Methodology

Forecasts come from TabPFN, a tabular foundation model. Election forecasting is a small-data problem — a few hundred comparable races per cycle, not millions — which is exactly where this class of model is strong. For each race we build a feature row from its polling (a recency-weighted average, the trend, how much pollsters disagree, how many polls) plus the seat's prior result, and the model predicts the win probability and final margin.

The model trains on every polled Senate, Governor, House and presidential race from 1998–2025. We validated it with a strict walk-forward backtest: to forecast any past cycle, the model only ever saw earlier cycles — never the outcome it was being graded on. No result data leaks into the inputs.

• Aggregating polls is the big win. Going from a single poll to a race-level average lifts winner accuracy from ~80% to ~88%.
• The model's edge is sharper margins and calibration, not the winner call (the poll average already calls most races right). Its predicted margins beat a naive average, and its probabilities are well-calibrated.
• Calibration matters more than bravado. Over many races, our 70% calls win about 70% of the time. We'd rather be honestly uncertain than falsely confident.

A fair question, asked two ways. We raced TabPFN against a 14-model field — random forests, XGBoost, LightGBM, CatBoost, a Gaussian process, a neural net, kernel and nearest-neighbor methods, regularized linears — on the same races and the same walk-forward rules. Nothing beats it, and the plain bias-corrected poll average — no machine learning at all — beats every model except TabPFN. TabPFN is built for exactly this small-data regime, which is why it's the one learner that adds value over averaging — the aggregation does the heavy lifting, and the model is a small, real bonus on top. See the full bake-off on the AI Models page.

Each pollster is rated on its final poll before each election — the call that actually matters — against the result (see the leaderboard). The score is sample-size-adjusted, so a lucky run over a few races doesn't outrank a steady record over hundreds. Ratings are walk-forward inside the model — a pollster's rating on any date uses only its earlier polls. We use them to de-bias polls (subtract a pollster's lean) rather than to weight by accuracy, which our backtest showed doesn't help.

We also de-bias by voter screen. On full-campaign data, registered-voter and adult-sample polls overstate the Democrat by ~2 points versus likely-voter polls of the same races late in the cycle, and correcting it cut race error ~6% out-of-sample. So we shrink the Democratic margin on RV/adult polls toward their likely-voter equivalent — but only as the election nears (where the effect is real and measured), not months out where it isn't.

Senate & Governor races are forecast from polls. The 435 House districts are almost never polled, so they're forecast from fundamentals — district partisan lean, incumbency, and the national generic ballot — using a model trained on recent House results. Chamber control comes from a Monte-Carlo that draws one correlated national swing across all races each run, so it reflects the real risk that polls miss in the same direction everywhere (independent draws look far too confident).

On the rankings board, where House races do have a poll or two, we blend the two: the poll model carries well-polled districts, but on thin-poll districts (≤2 polls) we mix in the fundamentals model, because one or two polls are noisier there than the district's lean and incumbency. Both are VotePredictor's own models — no borrowing from the forecasters we're ranked against — and the blend is what lifts our House win-call ranking. It's the same idea as the poll average itself: combine signals, and trust each where it's strongest.

A poll six months out is not a poll on election eve. Using the full-campaign poll feeds across five cycle-office series (President, Senate and Governor races in 2020, 2022 and 2024 — every poll with its field dates), we bucketed polls by how far they were from the election and measured how far the average missed the final margin.

Two lessons, and they pull in opposite directions. In a normal cycle (2024), polls converge: the average miss grows from ~4 points on election eve to ~6 points three-plus months out, so earlier polls genuinely carry less information. But lead time only shrinks noise, not bias: in 2020 the polls were off by 5–7 points at every horizon — the late polls were actually the worst — because a systematic miss can't be outrun by waiting. That is exactly why our live forecast widens its uncertainty the further it is from November, and why being early is a smaller risk than the chance the whole cycle is tilted.

Pollster ratings grade the people who take polls. This grades VotePredictor's own model against the other forecasters — FiveThirtyEight, Cook, the Economist, Sabato, DDHQ, Silver Bulletin and a dozen more — using each outfit's final pre-election win probability per race (2016–2024), archived by JHK Forecasts. Everyone is scored identically: Brier score — the squared error of the win probability — on the same races and the same outcomes, against VotePredictor's strictly out-of-sample (walk-forward) calls.

The honest read: VotePredictor lands in the pack but behind the leaders, and the gap is widest on the hard, genuinely competitive races. That tracks — it is a deliberately poll-only model (no fundamentals, no hand-adjustment), competing here with its raw probability rather than the better-calibrated production version. It's the clearest evidence we have that expert priors buy something exactly where polls are thin and noisy.

• Live polls from VoteHub (CC-BY 4.0); historical polls from FiveThirtyEight and the Silver Bulletin; results + partisan lean from FiveThirtyEight / MIT.
• VoteHub omits candidate party, so we attach it from a curated crosswalk that we cross-check against Wikipedia.
• House district lean is the 2022 vintage, so recently-redrawn districts are approximate.
• This is an independent research project, not affiliated with any campaign, and not a betting product.