Rankings

Difficulty-adjusted skill, with uncertainty

Our main board — pollsters, forecasters and VotePredictor together. The one metric they share is win-call Brier (a pollster's win probability is derived from its margin; forecasters publish one directly). A Bayesian two-way model adjusts for race difficulty, the bar is each entity's 90% credible interval (wide = few races), and the board is ranked by a conservative estimate so thin coverage can't top it. Pick an office; lower (left) is better. Sharp forecaster probabilities lead here — switch to Pollsters for the margin-accuracy view.

VotePredictor+ is a meta-forecaster: it stacks our own model with the published expert consensus, leaning on the human shops where they've historically been sharper (the House) and on the model where it has (Senate, President), with the blend weight fit on prior cycles only. Because it ensembles the very forecasters it sits beside, it's flagged separately and isn't part of the fair model-vs-forecaster comparison.

VotePredictor+ensemble goes one further and folds in the pollster consensus too — but it trails VotePredictor+, and that's the interesting part: the poll signal is already inside the model, which de-biases and recency-weights the same polls far better than a raw pollster average. Adding that average back is redundant and adds no lift — evidence that the gains live in the forecasters' judgment, not in re-counting the polls.