Methodology

Overview

This dashboard visualizes Mac Tan's forecast for the 2026 United States congressional elections. Unlike simple poll averages or pundit ratings, the model produces a full joint probability distribution across all House and Senate seats — not just a prediction of the outcome in each individual race.

In short, it combines what we know about a district from recent history with what polling is telling us right now about the national and local political mood, weights each source according how certain its prediction is, and then runs thousands of simulated elections to generate win probabilities and credible intervals for the outcome in each race.

Every number on this site is the direct output of that simulation. When you see "R Win: 63.4%", that means Republicans won the seat in about 6,340 out of 10,000 simulations.

The 2PV Metric

All forecasts are expressed as two-party vote share (2PV). Votes for third-party and independent candidates are excluded, collapsing every race into a zero-sum probability space on the interval [0.0, 1.0].

R2PV = R_votes / (R_votes + D_votes)

Win condition: R2PV > 0.5  →  Republican wins the seat
Win condition: R2PV < 0.5  →  Democrat wins the seat

Win probability in the final output is calculated as the fraction of posterior draws where a party's 2PV exceeded 0.50. This is a direct integration over the posterior distribution using Hamiltonian Monte Carlo.

P(R wins) = mean(R2PV > 0.5)  for sim_id in 1...N

Where N = 10,000 posterior draws per seat

National Swing Model

The national swing model produces a posterior distribution for each district, based on an estimate of its baseline partisanship and the national political environment. Several factors go into this model:

Partisan Baseline

Each district and state starts from a baseline 2PV result, which is the 2PV result in the previous regularly scheduled House or Senate election (excluding special elections). However, often the prior election result doesn't serve as a reasonable baseline: common reasons for this would be that the incumbent is retiring, the previous election wasn't contested, or (in the case of House elections) the district's boundaries have been redrawn since the last election. In these cases the model will rely much more heavily on the most recent presidential election result in the district or state to form a partisan baseline.

National Environment

The generic congressional ballot (the gap between voters who prefer a generic Democrat vs. a generic Republican) average provides an estimate of the national political environment and the extent to which it favors Democrats or Republicans. This is estimated from available polling at any given forecast date. When the generic congressional ballot swings toward one party relative to the previous election, that party will generally perform better in House and Senate races across the board, although polling and other district- and state-specific factors can strengthen, weaken, or overpower this effect.

Incumbency

Incumbents generally do better than generic candidates from their party, even if their advantages are weaker than ever. When there is no incumbent running for re-election, states and districts tend to revert to their presidential voting habits. A race is assumed to have an incumbent running for election until the incumbent has either suspended their campaign or lost renomination.

Midterm

District- and state-level election results tend to be more sensitive to swings in the national political environment during midterms than in presidential elections.

Redistricting

When a House district's boundaries are redrawn (as is happening in a lot of places these days), the previous election results are less useful as a partisan baseline. In districts that have been redrawn since the previous election, the model relies much more heavily on the most recent presidential election result instead of the previous House or Senate election result. The model will only consider new district boundaries once they have survived any legal challenges to them (so as of May 7, 2026 for example, House forecasts for California and Texas use the new district boundaries but House forecasts for Virginia and Florida use the old ones).

Polling Model

Polls go through adjustment and weighting before being averaged.

Partisan Adjustment

The poll partisanship adjustment is very basic: it adjusts polling results only if the New York Times poll tracker identifies the poll as a partisan poll. This is a hard criterion: it isn't enough for the pollster to generally produce results that lean toward one party; a poll is considered partisan only if it is conducted for a political party, explicitly partisan organization, or a candidate. Polls which are sponsored by one party have their results adjusted slightly away from that party and they are given much less weight in the average. For generic ballot polls, this is the only adjustment performed.

National Environment Adjustment

Polls are snapshots of the political mood at a point in time. A poll conducted six months ago might produce very different results if it were conducted today, even if it was conducted according to the same methodology on the same population using the same questions. To keep Senate and House district-level polling relevant, each state- and district-level poll's result is adjusted by the amount by which the generic ballot average has shifted since the poll was conducted. For example, if a Senate poll conducted six months ago showed a tie, but the generic ballot average has shifted six points more Democratic since then, the Senate poll would be adjusted to be six points more Democratic as well.

Poll Weighting Factors

MCMC Simulation

The final prediction is not a point estimate but a full probability distribution. Stan samples thousands of values from the posterior predictive distribution for each seat, capturing correlations across seats (e.g., a strong Democratic wave affects all seats simultaneously).

From these draws, this dashboard calculates:

r_prob     = mean(draw > 0.5)       // Win probability
r2p_avg    = mean(draw)             // Expected r2p
r2p_p05    = quantile(draw, 0.05)   // 5th percentile
r2p_p95    = quantile(draw, 0.95)   // 95th percentile

The massive raw posterior file (house_district_posterior.csv, ~124MB) is processed daily by the update_data.py worker script, which compresses thousands of rows per district into these five summary statistics and writes a lightweight JSON file for the dashboard to serve.

Aggregation

For each House and Senate race, the model produces at least a posterior distribution based on national swing as well as (potentially) a posterior distribution based on race-specific polling. These are averaged using inverse-variance weights, which is to say:

pred = (national_swing_pred/national_swing_var + poll_pred/poll_var) / (1/national_swing_var + 1/poll_var)

If var_poll → ∞ (no polls exist): pred → fund_pred
If var_poll → 0 (many perfect polls): pred → poll_pred

These weights are known to minimize the variance of the final prediction and are also consistent with Bayesian updating of a normal prior with normally distributed data.

Data Sources

Limitations & Caveats

All forecasts are probabilistic estimates, not predictions. A race with a 90% probability of one candidate winning is not a certainty — it means the model would expect the favored candidate to win 9 out of 10 times under similar conditions. Out of every 10 races where the leading candidate has a 90% probability of winning, it should not be a surprise when the underdog wins one.

Key limitations include the following:
- The model does not direct take into account things that happen in the news, like scandal or economic shocks. The forecast will reflect these things only to the extent that those news events are reflected in the polls.
- Polling in low-salience House races is sparse and often of poor quality: typically the only polls of a House race will be those conducted for one of the campaigns which the campaign decides to release. As you might imagine, this can lead to significant selection bias in what House polls we see.
- The model does not account for third-party and independent candidates who could tip results in close races. For the most part the impact of these candidates is limited, but in some races they may have a large impact or may even be running as the de facto Democratic nominee (see Nebraska's Senate election).

This dashboard is a visualization layer on top of publicly available research. The underlying model code is open-source and available at github.com/thisismactan/US-2026.