Who’s reading the US midterms 2026 correctly? Markets, models and polls tell different stories

The upcoming 2026 US midterm elections are entering their final stretch, but there is no single way to measure where the contest stands. Prediction markets such as Kalshi and Polymarket are continuously updating their prices based on what traders think will happen. Polling averages measure where voters currently stand. Silver Bulletin's FLIPR model combines polling with political fundamentals and expert race ratings to simulate possible outcomes. The Cook Political Report and Sabato's Crystal Ball, meanwhile, assess individual races using expert judgment and competitiveness ratings. All are looking at the same election. Yet they can produce noticeably different pictures. That difference is not necessarily a contradiction. The systems are answering different questions, using different inputs and operating on different timelines.

Prediction markets: What are traders pricing in?

Kalshi and Polymarket approach elections through prediction markets. Rather than asking voters directly, they provide markets in which participants trade contracts tied to future outcomes. Kalshi's own September 21 overview said traders had moved toward pricing Democratic control of both chambers, while stressing that market expectations can change as polling and campaign developments emerge.

Polymarket similarly expresses expectations through market prices. Its markets can cover everything from control of the House and Senate to individual races and broader election scenarios. A contract priced around 60 cents, for example, can be interpreted as the market assigning roughly a 60% implied probability to that outcome, subject to the market's rules and liquidity. That makes prediction markets different from polls.

A poll asks voters what they currently intend to do. A prediction market asks traders to put a price on what they believe will happen on Election Day. The distinction can become particularly important when a major development occurs. A new poll, debate performance, candidate controversy or economic development can move market prices before a large batch of new public polling becomes available. But market prices are not guaranteed forecasts. They reflect the participants, information and trading conditions in that particular market.

Silver Bulletin: Turning polls and fundamentals into simulations

Silver Bulletin takes a more traditional modelling approach, although its FLIPR system is considerably more than a simple polling average. FLIPR, short for Forecast with Leading Indicators, Polls and (Expert) Ratings, combines polling, non-polling fundamentals and outside expert ratings. The model then runs thousands of simulations to estimate the range of possible congressional outcomes.

Its September 28 update showed Democrats with a 9.2-point advantage in its likely-voter-adjusted generic congressional ballot average. The model also had Democratic candidates holding polling leads in six Senate states it identified as particularly important to the Senate tipping point: Texas, Maine, Ohio, Iowa, Michigan and Alaska. That does not mean FLIPR simply takes a 9.2-point national lead and converts it into seats.

The model accounts for the fact that congressional races are interconnected but not identical. It uses individual race polling, candidate and incumbency factors, the national political environment and other fundamentals. That is one reason a model can produce a different result from a prediction market even when both are looking at broadly similar political developments. Silver Bulletin's methodology notes that FLIPR typically uses 40,000 simulations and models House, Senate and gubernatorial elections jointly.

Also Read: RAF airbase incident: Netanyahu claims Israel warned Britain about Iranian terror plot

Cook Political Report: The race-by-race approach

The Cook Political Report uses another framework altogether. Rather than presenting every race as a probability, Cook classifies contests on a seven-point scale: Solid Democratic, Likely Democratic, Lean Democratic, Toss Up, Lean Republican, Likely Republican and Solid Republican. Cook says its ratings consider factors including a state's political makeup, candidate strengths and weaknesses, and the political environment at both the state and national levels.

As of September 29, Cook's Senate ratings had seven races classified as Toss Ups: Alaska, Iowa, Maine, Michigan, New Hampshire, Ohio and Texas. That is useful information, but it should not be read as equivalent to a prediction-market probability. A Toss Up means Cook considers the race highly competitive; it does not mean the race is mathematically 50-50. This is an important distinction when comparing election coverage. A headline that places a ‘Toss Up’ next to a market's 55% or 60% probability can make the two systems appear to disagree more than they actually do.

Sabato's Crystal Ball: Another expert lens

Sabato's Crystal Ball provides another race-rating framework, using expert analysis to classify competitive contests rather than relying solely on a statistical probability. Like Cook, its value lies partly in the race-by-race context. Analysts can consider candidate developments, state political conditions and the broader electoral environment alongside polling. That means an expert rating can move even when the polling picture has not changed dramatically if analysts believe other factors have altered the competitiveness of a contest. The broader point is that expert ratings are designed to communicate the state of individual races in a relatively accessible way. They are not intended to function as trading prices or as the output of a single statistical simulation.

Why can all five disagree?

The biggest reason is simple: they are not measuring the same thing. A poll measures voter sentiment at a particular point in time. A prediction market measures the price traders are willing to put on a future outcome. A statistical model processes multiple data sources and simulates thousands of possible elections. An expert rating assesses the competitiveness of individual races.

There is also a difference in how uncertainty is treated. Suppose several Senate races are polling within a few points. A model may simulate those races thousands of times and calculate how often each party reaches a particular seat threshold. A prediction market may assign a price to Senate control based on traders' collective expectations.

Cook may simply classify several of those races as Toss Ups. All three can therefore be describing the same underlying uncertainty without producing the same number.

So which number should voters look at?

There is no single number that answers every question. If the question is what are traders pricing right now?, prediction markets provide that information. If the question is what does a statistical model calculate after combining polling and other indicators?, FLIPR provides a model-based answer. If the question is which individual races are considered competitive?, Cook and Sabato provide race-rating frameworks.

And if the question is where voters currently stand, polling remains the direct measure of voter sentiment, although polls themselves carry sampling and methodological uncertainty. The most useful way to read the 2026 election landscape, therefore, may be to understand what each system is actually designed to measure rather than treating them as competing versions of the same forecast.

With Election Day approaching on November 3, market prices can continue moving, new polls can change averages, models can update their simulations and race ratings can shift as campaigns develop. The disagreement between them is not necessarily a flaw. It is a reminder that an election is not a single number, and that different tools can turn the same political information into different measures of uncertainty.