A polling average combines every recent poll of a race into one estimate, so a single odd survey can’t swing your read of the race. The aggregator decides which polls count, gives newer and larger polls more weight, sometimes adjusts for pollsters that lean one way, and reports the result as a number and a trend line. The details differ from site to site, which is why two averages of the same race rarely match to the decimal.
Why is an average better than a single poll? #
Any one poll is a sample of perhaps 600 to 1,500 people, and its margin of error on each candidate’s share is typically 3 to 4 points. On the gap between two candidates the uncertainty is roughly double that. A poll showing a 3-point lead is often telling you the race is close and not much else.
Polls also disagree for reasons beyond chance: different populations (adults, registered voters, likely voters), different contact methods, different weighting choices. Averaging many polls cancels out much of the random noise and some of the method differences. What it can’t cancel is an error every pollster shares, which is how 2016 and 2020 state polling missed in the same direction. More on that below.
How is a polling average calculated, step by step? #
Most averages follow the same five steps, even if the math inside each step varies.
1. Decide which polls get in #
Every aggregator has inclusion rules. Common ones:
- The poll must disclose who ran it, when it was in the field, the sample size and the population.
- Straw polls, website pop-up polls and social media polls are excluded because respondents choose themselves.
- Tracking polls that release overlapping samples every day are handled so the same respondents aren’t counted several times.
- Campaign and party internal polls are either excluded or kept with a penalty, since they tend to be released when they look good for the sponsor.
2. Weight by recency #
A poll from last week says more about today than a poll from last month. Aggregators either drop polls after a fixed window (often 2 to 4 weeks) or fade them out gradually. A linear fade is the simplest version: a poll from today counts fully, a poll from halfway through the window counts half, and a poll at the edge counts close to nothing.
3. Weight by sample size #
Bigger samples have smaller sampling error, but the gain shrinks as samples grow: a 2,000-person poll is not twice as precise as a 1,000-person poll. So many averages weight by something like the square root of the sample size, and some cap the weight so one huge online poll can’t dominate.
4. Handle prolific pollsters and house effects #
If one firm polls a race every week and everyone else polls once a month, a naive average becomes mostly that firm’s opinion. Aggregators deal with this by counting only a pollster’s latest poll, or by splitting its weight across its releases.
Some also adjust for house effects, a pollster’s consistent lean relative to the consensus. If a firm has run 2 points more Republican than the average all cycle, its numbers may be shifted 2 points before they go in. We cover the idea in what a pollster house effect is.
5. Weight by pollster quality (optional) #
Some averages give more weight to firms with better track records or more transparent methods. This is where averages differ most, because quality ratings are judgment calls. Pollster ratings explains how those grades are usually built.
A worked example of a weighted average #
Take three polls of the same Senate race, all showing Candidate X’s share.
| Poll | X’s share | Age | Recency weight | Size weight | Combined weight |
|---|---|---|---|---|---|
| A | 50% | 2 days | 0.93 | 1.0 | 0.93 |
| B | 46% | 10 days | 0.67 | 0.8 | 0.54 |
| C | 54% | 25 days | 0.17 | 0.6 | 0.10 |
The recency weights here are linear across a 30-day window (1 minus age/30). A plain average of the three is 50.0%. The weighted average is:
(50 × 0.93 + 46 × 0.54 + 54 × 0.10) ÷ (0.93 + 0.54 + 0.10) = 76.74 ÷ 1.57 ≈ 48.9%
Poll C’s 54% barely moves the result, because it is old and small. If a fresh large poll showed 54%, the average would move a lot more. That is the whole point of weighting: the average should follow the newest, best information without jerking around every time an outlier lands.
Simple averages vs. weighted models #
| Simple average | Weighted model | |
|---|---|---|
| How it works | Mean of the most recent polls in a window | Recency, size, pollster and sometimes house-effect adjustments |
| Transparency | You can check it with a calculator | Depends whether the method is published |
| Weakness | A burst of polls from one side can tilt it | Adjustments are judgment calls you may not see |
| Speed | Jumps quickly when a new poll arrives | Smoother, can lag a real shift by days |
RealClearPolling is the best-known simple average. Silver Bulletin, The New York Times and Decision Desk HQ publish weighted averages. FiveThirtyEight, long the most cited weighted average, was shut down by ABC News in March 2025, so older articles that link to it now point nowhere useful.
Neither approach is “right.” A simple average is easy to audit but vulnerable to whoever publishes the most polls. A weighted model is harder to game but asks you to trust choices you often can’t see.
How the Election Tracker average is built #
Because readers ask where our numbers come from, here is the method in Election Tracker, the free iPhone app this blog supports. Each race with enough polling gets a 30-day average built from three stated rules:
- Duplicates collapse. If one pollster polls the same race several times in the window, it counts once, using its largest sample. One busy firm can’t swamp the average.
- Newer polls weigh more. Linear recency weighting across the 30 days.
- A minimum number of appearances. A candidate has to show up in enough polls before being averaged at all, so a name tested in one survey doesn’t get a misleading line.
The app shows the average for Senate races, the generic congressional ballot, governor races and both parties’ 2028 primary fields, with a trend line once a race has enough polling over enough time. Every poll behind the average sits underneath it with the pollster, sample size, population and field dates, so you can see what you’re looking at. It doesn’t adjust for house effects or grade pollsters; if you want that, use it alongside a weighted model.
Where polling averages go wrong #
- Shared errors. If every pollster under-samples the same group, the average inherits the miss. In 2020 the average in several Midwestern states overstated the Democratic margin by several points because most polls made the same mistake.
- Herding. Pollsters who nudge their results toward the consensus make the average look more certain than it is. See polling herding.
- Lag. An average built from polls fielded over the last two weeks can’t show a shift that started yesterday.
- Thin races. A House district or a down-ballot governor race may get two or three polls all year. An “average” of two polls is not much better than one.
How to read a polling average without fooling yourself #
- Look at the trend, not the day. A 0.5-point move overnight is noise. A 3-point move sustained over three weeks is a signal.
- Check how many polls are in it. An average of 35 polls is sturdier than an average of 4.
- Check the population. Averages mixing registered-voter and likely-voter polls can shift when pollsters switch to likely-voter screens in the fall. Registered vs. likely voters explains why.
- Remember undecideds. A 46-44 average with 10% undecided is less settled than 50-48.
- Treat it as a description, not a forecast. An average says where the race stood when the polls were taken. Forecast models add fundamentals and uncertainty on top.
For the 2026 midterms on November 3, the practical routine is to watch a handful of averages rather than chase individual releases. Election Tracker lets you pin the Senate and governor races you care about to the top of their tabs and share an average as plain text with its poll count and dates attached, which is handy when a friend sends you a single scary poll.
Frequently asked questions #
Why do different sites show different polling averages for the same race? #
They include different polls, use different time windows, and weight recency, sample size and pollster quality differently. Some adjust for house effects and some don’t. Gaps of 1 to 2 points between reputable averages are normal; bigger gaps usually mean one of them is including or excluding a particular pollster.
How many polls does a polling average need? #
There’s no official minimum, but an average of fewer than about five polls is fragile, since one new poll can move it several points. Statewide Senate and governor races in competitive states usually have plenty by the fall; many House districts never do.
Are polling averages accurate? #
They are usually closer to the result than a typical single poll, but they can miss by several points when pollsters share the same blind spot. Treat a lead inside roughly 3 points as uncertain, especially in a state where polls have missed in the same direction before.
What happened to the FiveThirtyEight polling average? #
ABC News shut FiveThirtyEight down on March 5, 2025, and its averages stopped updating. Weighted averages from Silver Bulletin, The New York Times and Decision Desk HQ, plus simpler averages such as RealClearPolling, are the common alternatives.
Is a polling average the same as a forecast? #
No. An average describes current polling. A forecast turns polls, and often economic and historical data, into a probability of winning. A candidate can lead the average by 2 points and still be close to a coin flip in a forecast.