In the Silver Bulletin/FiveThirtyEight tradition, we’ve always given names to our sports models (PELE, ELWAY) — but never to our election forecasts. I suppose we’d like to think that our presidential model is renowned enough at this point that it doesn't necessarily need any further identifier. But we’ve decided we do need a name for what we informally call our “midterms model.”
That’s because “midterms model” is actually a misnomer. The model forecasts both Congressional and gubernatorial races, but it’s trained on both midterm years like 2022 and years where these races coincide with presidential elections.
“House, Senate and gubernatorial model” is a mouthful, however. So now we’re calling our model FLIPR 🐬 instead.1 In line with our tradition of silly backronyms, FLIPR stands for Forecast with Leading Indicators, Polls and (Expert) Ratings.
We like this name because it’s sort of a triple entendre:
The principal goal of these elections is flipping individual seats, and ultimately the House and Senate.
FLIPR is a hat tip to the inherently probabilistic nature of the forecast — we run 40,000 simulations every time the model is updated. It’s not as simple as “flipping coins”, but the simulations are necessary because of the complex relationships between different types of races.
And the FLIPR acronym is also rather descriptive. FLIPR essentially consists of three different layers stacked on top of one another: polls, non-polling leading indicators (e.g., fundraising, incumbency) or what we sometimes call “fundamentals”, and then a final layer of race ratings from longstanding outside expert groups like the Cook Political Report.
In fact, there are three versions of FLIPR that correspond to each layer:
Lite derives as much information as possible from polls alone;
Classic is polls + fundamentals;
And Deluxe is polls + fundamentals + expert ratings.
While Deluxe is our default version, almost all of the charts and tables on the FLIPR landing page allow you to, uh, flip back and forth between these versions. The way FLIPR is designed, the expert ratings layer is probably the least important — the proverbial “icing on top.” But if you want a forecast without any expert ratings, we have you covered with Lite and Classic.
FLIPR is part of the FiveThirtyEight model lineage
FLIPR has a new name — but is it a new model? I’d say probably not. If you’re a former FiveThirtyEight reader — I retained the rights to all of the models I designed for them per my deal with Disney — you’ll recognize that some of what I’ve just described sounds familiar. In fact, some of the code in FLIPR just celebrated its 16th birthday, having survived since the first midterms forecast I built back in 2010 when FiveThirtyEight was stationed at the New York Times.
We think this track record is one of the things that makes FLIPR unique — competing models come and go, succeed and fail, but few have been around long enough to truly have any sort of long-term track record. And our midterms forecast performed quite well in both 2018 and 2022.
However, FLIPR and its predecessors have a lot of moving parts. And we’ve made refinements every year. While the basic foundation is similar to 2010, nearly every component of the model has been swapped out for a new version at some point. It’s a little bit like the paradox of the Ship of Theseus: if every component of a model has been upgraded at some point, is it still the same model?
The model underwent a near-complete refit in 2018 when the Lite/Classic/Deluxe concept was introduced. It also underwent another suite of changes in 2022 to reflect increasing political polarization. And even though we really liked the 2022 version of the model, we wound up making more changes and improvements this year (2026) than I was originally expecting — rather than just “turning it back on.”
These changes have a common theme: they don’t change the underlying structure or philosophy of the model so much as seek to execute on our original design with more precision, robustness, and internal consistency.
I have complicated feelings about AI coding tools — and I very much disagree with the idea that they’re a substitute for human judgment. But as a complement to human judgment, they’re pretty nice. Probably, in fact, they’re better for reviewing and refining an existing model than for building one from scratch. Basically, every single line of code from our 2022 model2 has been audited and, in many cases, improved.
Many of these improvements are not highly visible: they concern things like faster program runtime, better error detection, and better “code hygiene” that makes FLIPR easier to audit and update.
But having software that speeds up some of the most tedious aspects of coding also freed up time to test out different hypotheses and clear a long list of items off of our punch list. Some of these upgrades fell under a longstanding category of “would be nice to have, if only we had the time.” For example, FLIPR now explicitly simulates ranked-choice voting, as used in Alaska and Maine. Building that code took a couple of hours and relied heavily on our expertise; it’s not nearly as simple as telling an AI coder “ok do ranked-choice voting now.” It probably would have taken a couple of days instead of a couple of hours in the world before Claude and ChatGPT, however.
Still, ranked-choice voting is a fairly low-stakes problem since the other 48 states don’t use this process. Other questions we asked of the data are more consequential. For instance, as elections become more predictable on the basis of overall partisanship with less ticket-splitting than ever, while at the same time the polls have had some rough years3, should our model increase the relative weight on “fundamentals” as opposed to polls? The answer is basically yes4, and FLIPR now does that. Another question: does the political environment now “lock in” more quickly, such that indicators like the generic ballot are more predictive even months ahead of time than they once were? Again, the data says “yes”. In an era of intensive partisanship, races don’t necessarily shift that much between August and November, and there are fewer unexpected shifts. FLIPR now accounts for that, too.
Why forecasting midterms is hard
As we’ve said, FLIPR forecasts congressional and gubernatorial races based on a combination of polls, fundamentals and expert ratings. The model is fit primarily based on Senate races since 1990, and House and gubernatorial races since 1998.
Compared with our presidential forecast, which mostly defaults toward being “polls only” by election day, FLIPR is more of an even blend between polls and fundamentals (and in Deluxe, expert ratings, though these have less influence than the other major components). This is both because non-polling data is richer for House, Senate, and gubernatorial races than for the presidency — there are around 500 races combined in each midterm, rather than just one presidential race — and because polling data is sparser. Many important House races receive little to no nonpartisan polling, for example, so we need other methods to forecast these races.
I’ve emphasized the multi-layered nature of FLIPR: polls, fundamentals, expert ratings. But there’s also another dimension that cuts at a 90-degree angle to this. In FLIPR, there are both “micro” indicators that apply to one election (e.g., the polls in a particular Senate race, or how much money each candidate has raised in that state) and “macro” indicators like the generic ballot and polarization5 that affect all of them.
The middle zone between the micro and the macro is a rich and sometimes turbulent environment, however. You can’t just forecast each race individually and apply some sort of Monte Carlo simulation on top because race outcomes are somewhat correlated. There are years like 2016 when a party (in that case, Republicans) beats its polls across the board in a way that indicates systemic bias in the forecast rather than variance in individual races.
Still, the correlation among congressional races is less than that in presidential elections — the major difference being that you don’t have the same candidates on the ballot in every race. In fact, this is especially true in midterm years as opposed to presidential years, when the presidential race serves as more of an anchor on everything. The 2018 midterm was strong for Democrats, but also featured pockets of Republican strength, like in Florida or in Senate races in rural states that ousted Democratic incumbents. The 2022 map was patchy, too, with essentially a mini “red-wave” in New York and New Jersey that didn’t extend to most other parts of the country.
The rest of this document is more practical and less philosophical. I’ll explain how FLIPR works in seven high-level steps. Then I’ll give you a fairly exhaustive list of what’s new or different in the model in 2026.
FLIPR in 7 Steps
Step 1. Calculate polling averages for individual races and the generic ballot
The common thread between many of our politics models is our polling average algorithm. The code we use to calculate our Trump approval ratings and generic ballot averages is mostly the same6, for instance. And mostly the same process is also used in FLIPR:
Polls are weighted based on recency, sample size, and their Silver Bulletin pollster rating;
The polls are adjusted in various ways, such as for house effects;
We take care to ensure that the model responds to new polls well, neither being “too bouncy” or too impervious to new evidence; we’ve found the best way to do this is to calculate what are essentially various LOWESS regression curves, and then average more aggressive and more conservative versions.
The difference between FLIPR’s polling averages and others like our presidential approval ratings is that FLIPR makes a more comprehensive series of adjustments:
FLIPR’s house effects adjustment — house effects are persistent differences in polling firms tending to rate Republicans or Democrats more highly — is more robust because FLIPR has more data to rely upon: it can derive these both from the generic ballot and from polls of individual races. It’s also more punitive to expressly partisan or internal campaign polls when these are released to the public, essentially starting out with a prior that such polls are biased.7
FLIPR applies a likely voter adjustment whereas our regular polling averages do not. Basically, every poll of registered voters or all adults is adjusted toward what we think the poll would say if it were conducted among likely voters instead. This is calculated by comparing polls that publish both registered voter (or all adult) and likely voter versions of the same survey. In the large majority of such surveys this year (2026), for instance, Democrats gain ground on the LV versus the RV version, probably reflecting greater enthusiasm. Thus, RV/adult polls are shifted toward whatever party tends to win these comparisons — so far this cycle, Democrats — although the magnitude of the shift is shrunken toward a prior8 until we have enough such polls that FLIPR deems the prior can be entirely phased out.
Finally, FLIPR applies a timeline adjustment based on changes in the generic ballot. For example, if a race was last polled in March, and Republicans have gained 3 points in the generic ballot since then, the model will take a fraction of those 3 points9 and add it to the Republican candidate’s margin. Because our generic ballot average uses fairly conservative settings, this adjustment isn’t usually that pronounced. But it does mean that our polling averages in individual races can shift slightly even if there is no new polling in that contest.
So basically, FLIPR seeks to estimate what a poll of a given race would say today if it were conducted on a likely-voter basis. In this way, it hopes to avoid relying on “stale” polling.
FLIPR also applies the likely voter adjustment to its averages of the generic ballot and presidential approval, which our regular averages do not. This makes a material difference. As of early August, for example, Democrats are up by about 6.5 points in our default generic ballot average, but that advantage expands to around 8 points after likely voter adjustments, which is the version that FLIPR uses. (The likely-voter adjusted version of the generic ballot can be found on the FLIPR landing page.)
Step 2. Estimate the national environment and calculate CANTOR similarity scores to derive the “Lite” snapshot of each race
The “Lite” version of our forecast is supposed to be “polls only”. But that creates a problem for races that have little or no polling. Our solution is to impute polling-like estimates for these races based on FLIPR’s calculation of the national political environment, and polling in similar races using an internal system we call CANTOR (Congressional Algorithm using Neighboring Typologies to Optimize Regression).
FLIPR’s estimate of the national environment is based partly but not entirely on the generic ballot. It also calculates a “bottom-up” estimate of what polls of individual Senate and House races imply about the national picture.10 For example, say that the generic ballot has Democrats up by 8 points, but polling of key races for Congress is more consistent with something like a D +4 environment instead. FLIPR will basically take an average of these11 , pegging the environment at something like D +6 instead.
Implicit in the bottom-up calculation is what we call a “partisan lean score12” for each state and district, which is based on how it voted in the past two presidential elections relative to the country overall, as well as in recent state legislative elections. For congressional districts, estimates reflect new boundary lines after redistricting.
Texas has a partisan lean of around R +11, for example. So if, say, Democrat James Talarico leads Republican Ken Paxton in our Texas polling average by 2 points, that reflects a 13-point outperformance relative to the baseline. Individually, that’s a noisy estimate — and Paxton is arguably a poor candidate. But when you apply this technique to all polled Senate and House races, these differences tend to cancel out. (Though the implementation of this is more complicated in practice than in this simplified example.13) Newly this year, the bottom-up estimate also places more weight on states and districts that are similar to most others nationally, so a poll of a swing district in Pennsylvania will have more impact on the calculation than a race in Hawaii or rural Alabama.
While you could just calculate a reasonable proxy forecast based on the national environment + the district’s partisan lean score, CANTOR takes a couple of further steps. One is to adjust this prediction based on something we call elasticity: how sensitive a state or district is to changes in the national environment. Canonically, a state like New Hampshire is very “swingy” — it has a lot of cross-pressured, moderate voters — while a state like Georgia is more bifurcated between strongly Democratic (e.g., Black voters, young professionals) and strongly Republican groups (evangelical whites). This elasticity calculation, which is derived from voter-level microdata14, is employed at various points by FLIPR, although for the most part it has a relatively minor impact.
CANTOR itself is a more involved process. Basically, it calculates similarity scores between any given combination of states or districts, creating what’s essentially a set of 48515 x 484 pairwise comparisons. These similarity scores, in turn, are based on a variety of political (e.g., partisan lean score), demographic (racial composition, religious composition, college-educatedness, etc.), and geographic characteristics (e.g., region16, geographic distance, urbanization) and how similar states or districts are across the entire set of these factors.
CANTOR then evaluates whether polls of individual races tend to be stronger or weaker than the baseline in similar states and districts. For example, say that CANTOR is calculating its estimate for Iowa’s 2nd congressional district in the northeastern part of the state, one of the most competitive races this year. Democrats are mostly outperforming presidential-year partisan baselines in every race this year, but does that overperformance tend to be concentrated in states and districts that are similar to IA-217? If so, CANTOR will also project further outperformance there.
Step 3. Add fundamentals to calculate “Classic”
“Fundamentals” basically get at the exact opposite question from a “polls only” model. Suppose that there were a national moratorium on state and district polls.18 How would you predict individual races? Well, you could get pretty far by looking at things like district partisanship, incumbency, and fundraising, especially if you were also able to use national polling series like the generic ballot (even if you didn’t have any local polling data).
FLIPR actually calculates two different versions of the fundamentals, however. What we call “simple fundamentals” is just the generic ballot plus each state or district’s partisan lean score. “Fancy fundamentals” uses a whole host of factors based on a large-scale regression analysis of past races since 1990, conversely. Many of these factors fall under the category of “candidate quality,” such as experience in elected office and even the presence or absence of scandals.
Prior to 2022, the model only used “fancy fundamentals”. But the more that polarization increases — and it’s higher than ever now — the more that partisanship crowds out everything else. “Simple fundamentals” can increasingly be equally if not more predictive than “fancy fundamentals”.19
The precise mix of “simple fundamentals” and “fancy fundamentals” varies based on overall national partisanship, as well as the type of election. For Congressional races, “simple fundamentals” now get almost half the weight in many cases. Roughly speaking, this means that candidate quality factors, including incumbency, are only about half as important now as they were over the broader period since 1990. However, “simple fundamentals” tell you comparatively less in gubernatorial races, where there is still much more ticket-splitting; e.g., Republican Gov. Phil Scott will probably be re-elected again in Vermont.
New this year, FLIPR also looks at how much “simple fundamentals” and “fancy fundamentals” diverge. In Vermont, for instance, partisan lean and the national environment would imply that a Democrat would almost certainly be elected governor, while “fancy fundamentals” account for Scott’s strong popularity as judged by his large re-election margin. In cases where “simple fundamentals” and “fancy fundamentals” are not aligned, the model trusts fundamentals less in general. In these cases, it both puts more emphasis on polls and (in Deluxe) expert ratings, and increases the overall uncertainty associated with the race.
The specific factors used in the “fancy fundamentals” calculations are as follows. Some factors can only be used in races with incumbents20 or only in Congressional races (not governor), and these are designated accordingly:
Partisan lean scores for the state or district;
The national political environment, as derived in Step 221;
Incumbency; in fact, we technically fit separate regressions for incumbent and open-seat races. FLIPR’s handling of incumbency is slightly complicated:
It uses congressional approval polls to calculate whether the overall political environment is strong or weak for incumbents;
It assumes the incumbency advantage is larger in smaller, more idiosyncratic22 states and districts, where constituent services tend to be more important;
It uses the incumbent’s margin in her last competitive race, adjusting for the political environment at the time of the last election and changes in district boundaries;
It looks at whether the incumbent is in their first elected term, and whether they defeated an incumbent in their previous race;23
And for states that have redistricted, the incumbency advantage is less when there’s little overlap between a member’s old and new districts.
Fundraising, based on individual contributions only (so not party or PAC money) with a bonus for contributions from within a candidate’s home state;24
Elected experience on a 4-point scale from 0 (never elected to anything) to 3 (current or former elected senator or governor);
Party loyalty scores based on congressional voting records. Candidates who vote against their party more often actually outperform electorally, other factors held equal. This calculation is only used for congressional incumbents; there is no proxy for it for governors or for challengers (even if they’ve served in Congress before);
Scandals25, where new scandals are treated as being more important than “old” scandals when a candidate has already survived an election with the scandal having been known to the public.
Also, Classic does use a pinch of CANTOR’s proxy estimates for polling, though at a relatively modest weight. While CANTOR is a cool system, it’s generally less accurate than either fundamentals or polls directly from the state or district.
Step 4. Add expert ratings to get “Deluxe”
Every year, we have a debate about whether to retire the Deluxe version and stick to a model that uses purely objective inputs. Even though the expert ratings reflect other people’s judgments and not ours, there are some undesirable properties from including them. For instance, if experts are looking at Silver Bulletin forecasts to help calibrate their ratings, the entire process becomes somewhat recursive. Furthermore, in recent elections, the Deluxe version of our forecasts typically only “calls” one or two additional races correctly across the ~500 races we forecast every year compared to Classic.
But we’ve decided to continue publishing Deluxe and indeed treating it as the “default” forecast you see on the page — you can always click on the Classic or Lite versions if you prefer them.
That’s because, in the original design of the model, expert ratings were intended to serve a fairly narrow purpose: basically to account for idiosyncratic factors in each race that are not necessarily well-captured by the fundamentals, basically to shift around a little bit of vote margin from state to state without changing the macro picture much. Expert ratings are not a good way to estimate the national environment. Although the expert forecasts are good at discriminating between relatively stronger and weaker races for each party, they have a weird and persistent bias. Namely, expert raters tend to implicitly assume a neutral political environment (i.e., a generic ballot of about zero) until quite late in the cycle. As of early August, for instance, the unadjusted expert ratings imply a political environment of around D +2 versus the roughly D +7 environment calculated by FLIPR Classic. In a D +2 environment, a candidate like Talarico probably wouldn’t win despite Paxton’s weaknesses, but a D +7 would require much less overperformance.
FLIPR’s predecessors have always been aware of this tendency and “de-biased” the expert ratings before applying them. But in auditing the code this year, we discovered some shortcomings in the way this de-biasing process had been applied by the model. Basically, because of some artifacts in how races are rated, there was a bit of leakage from the experts’ implicit assumptions about the national environment (i.e., wrongly assuming the generic ballot = 0) into the Deluxe forecast. And we specifically do not think the experts’ views of the national environment provide much utility at all; their value is solely in the rank-ordering of different races relative to the national environment. In working on the model this year, we developed better techniques for removing this bias completely.
So the steps in going from Classic to Deluxe are as follows:
Translate the expert forecaster categories (e.g., “Lean Democratic”) into an implicit forecast of the margin in the race (e.g., D +6).26 The expert raters used in this calculation are Cook Political Report, Inside Elections and Sabato’s Crystal Ball.
Estimate the national environment implied by the translated expert ratings.27
De-bias the expert ratings by shifting them to FLIPR’s estimate of the national environment. For instance, if FLIPR estimates a D +7 national environment and the experts’ implicit national environment is D +2, we’d shift the expert ratings by the equivalent of 5 points toward Democrats.
Blend the de-biased expert margins with Classic. The weight assigned to expert ratings is generally smaller now than in past years. It’s typically about one-sixth of the forecast for Congressional races with polling and one-quarter of the forecast for gubernatorial races with polling28, though higher for races without polling. Before, expert ratings got closer to one-third of the weight in Deluxe across the board.
We believe this better serves the model’s original intent of using expert ratings to gently nudge the forecast in individual races — giving them more say in races where other indicators are less reliable, and they have a particularly strong opinion — without letting the experts change the overall national picture.
Step 5. Translate the current snapshots into forecasts, accounting for projected changes in the national environment
This is one of the simpler steps. FLIPR calculates an estimate of the House popular vote based on macro-level factors in elections since 1946: presidential approval ratings, the House popular vote in the previous election29, polarization, and whether the election is a midterm. As of early August, for example, this calculation forecasts Democrats to win the popular vote by around 8.5 points (mostly because of Trump’s very poor approval rating).
Meanwhile, Classic and Deluxe have their own estimates of the current national environment, which is currently closer to D +7: we call this the snapshot. The snapshot is then nudged toward the prior — so this year, for example, Democrats are expected to gain slightly more ground between August and Election Day. We don’t apply this adjustment in Lite, since Lite is supposed to be a “polls only” model. Instead, in Lite, the forecast is basically the snapshot, but with greater uncertainty.
I say “nudged,” however, because the effect of this adjustment is relatively gentle, and essentially fades out by Election Day. Indeed, as the generic ballot has become more stable, the prior has somewhat less utility than it used to in forecasting elections even months in advance.
Step 6. Account for uncertainty and calculate correlations between races
A candidate with a 3-point lead in a race where the margin of error is 5 points will win her race 88 percent of the time. This might be a good estimate in a well-polled Senate or gubernatorial race under stable conditions on the eve of the election. If the margin of error doubles to 10 points, however, because the election is a long way off, polling is sparse, or the data is contradictory, she’ll only win 72 percent of the time instead. That’s much better than a toss-up, but still a considerably less confident forecast.
Because all of our models are probabilistic, we tend to care a lot about this kind of stuff. It’s not just that elections are uncertain: the degree of uncertainty is predictable, at least to some degree. Abdul El-Sayed’s near loss in the Michigan Democratic primary despite a large polling lead wasn’t that surprising given that it’s not uncommon for primary polls to miss the mark by double digits, for example.
So in FLIPR, uncertainty is higher when:
There’s less polling in the race;
There are a lot of undecided voters in the polls
There’s more time to go until the election;
The various subcomponents of the model disagree with one another, e.g., there’s a big divergence between fundamentals and polls.
However, this is only half the battle — and it’s the easier half. As I’ve said, the errors in forecasts of congressional and gubernatorial races are somewhat correlated with one another, although less so than in presidential elections. So FLIPR not only has to estimate the uncertainty associated with each of the ~500 races that it forecasts, but also how the uncertainty in one race relates to another one. If Democrats beat their forecast in the Iowa Senate race, for example, they will probably also do so in the Iowa gubernatorial race and in Iowa’s 2nd congressional district — but we wouldn’t expect a race in Hawaii to be particularly correlated with Iowa.
FLIPR approaches this by breaking uncertainty down into 3-5 interlocking subcomponents, depending on how you count them:
National error affects all races in the forecast roughly equally. If, in one simulation, the overall map is 3 points more Democratic or more Republican than in our forecast, we basically30 just shift every race by 3 points in that direction. This, in turn, can be broken down into two further subcomponents:
“Drift”, our term for the change in the national environment between the current forecast and Election Day. There tends to be less drift — fewer macro-level shifts in the polls — in political climates with high polarization like the one we’re in now. And drift zeroes out as Election Day approaches.
Election Day error: basically the possibility of a systematic polling/forecasting error on Election Day itself.
Local error, conversely, applies only to one race or one state. In fact, we now account for both of these conditions:
In each simulation, we take one random draw that applies to all races (House, Senate, governor) in a particular state. (This is a new wrinkle for 2026.)
And then we also draw a random error that applies only to exactly one race.
Finally, correlated demographic error is an intermediate layer that sits between these and is the most complex layer. Basically, we use all the variables used in CANTOR to simulate the possibility of underperformance or overperformance among particular subgroups. In one simulation, for example, Democrats might do better than expected among Hispanic voters, but worse than expected in the Northeast, and worse than expected among Catholics. This can produce the sort of geographic or thematic clustering of errors that we’ve seen in recent elections like 2018 and 2022 and FLIPR tries to create some maps like these. The intuition is simply that errors are more correlated across similar states and districts.
Step 7. Simulate the election 40,000 times, including projecting turnout
Since almost all of the parameters that FLIPR needs have been generated by this point, the simulations themselves aren’t particularly complicated from an algorithmic standpoint.31 There are a few small wrinkles, however:
FLIPR uses slightly fat-tailed distributions to account for the possibility of rare, truly-out-of-sample events, although this effect has been dialed down from previous years;
We use variance reduction techniques that make the simulations more stable, so the effective sample size is larger than 40,000 simulations;
And the simulations do get more complicated in the case of multi-way races:
In races with prominent independent candidates, FLIPR simulates the independent’s vote share using a right-tailed statistical distribution in which the independent usually doesn’t come close but has some outside chance of gaining momentum and being a real factor in the race;32
In Maine and Alaska, FLIPR now carries out a fairly faithful simulation of ranked-choice voting;
In states like Georgia where elections go to runoffs if no candidate gets 50 percent of the vote, FLIPR simulates this too;
And FLIPR simulates the complex “jungle primary” process in Louisiana’s House races, which are using it for the last time this year.
FLIPR also probabilistically simulates election turnout and calculates an estimate of the national House popular vote. The turnout simulations are essentially a post-processing step; they don’t actually affect our calculations of vote shares, or the probability of a given party winning a particular race or the House or the Senate. Nevertheless, the design is fairly elaborate:
Turnout in individual states and districts is benchmarked based on the competitiveness of the race, whether there’s a Senate or governor race in the district, past turnout in the state or district, and FLIPR’s estimate of the voting eligible population (VEP), adjusted for population growth.
Overall national turnout is forecasted based on polls that ask voters whether they’re more excited or less excited than usual to vote in the upcoming election.
The local and national estimates are then reconciled in various ways. FLIPR requires turnout in House, Senate and gubernatorial elections in the same state to remain within designated bounds. It applies a modest correction if it infers from its estimate of the national political environment that one party’s voters are likely to turn out in greater numbers as compared to past elections. And it treats its national turnout estimate as authoritative; if, say, summing up its projections in individual districts produces an estimate that overshoots the national projection by 2 percent, it will just give every district a 2 percent haircut.
Like basically everything else in FLIPR, these turnout estimates are treated as probabilistic rather than deterministic; we estimate both the mean and the variance.
Changes since 2022
Since the rest of this document is quite long, I’ll be brief here. But you’ll basically see a few major themes.
Thinking more carefully about how increasing polarization affects all of the parameters in our model. We changed some of this in 2022 but left others alone.
Thinking more carefully about the relationships between House, Senate and governor races within particular states.
Accounting for some predictable differences in behavior between midterms and HSG behavior in presidential years. For instance, races are more correlated in presidential years.
Here is a fairly exhaustive list of changes:
The national environment and forecast uncertainty
We’ve now more carefully calibrated the amount of “drift” in the forecast from the current day to Election Day. Previously, the model blended drift and Election Day uncertainty together. The new specification is more consistent with our presidential forecast. As part of this process, we discovered that the national picture now “locks in” earlier than it used to.
FLIPR’s estimate of the national political environment now blends the generic ballot with a “bottom-up” estimate derived from state and district polling, rather than relying on the generic ballot alone. This was already being done for CANTOR but had not been applied to the rest of the model.
In calculating its prior for the national popular vote, the model now makes a bigger distinction between presidential years and midterms. Midterms tend to produce more lopsided outcomes than are typically seen in congressional elections in presidential years.
The model now assumes that races are slightly more correlated with one another in presidential years than they are in midterms.
We’ve now introduced a statewide variance layer in addition to local and national variance.
Fat tails are less severe than in past years, as we now have more cycles without huge outlier outcomes.
We’ve cleaned up some inconsistencies in the projection of the national political environment between Lite, Classic and Deluxe.
Variance-reduction techniques are employed to keep the simulations more consistent.
Fundamentals, CANTOR and expert ratings
The overall weight given to fundamentals is now differentiated between congressional and gubernatorial races. Fundamentals tend to be more predictive in the former than the latter because congressional races more closely follow the partisan lean of a state or district and the data is generally higher-quality.33 As compared to the 2022 build of the model, fundamentals are considerably upweighted in our congressional forecasts but slightly downweighted for governors.
The ratio of “simple fundamentals” to “fancy fundamentals” in Classic is now tied to our polarization coefficient rather than using a fixed ratio.
“Fancy fundamentals” now use FLIPR’s overall estimate of the national political environment, including imputations from state and district polls, rather than just the generic ballot. “Simple fundamentals” are still based on the generic ballot alone.
When “fancy fundamentals” and “simple fundamentals” disagree, the model now puts less weight on fundamentals and tends to trust the polls more. When they agree, it increases the weight on fundamentals.
CANTOR and the “bottom-up” estimate of the national political environment have been improved in various ways, including by weighting races that are more representative of the national political environment more heavily and more explicitly accounting for selection bias in which races get polled.
A wider array of demographics are now accounted for by CANTOR, particularly including religious groups (e.g. Jewish voters and Muslim voters) that were lumped in with other groups before.
The weight given to expert forecasts versus the other components is generally now lower than it previously was in Deluxe, and now varies inversely with how reliable FLIPR thinks the fundamentals are in each race. Essentially, the expert ratings are used as more of a fallback in cases where other indicators may be unreliable.
In Deluxe, the model is now much truer to its original intent of not allowing expert forecasts to affect the national environment — only the forecasts of individual races.
Gubernatorial races have always been treated as being less partisan than congressional races by the model, but this distinction is now more complete. For example, gubernatorial forecasts rely less on the “fundamentals” because partisan lean is a less reliable signal for them.
Polling and turnout
As turnout patterns change, the model has discarded its prior that the shift from registered voters to likely voters typically favors Republicans. It now uses a more neutral prior, but more quickly overrides it as more polls that directly compare likely-voter and registered-voter results become available.
The model now calculates its own native version of presidential approval ratings that makes a likely-voter adjustment.
Turnout estimates within a state are now reconciled between House, Senate and gubernatorial races to ensure they match one another within certain bounds.
Turnout estimates are also now partly adjusted to better match FLIPR’s estimate of the national political environment. This adjustment is capped so that it cannot overwhelm other information in the turnout model, and the resulting estimates are still allowed to diverge because of Senate and gubernatorial races, ballot structure and other local factors.
Miscellaneous
In 2022, we introduced a bonus in our fundraising data to account for contributions from within the candidate’s state, which tend to be more predictive of the eventual outcome. But this was implemented clumsily and excluded small-donor contributions where the donor’s state isn’t aggregated by the FEC. We now look at the ratio of in-state to out-of-state contributions for donors where this information is available, and then extrapolate it to all types of individual contributions, including small-donor contributions.
In certain rare circumstances, we now allow the fundraising estimates to be overridden and calculated by proxy instead. Namely, these circumstances are when a candidate (say, Troy Jackson in Maine or Darline Graham in South Carolina) is a late replacement and hasn’t yet had time to file a fundraising report, or when a candidate (say, JB Pritzker in Illinois) eschews or discourages individual contributions to their campaign. Because FLIPR does not account for self-funding, these candidates would previously have been treated as raising zero money, but treating this as missing data is more appropriate.
The model now explicitly simulates ranked-choice voting.
It also now explicitly accounts for Alaska’s top-four primary system, in which more than one candidate from the same party can appear on the general-election ballot.
Dolphin emoji optional.
We didn’t run a Congressional forecast in 2024.
Though we’d dispute the extent of this, and it’s been far more of an issue in presidential years than in midterms.
Gubernatorial elections remain considerably less correlated with partisanship as determined by presidential voting than congressional elections, and FLIPR now accounts for this. It also now gives more weight to the fundamentals in races where the fundamentals are more self-consistent and less when they’re more contradictory.
Our indicator for polarization is the difference in party means from Voteview. This has perpetually been growing, meaning that the parties now usually vote as unified blocks with almost no crossover.
The main differences pertain to how aggressive the respective averages are. Presidential approval ratings tend to tolerate more aggressive settings, while the generic ballot average needs to be fairly conservative or will bounce around too much — just trust me on this one since I’ve tried it the other way.
On average in the historical data, this bias is about 4 points toward their preferred candidate.
In previous years, the prior tended to assume that Republicans would benefit from the shift to likely voters, but this is much less clear now as Democrats’ college-educated coalition tends to turn out at high rates.
We’ve found empirically that a partial correction works better than a full correction.
But FLIPR does not use gubernatorial polls in this national environment calculation, as they’re considerably less correlated with overall state partisanship than polls of congressional races.
Though the average is weighted based on the relative amount of generic ballot vs. individual race polling.
Partisan lean scores are analogous to Cook Political’s PVI, though denominated differently. If we say a district is R +10, that means the vote margin in that district is 10 points more Republican than in the country overall. Cook PVI would call the same district R +5 because it’s based on the vote share for one party, rather than the margin which is usually about twice as large.
That’s because there is significant selection bias in which races get polled. A R +12 state or district that is expected to be competitive might get a lot of polling, while one where the Democratic challenger is weak might not. CANTOR has methods for correcting for this selection bias, which we reviewed and slightly revised this year.
Specifically, from the Cooperative Election Study waves from 2016 through 2025.
i.e., 435 districts + 50 states
We define the four major regions of the country as Northeast, South, Midwest, West, though we’ve had our own longstanding internal definitions of these rather than using the Census Bureau’s, and some states are split between regions.
i.e., predominantly white, Midwestern, suburban/small town, etc. districts.
This is phrased carefully because fundamentals do use generic-ballot polls.
Even though “fancy fundamentals” also include the national environment and partisan lean scores.
Races with appointed incumbents as opposed to elected incumbents are treated as open-seat races.
In line with their names, “fancy fundamentals” uses our fancy estimate of the national environment derived from both the generic ballot and imputations from local race polls, while “simple fundamentals” uses the generic ballot alone. This is another change for 2026; in previous years, both versions used the generic ballot only.
That is, states and districts with low pairwise CANTOR scores, meaning they don’t resemble most other districts.
Being in your first elected term is actually an advantage in the model, because it implies that you didn’t have the incumbency advantage in your last race but now you do. Likewise, having defeated an incumbent in your previous election implies greater electoral strength.
Home-state contributions get a 5x weight.
The list of scandals was originally compiled by my former FiveThirtyEight colleague Nathaniel Rakich. Eli and I have done our best to keep it updated. This is one of the more subjective parts of the model, though.
This requires more caution for races rated as “Solid/Safe R” or “Solid/Safe D” because these span a very wide range of outcomes. A race where the Republican is expected to win by 15 points might be classified as “Solid R”, but so would one where she’s projected to win by 50. Basically, the model now uses a Bayesian process for handling these cases. A “Solid R” rating is not informative unless, based on Classic’s forecast, the race was plausibly near a boundary where it might have been rated as something other than Solid R.
This is calculated separately for each expert rating outlet.
The weights now depend partly on how reliable the model thinks the fundamentals and the polls are. Expert ratings get higher weights in races where FLIPR thinks these other methods are less reliable. That’s why they get more weight in gubernatorial races, for instance; the fundamentals are generally less reliable in governor races than in races for Congress.
For recent years, this is adjusted for “blank districts” in which a party did not field a candidate.
Although the shift is multiplied by the state or district’s elasticity rating.
The bigger challenge is really in the engineering/programming to make the simulations run quickly.
But note that an increasing number of independent candidates are really Republicans or Democrats in everything but name. In several races such as Nebraska this year, for example, Democratic candidates have actually abandoned their races to make way for left-leaning independents. In cases like these — or cases like Bernie Sanders in Vermont — the model treats these candidates as de facto Democrats.
Particularly when it comes to fundraising data.
Ken Paxton is "Arguably a poor candidate"? Is ANYONE arguing that he's a good candidate?
Raise your hand if you'll be calling your next fantasy sports team the Fancy Fundamentals