary algorithm librariesIn development
Our own modeling libraries, now in development: a Bayesian blend that weights each source by its measured accuracy for that type of race, and multilevel regression (MRP) for state and district estimates from panel data. Nothing from them feeds the live number until it is published here with a version date.
Additional field modesIn development
SMS-to-web surveys, interactive voice response (IVR) and selective live telephone interviews for hard-to-reach groups, to supplement the online panel.
Every poll we release will disclose
- Sponsor, if any, and who paid for the poll
- Field dates, mode and full question wording
- Sample size, population and margin of error
- Weighting targets and key crosstabs