How WetherVane Works

Most forecasting models treat counties as independent: average the polls, adjust for house effects, output a number. WetherVane starts from a different question: what is the underlying structure that makes places move together?

Thousands of places share hidden behavioral patterns. A rural evangelical county in Georgia moves with rural evangelical counties in Iowa — not because anyone coordinates, but because the same forces act on similar communities. Discovering that structure — not just reading the surface results — is what makes prediction defensible.

The model is built for readers who want to understand electoral dynamics, not just consume a top-line forecast. If you read FiveThirtyEight for the methodology write-ups, this is for you.

All metrics are on the 2024 presidential election. LOO (leave-one-out) cross-validation excludes each county from its own type mean before predicting it — this is the honest generalization metric that cannot be inflated by self-prediction.

The standard holdout r (0.698) is inflated by ~0.22 because counties help predict their own type means. LOO r (0.731) is the correct metric for evaluating generalization. Both are reported for transparency.

Cross-election validation tests whether the type structure discovered from historical shifts generalizes to new cycles. Across four presidential election pairs, mean LOO r = 0.476 ± 0.10.

Not all cycles are equally predictable. 2020→2024 (r=0.38) was the hardest — the Harris-Trump dynamic produced unusual cross-type movement, particularly among Hispanic communities. The 2012→2016 transition (r=0.52) was most predictable: Trump's initial surge followed existing type fault lines closely.

The cross-election LOO metrics above measure whether the type structure generalizes to new presidential cycles. But the harder test is this: does the model generalize to a completely unseen election type — a midterm it has never trained on?

To test this, we retrained the model from scratch after removing all 2022 data — no 2022 governor shifts, no 2022 Senate shifts, none of it. We then used the 2020 presidential Dem share as a county-level prior, and predicted 2022 outcomes purely from the structure the model had learned through 2020.

The Senate result — county r = 0.97 across 1,880 counties — is the strongest evidence that the partisan geography structure the model learns is real and durable. Correlations this high mean the model gets the relative ordering of counties nearly perfect: which ones lean Democratic, which lean Republican, and by roughly how much. It does not mean point estimates are perfect — the RMSE of ±4pp on Senate races reflects real uncertainty about national environment and candidate effects.

Governor races are harder (r = 0.89) because governor outcomes depend more heavily on candidate-specific factors and state-level dynamics that a national type model cannot fully capture. The type structure still explains 79% of the variance in county-level governor outcomes — but the remaining 21% is genuine candidate effect.

Seven competitive Senate races, never seen during training

The model predicted these outcomes using only the type structure learned from 2008–2020. Errors are predicted minus actual Democratic share (state-level, vote-weighted average across counties).

Five of seven races land within ±2.5pp. The type-mean baseline (r = 0.68) predicts these same races with errors up to ±17pp — showing that the presidential-prior enrichment is doing substantial work beyond just knowing which communities lean which way.

Extended multi-election backtest

Beyond the 2022 holdout, the model has been validated across 11 elections spanning presidential (2008–2020), Senate (2014–2022), and governor (2018, 2022) cycles using year-adaptive Ridge priors. The combined backtest achieves r = 0.939 with direction accuracy of 88–100% across all elections. The model shows an expected temporal gradient — stronger on recent elections where the political geography more closely matches the training era — but maintains predictive power even for elections 16 years in the past.

• Structure from behavior, not demographics. Types are discovered from how places shift electorally. Demographics describe the types after discovery — they do not define them. This avoids baking in assumptions about which demographic groups drive politics.
• Cross-state information sharing. Because types cross state lines, a poll in one state informs predictions in another. Most models treat states as independent. WetherVane treats the country as one connected landscape.
• Full uncertainty quantification. Every prediction comes with 90% credible intervals. Intervals widen where the model has less data and tighten where type signals are strong.
• Transparent and interpretable. Every prediction traces back to specific types, their shift patterns, and the polls that influenced them. Not a black box — inspect it on the map.
• Free data only. No proprietary datasets, no paid subscriptions. Every source listed below is publicly available, making the model fully reproducible.

WetherVane uses exclusively free, public data. No proprietary datasets or paid subscriptions.

WetherVane is in active development, targeting the 2026 midterm elections. The model currently covers all 50 states and DC, tracking 33 Senate races.

The poll scraper runs weekly, ingesting new polls and updating race forecasts automatically. Individual race forecasts are available on the forecast page.

Planned improvements: BEA regional economic data, FEC donor density features, richer poll ingestion with crosstab disaggregation — crosstabs tell us which types were sampled, so a poll oversampling college-educated voters should pull harder on types with high college-educated membership.

Built by Hayden Haines.

Methodology inspired by The Economist's 2020 presidential model (Heidemanns, Gelman & Morris). Type-covariance architecture adapted to shift-based community discovery.

Election return data from MIT MEDSL and Algara & Amlani (Harvard Dataverse). All other data sources are listed above.