The goal of the model is not to tell you who's going to win, exactly. The goal is to only tell you something has an 80% chance of happening if, in practice, it happens about 80% of the time.
Every number on this site is the output of a framework that starts with a baseline estimate of each race based on historical data, then updates that baseline as new evidence comes in — including district, state, and national polls, expert ratings, fundraising, and demographic data. This is applied across all 470 federal races in the 2026 midterms. Then, two million possible election outcomes are simulated to produce the probabilities and ranges you see on every page.
Fundamentals
Before a single candidate-specific poll is conducted, the model already has a pretty good understanding of most races. That prior is built from what we call the fundamentals: how have these voters voted in the past? Do they behave like swing voters or party loyalists? Do they love their incumbent officeholder or are they an underperformer relative to the partisanship of their state or district? Then, we factor in the national political environment, derived from generic ballot polls. The tide that lifts or sinks all boats. For the stuff you can't reliably put into numbers (a candidate with a scandal, a regionally specific issue, just a general vibe), we incorporate race ratings from Cook Political Report, Sabato's Crystal Ball, and Inside Elections. For a safe red or blue seat, the expert signal doesn't add much. But if they're telling us a House seat in a Trump +16 district is competitive and we don't have any polls of it... we should probably hear them out.
Polling
When polls start to flow in, the model perks up. Each poll updates the prior for its corresponding race, weighted by sample size, recency, and the quality of the pollster based on their past performance and documented methodological practices. Here's the thing, though: most House races are never polled, and sometimes a district that isn't on anyone's radar can shock the nation during a wave. For those races, the fundamentals and expert ratings have to carry more than their fair share of the weight. To help, the model also borrows information from polls in states and districts with similar demographics and voting patterns, so even unpolled races benefit indirectly from polling elsewhere.
Simulation
Then we simulate the election. Then we simulate it again. And again. And again. Two million times, actually. Why two million? We (I, really, but we've committed to the royal we at this point) determined that two million was about as high as we could go while keeping the model quick and computationally inexpensive. Because elections are correlated, each simulation draws an outcome for all 470 races simultaneously. A good night for Democrats in Ohio probably means a good night in Michigan too. A national polling miss doesn't just hit one race. We account for this by applying shared shocks at the national, regional, and demographic level, which creates some really funky maps if you look at any one individual simulated outcome. In aggregate, however, they paint a much clearer picture, communicated through win probabilities for each race, seat count ranges for each chamber, and the probability that either party controls the House or Senate.
Calibration
Here's the part where we tell you whether this thing actually works. The model was tested against the last two midterm elections, which were deliberately held out of the dataset during development. After locking down every parameter, we ran it as if we were forecasting those elections in real time, using only the information we would have had at the time. In those simulations, when the model said something had an 80% chance, it happened about 82% of the time. It did better than most competing forecasts in some years and a little worse in others. I don't think it's a perfect model. It's a lot closer to perfect than the 2024 version, that's for damn certain. But we're calibrated, which is not nothing.
Limitations
The model doesn't know everything. There are things that might be marginally clarifying but go beyond the scope of my technical knowledge or my budget for data procurement. There is only so much that can be done when certain pollsters drop a ton of data in a window where others do not. Mid-cycle redistricting smudges the signal from historical voting patterns in a population. Should this not completely blow up in my face and I continue modeling in 2028, I can only imagine which experimental tools end up yielding meaningful improvements. But as Lorne Michaels says of Saturday Night Live, the show doesn't go on because it's ready, it goes on because it's 11:30.
What the Numbers Mean
When the forecast says a candidate has a 70% chance of winning, that means they win in 70% of the simulations and lose in the other 30%. It's not technically a prediction. It's a distribution. Think of it like Doctor Strange peering at millions of possible futures. We looked at them all and can present to you the universe of plausible outcomes.