A pollster made up its numbers. Here is what that did to our forecast, and what we changed

On October 20 a polling firm called Wedgewood Polls admitted that its numbers were made up.

The admission came after FiftyPlusOne, a polling aggregator, asked the firm for the raw data behind its October 12 to 14 survey of the Mississippi Senate race. Wedgewood said the survey had been run through Pollfish. According to FiftyPlusOne, the file it sent did not look like real Pollfish data: the respondent IDs and timestamps were wrong. Before FiftyPlusOne published, Wedgewood posted a statement calling itself a "fictional entity" and the whole operation "an experimental project to demonstrate the vulnerabilities, lack of oversight" in polling. It said the "vast majority" of its releases had no real data behind them. The one poll it still says was real is a crowdfunded survey of Nebraska.

Wedgewood had been publishing since June. It said it was based in Boston, but Massachusetts business records show no such company, and it gave two names, Jack Sims and Andrew Prater, as the people running it. The New York Times removed 15 Wedgewood polls from its database and Decision Desk HQ 17. Silver Bulletin, VoteHub and FiftyPlusOne dropped them too.

The polls did real damage. In June Wedgewood put Mallory McMorrow at 9% in Michigan's Democratic Senate primary, with only 6% undecided. At the same time Susquehanna had 49% undecided. McMorrow ended her campaign on July 5. She says donations dried up after the Wedgewood numbers came out.

Wedgewood was not the only one. Sims previously ran SoCal Strategies, which, according to its state filing, he dissolved in May. Polls kept coming out under SoCal's name afterward. After reporters asked about them, its website and social media accounts disappeared on October 22, and FiftyPlusOne stopped collecting its polls. In August a third firm, Median Strategies, admitted it had made up its polls of the Los Angeles mayoral race and the Wisconsin governor's primary.

What it did to our forecast

We held 25 Wedgewood polls. They came in the way most polls reach us, through Wikipedia's poll tables and VoteHub's list. Of those 25, 22 were general-election polls of Senate, governor and House races and 3 were primary polls.

In most races they barely counted. Ohio's Senate race has nearly 30 polls in our average and Texas's about 50, and in those races a Wedgewood poll carried less than 5% of the weight. Taking them out moves those races by a fraction of a point.

Mississippi was the exception. Our Senate average there rested on three polls, and Wedgewood's (Cindy Hyde-Smith 52%, Scott Colom 48%) carried 52% of the weight. A real poll that close would matter, because it is far closer than Mississippi usually votes. With it in our forecast, Colom had a 15% chance. Without it, he has 8% today.

Three other races moved by a point or more:

• Kansas Senate: Adam Hamilton had 45% with Wedgewood's poll in; he has 31% today without it.
• Kansas governor: Cindy Holscher had 18%; she has 10% today.
• Nebraska governor: Lynne Walz had 11%; she has 8% today. Here the fake poll ran against the Democrat. It had Jim Pillen 17 points ahead, more than any other poll of the race.

Control of the Senate barely moved: the Democrats had a 57% chance on Sunday with the Wedgewood polls in, and have 38% without them. Only one of the polls was of a House race, Nevada's 2nd district, and the Democrats' House chance, 77%, did not move.

One thing stands out. Wedgewood's general-election numbers were not lopsided. In 13 of its 17 polls of races we forecast, its margin was within 3 points of the other polls of the same race. It leaned to the Democrats in 8 of them and to the Republicans in 9. Made-up polls that look like everyone else's are hard to spot from the numbers. They were found by asking for the raw data.

What we changed

The polls are gone. We have removed all 25 Wedgewood polls. We have also removed the 9 SoCal Strategies polls fielded since the firm was dissolved in May. Its 14 polls from 2024 and 2025 stay, because nothing has been shown to be wrong with them.

They stay gone. Our readers now refuse both firms, so neither can come back through Wikipedia or VoteHub, however long they stay listed there.

New pollsters are marked. On every race page, a poll from a firm with no record now carries the label "New, no record". That means no poll in our records before December 2025, no final polls at past elections, no membership of AAPOR's Transparency Initiative or the Roper Center archive, and not a university. Every poll already shows its share of the average, so you can see when one poll carries half of it, as Wedgewood's did in Mississippi.

What we tested and did not change. We tested whether new firms should count for less. We took the final polls of 230 Senate, governor and presidential races from 2018 to 2025 and marked each firm as new or established as of that election. New firms' final polls were further off: on average they missed the result by 5.2 points, against 4.6 for established firms. But counting them at half weight did not make our averages more accurate. The race averages missed by 4.41 points instead of 4.40, and leaving new firms out altogether made it 4.48. New firms are sometimes right where the established ones are wrong, and averaging already limits what one poll can do. So new firms keep full weight.

The weak spot is a race with few polls. Our forecast already allows for more error when an average rests on one or two firms, which is why Colom's 15% was not higher. But one poll can still carry half of a thinly polled race. Where it does, check who that poll is from. Our method page sets out the weights.