Why averaging reduces noise
Individual polls contain sampling error. A poll of 1,000 voters has a margin of error around 3 percentage points simply by chance, even if conducted perfectly. When multiple polling firms survey the same race over days or weeks, their results bounce around a true underlying value. Averaging many polls reduces this noise: a 10-poll average of the same race is far more stable than any single poll.
This is why news organizations and political analysts report poll averages rather than individual surveys. A single poll showing a 10-point lead might be an outlier; if five polls show the same 10-point lead, the aggregate is reliable.
Weighted aggregation and methodological bias
Simple averaging works better than individual polls, but it assumes all polls are equally reliable. Weighting the average accounts for differences in sample size, polling methodology, and historical accuracy. Some firms consistently oversample one party or systematically bias their results; weighted aggregation penalizes these firms. Recency weighting also matters: recent polls reflect current conditions better than old ones.
Advanced aggregators like FiveThirtyEight adjust for house effects (systematic biases of certain pollsters), correlations between polls (polls don't move independently), and trends over time. These refinements further reduce error but require care: overfitting to historical patterns can introduce new errors if the current election is truly different.