Election polls

Analysis methodology

An open description of data selection, pollster balance and recency weighting. These free analyses are experimental and descriptive.

Comparability comes first

We include source-checked polls with a known sample, fieldwork dates, publication date, method and treatment of undecided voters. Telephone, face-to-face, online and mixed methods are separate groups, as are redistributed, separately reported and excluded undecided responses. Matching metadata alone does not establish that questions and target populations match; sources also require editorial review.

Three methods, the same polls

Equal weighting counts each study once. Pollster balancing averages within each pollster first, then combines pollsters equally. Recency weighting uses a 30-day half-life from fieldwork end within each pollster, then combines these summaries using the recency of the pollster's latest study. Frequent publication alone does not multiply a pollster's contribution.

Formula and version

Time weight = 2^(-days since fieldwork end / 30). Pollster share = the time-weighted average of its studies. Pollster contribution = the time weight of its latest eligible study divided by the sum of pollster weights. The combined result sums pollster shares using these contributions. Model: organization-time-v1. The 30-day half-life is a modelling choice, not a claim to the scientifically optimal window.

Why are reliability coefficients not available yet?

We do not convert external ratings directly into accuracy coefficients. Historical accuracy requires sourced pre-election polls, comparable election types and windows, final results and out-of-sample validation. Pollsters with limited history are not automatically penalized. This version shows recency and pollster balance; large samples do not automatically earn an accuracy bonus.

Missing data and historical views

Missing party shares are not treated as zero. Duplicate records with the same pollster, fieldwork, sample and results are counted once. Each point uses only polls known to have been published by that date; future records cannot enter earlier points. The chart is calculated from today's corrected archive, rather than a record of forecasts we published in the past.

Interpreting the result

Aggregate series require at least 3 polls from 2 pollsters. This is a display threshold, not a guarantee of representation or statistical sufficiency. Lines do not bridge insufficient periods. Differences between methods are neither confidence intervals nor election win probabilities. When filters change, first check which records changed; do not attribute the entire difference to political movement.

AAPOR Disclosure Standards · A reference for transparency about sources, questions, populations and methods; it does not endorse this model.