After eight months of work, the first-ever gubernatorial and state legislative forecasts at State Navigate are now out. With the Virginia and New Jersey elections just 98 days away, we’re excited to use our wide breadth of experience from previously working at CNalysis in being the first outlet to publish a quantitative forecast of this year’s elections in Virginia.

Meet the Team

Executive Director Chaz Nuttycombe and Development Director Jack Kersting, the two lead forecasters at State Navigate, created the most accurate forecast of the 2023 Virginia House of Delegates and State Senate elections, accurately predicting the winner of every district in the Virginia legislature that year. Nuttycombe has been forecasting Virginia elections since 2017, and the Virginia House of Delegates elections that year was the first chamber he ever created a forecast for. Compared to Kersting and Nuttycombe’s work at CNalysis, the forecasting team has grown slightly larger, therefore augmenting our forecasting capabilities and bringing even more expertise onboard.

Elections Coordinator Michael Foley has provided election data down to the precinct level across the Commonwealth, and has made State Navigate the only outlet that has published accurate results of last year’s elections in Virginia by using a formula to allocate same-day registration votes to districts and precincts, of which there are many in highly competitive college-town districts such as House District 41 in Blacksburg.

Mary Radcliffe, State Navigate’s Data Manager, has created a custom polling database for State Navigate to help keep the forecast up to date with the latest information. Political scientist Steve Rogers has calculated the measured impact of incumbency and campaign finance spending in the House of Delegates, making this the first-ever state legislative forecast in history that quantitatively applies campaign finance data to district outcomes.

Together, this team has constructed a forecast that is largely quantitative but also can have qualitative analysis applied when required. Our main focus, of course, are state legislatures. As such, our attention is mostly on the Virginia House of Delegates. We hope to create a forecast that accurately predicts the winner of at least 97/100 seats in the House of Delegates this year. Both Kersting and Nuttycombe’s 2021 House of Delegates forecast and Nuttycombe’s 2017 forecast accurately predicted the winner of 96 seats in the chamber.

Our Three Lessons from 2024

Across all state legislative seats that weren’t left uncontested, the CNalysis forecasts created by Nuttycombe and Kersting successfully predicted 94% of seats nationwide. When excluding multi-member districts, that number rises to 95%.

Our state legislative forecast success rate for Virginia has always run ahead of the percentage of seats that CNalysis would correctly predict nationwide in even-numbered years due to two reasons: Familiarity and connections in the Commonwealth, and the sheer amount of previous election data that we can easily get in Virginia.

Still, we always focus on our losses instead of our wins to see where we went wrong and how we can continue to improve our forecasts. After analyzing 2024, we came away with three lessons in state legislative forecasting.

1. To this day, State Navigate is the only organization in the country that forecasts the outcome of every state legislative seat in the country, and it’s a bear to do. In 2022, there were few states where we calculated what the state legislative vote was in the 2020 elections. In 2024, that remained true, and we largely relied on the state legislative vote data from 2022. In comparison, the 2023 Virginia forecasts calculated every single bit and byte of election data under the new redistricting lines going back to the 2017 elections. This difference made the 2024 forecasts more reliant on statewide and presidential data instead of state legislative data, while the 2023 Virginia forecasts were reliant on the actual state legislative vote itself. 
2. Chaz has used formula-based modeling to create qualitative ratings since the 2022 forecasting season. This has made us much more accurate compared to our forecasts previously– a ‘base rating’ is created using hard election data, with qualitative analysis applied thereafter. This includes campaign finance, which is now quantitatively applied, but also what the qualitative state-level ‘conventional wisdom’ is and any stories (i.e. scandals) that may affect the outcome. While Chaz’s qualitative analysis proved to be useful in the forecasts last year, we missed nearly fifty districts by following ‘conventional wisdom.’ A large chunk of districts were given a more friendly rating to the eventual losing party due to both Republicans and Democrats misreading the environment. As such, going forward the only effect political insiders may have on the forecasts are by providing hard data, such as polling, that they’re willing to share with State Navigate’s forecasting team.
3. Special elections (or contests such as the Washington Primary) are no longer useful tools for predicting elections in presidential years, but until proven otherwise we assume they are still a useful tool in measuring the outcome of non-presidential year elections. The propensity of the Republican and Democratic voting bases has been upended, even since the 2017 gubernatorial election in Virginia– Democrats have gained with wealthier and more educated voters that are more likely to turn out in special elections, which created a ‘blue mirage’ of a Democratic-friendly environment in 2024. We’re appropriately applying a rolling average of special elections compared to their presidential margin to the forecast.

Okay, so what’s different about this year’s forecasts compared to 2023 and 2024?

To recap the aforementioned adjustments in Chaz’s work in the ratings creation process via formula: We’re applying campaign finance data quantitatively rather than qualitatively, political insiders can only influence ratings via providing ‘hard’ internal data rather than ‘soft’ conventional wisdom. We also aren’t publishing ‘ratings’ on our forecast pages and instead are only providing the odds and projected margins.

Digging into what’s new on Jack’s end: In 2024, the CNalysis state legislative forecasts had two models: The ‘classic’ model, which only uses handmade ratings made by Chaz that Jack runs through thousands of simulations to calculate overall chamber odds, and the ‘expanded’ model, which simply put took recent election results and adjusted them to the 2024 environment, which will be detailed more below. This time, we’re only publishing one model, which is an updated version of 2024’s expanded model.

The core of this year’s model are recent elections in the district. For the forecast, we use a handful of elections we feel best represent the data and weigh them according to the most recent and comparable elections.

For the HoD in 2025, the biggest weight is given to 2023 HoD results, followed by Gov. 2021, President 2024, and a pinch of President 2020. This helps encapsulate incumbency, 2023 HoD, statewide race in the year following a presidential election, Governor 2021, and the most recent presidential election. This makes up roughly 85% of the forecast, the other 15% are from Chaz’s ratings. 

Aside from the weightings, the model has to make adjustments to the margins of the elections to match the current environment. Each election is adjusted in accordance with if the state was tied, to zero out the margin.

For example, in 2024 Trump won the 40th district by 9.2%, but lost the state by around 5.8%. This gives the 13th a zeroed margin of 15%, with rounding. On top of the normalizing of margins, some other adjustments for beating incumbents, retirements, and previously uncontested districts are made. The adjustments either adjust the zeroed margin, or remove that election from the calculation in that seat.

The zeroed margins are then added to the projected statewide margin, the governor forecast, plus regressed Republican overperformance in recent VA elections. Once adjusted, they are finally combined with Chaz’s ratings, and weightings stated previously to get a projected margin for each district.

Introducing: State Legislative Electoral, Demographic Similarity Index (SLEDS)

Once each district has a projected margin, the next step is to simulate the election. The simulation is a combination of a random walk, statewide margin, and a multivariate distribution, based on district similarity. State Legislative Electoral, Demographic Similarity Index, or SLEDS, is the similarity index based on the electoral history and demographics of the district.

This index is used in the multivariate distribution to determine correlations between districts. The uncertainty for each district was lowered after 2024, as the model was overly uncertain last year. For each simulation, the state moves x points in either direction, derived from time to election day and possible shifts in statewide fundamentals and polling.

This is combined with the multivariate distribution, that is derived from the uncertainty of the model on election day, and the similarity between districts. The simulation then counts the total number of seats for each party and then puts it into one of the majority categories. This is repeated 25,000 times to get a range of outcomes, giving you the forecast you see on the site.

The Statewide Forecasts

For the statewide elections, it is a bit simpler. All of the forecasts are derived from the gubernatorial forecasts this year. The governor forecast is made up of 2 factors, polling and the ‘Fundamentals’ of the race. The fundamentals use two indices, Trump approval change from the start of the term, and an economic index.

The two indices are then regressed against previous shifts from Virginia gubernatorial elections and the preceding presidential election. The polling is just a simple polling average, with adjustments for sample size, recency, population and partisanship of the poll. These factors combine to give a projected margin for the race.

The uncertainty of the model is based on possible movements of the factors and polls between now and election day, as well as the underlying error of the fundamentals and polls. For the races for Lieutenant Governor and Attorney General, the polling is the same, but with LG and AG polls, the fundamentals & variance are slightly different. Chaz has calculated an expected overperformance for each of the state row officer races vs. the Gubernatorial race based on incumbency and previous ticket-split rates. The model is also slightly less certain of the projections for these races as well.

As for the locality projections, it is identical to how the districts are zeroed out and then adjusted back based on the projected margin. 

This is a quick run of models and the data that goes into them, and how they work. The model is very similar to the previous forecasts we have created, but with minor refinements to them after a full cycle of them being implemented.

We hope that you enjoyed the rundown of our process as we head into election season here in the Commonwealth. It’s taken a lot of time to set things up just the way we like them, and we’re always collaborating and looking for ways to improve our forecasting building and methodology to make State Navigate as accurate as possible. While we are confident that our process is truly top of the line, we won’t stop until our powers of prediction are something close to supernatural. We hope you’ll stay with us as we continue to work our magic.