A candidate trading at 70% is not expected to win 70% of the vote. A candidate trading at 90% has not won. And a late swing in prediction market prices is not evidence that ballots have changed.
Those distinctions can disappear when election odds travel across social media as red-and-blue maps, screenshots and declarations that one party is headed for a sweep.
With five days until the midterms, Kalshi and Polymarket are offering a constantly updating assessment of what their traders expect. That can be useful information.
However, it is not representative polling, an official count or a declaration that a race is over.
An Oct. 26 analysis from the University of Washington’s Center for an Informed Public warns that prediction markets increasingly borrow the visual language of professional election forecasting while presenting a different kind of data. Its concern is not simply whether the favorite wins: It is whether audiences understand what the numbers actually measure.
The percentage is about winning, not vote share
Consider the South Carolina Senate market discussed in DeFi Rate’s latest update.
Republican Sen. Darline Graham was at 72% on Kalshi on Thursday, even as a New York Times/Siena poll showed her tied with Democrat Annie Andrews at 48%. In Florida, Republican Sen. Ashley Moody was trading at 90% while leading Democrat Angie Nixon 50% to 44% in the same polling series.
Those numbers answer different questions.
The polls estimate voter preferences. The winner markets price whether Graham or Moody ultimately wins under the contracts’ rules.
A candidate could receive 50.5% of the vote and win a contract that traded at 90%. Another could receive 49.5% and lose despite having been the market favorite. Neither outcome requires the candidate’s vote share to resemble the probability displayed before the election.
The same distinction applies to conventional forecasting models. A model giving someone a 70% chance of winning is not predicting a 70% vote share, either. Probability assesses uncertainty about an outcome, not the expected margin.
Kalshi itself offers different contracts for winners, margins and congressional control. Treating those products as interchangeable obscures what traders are actually buying.
An upset does not automatically make a probability wrong
A 70% favorite can lose without the outcome being inconsistent with the forecast. The remaining 30% is not decorative.
In a well-calibrated set of forecasts, events assigned a 70% probability should occur roughly 70% of the time across many comparable predictions. The other outcomes should occur about 30% of the time.
That means one favorite losing cannot, on its own, establish that a forecasting method is useless. Equally, one favorite winning cannot prove the number was well calibrated. A 55% forecast and a 95% forecast can both pick the eventual winner while expressing very different degrees of confidence.
The useful question is not only, “Did the market pick the winner?” It is also, “Over many races, did outcomes occur at rates consistent with the probabilities?”
The University of Washington analysis emphasizes that both conventional forecasts and market odds contain uncertainty. Fast updates and precise-looking percentages do not remove it.
Trading volume is not a voter sample
The large dollar figures attached to election markets can make them look more authoritative.
But $47 million in trading does not mean 47 million voters participated, or that the market conducted an unusually large poll. Contracts can change hands repeatedly, and participation is not organized to represent the electorate.
The Center for an Informed Public also points out that choosing the candidate a person expects to win can be entirely different from choosing the candidate that person supports. A trader can dislike a candidate and still buy a position on that candidate winning.
That limits a common interpretation of the markets: they are not direct measurements of national political preference.
Prices may reflect useful analysis of polls, turnout and campaign developments. They can also reflect shared assumptions, incomplete information and trading behavior. Money creates incentives to identify mistakes, but it does not guarantee that mistakes disappear.
The careful description is “traders favor this outcome,” not “voters favor this candidate by this percentage.”
Market odds can change what audiences believe
The distinction matters because the numbers can shape perceptions of a race.
A Brennan Center-funded survey experiment involving 1,994 self-identified registered voters found that prediction-market odds affected which candidate respondents expected to win, even when the polling information pointed elsewhere.
Compared with a hypothetical candidate given 50% Kalshi odds, a candidate at 67% was about 10 percentage points more likely to be identified as the expected winner. At 84% odds, the increase was approximately 20 points.
That finding concerns perceived electability. It does not establish that respondents changed their votes, that turnout changed, or that markets altered an actual election result.
Still, it shows why presentation matters. An audience can read a market price as authoritative evidence about an election even when the underlying product is a tradable contract, not a voter survey.
The University of Washington researchers warn that familiar maps and forecasting-style graphics can blur that distinction.
Election-night movement is not evidence of fraud
Prices will likely change as returns arrive. A candidate who looks strong early can become an underdog as different counties report, more ballots are counted or new information changes expectations.
A market swing establishes that the price changed. It does not establish that votes were added improperly, removed or altered.
The Center for an Informed Public warns that screenshots of changing odds could become material for election-integrity rumors. It also cautions that the relationship can run in the other direction: Rumors about interference can move markets if traders believe they might affect the result.
That creates a risk of circular reasoning. A rumor moves prices; the movement is then presented as proof that the rumor was true.
Evidence for an election claim must come from underlying records and verified reporting, not from whether a contract rose or fell.
Settlement is not an election call
The market favoring a candidate is distinct from the exchange resolving the contract.
Kalshi explicitly says it does not call elections. In its Oct. 6 transparency announcement, the company said its election settlement framework would rely on at least four of eight major news organizations, including the Associated Press, calling a race with none disagreeing. It said recounts or legal challenges could keep markets open until the winner is sworn in. Traders still need to read the rules of the specific contract.
A 99% price is therefore not the same event as settlement. And settlement under a contract’s rules is not the same function as an election authority certifying the vote.
Polymarket contracts likewise have written resolution conditions that users should distinguish from the way an election develops in real time, the University of Washington analysis notes.
Useful signals, not declarations
Prediction markets can contribute to election coverage by showing how traders assess new information and how their expectations differ across races. Their probabilities are forecasts in that limited sense, not promises, vote-share estimates or official calls.
The appropriate response is neither to dismiss them nor to treat them as an oracle.
For readers watching the final days, the central distinction is uncertainty. A market favorite can lose. A tight poll can coexist with a clear market favorite. Millions of dollars can trade without producing a representative measure of voters.
On Nov. 3, the prices will tell us what traders expect next. The ballots will determine who wins.
