TI was in Detroit the Wednesday before election day. After months of interviews, focus groups, and surveys, I asked my last question to a voter in this election.

Sheree is a 31-year-old black single mother. He voted for Hillary Clinton in 2016, voted in 2020, and remained undecided in 2024. During our 90-minute interview, we discussed his concerns about Donald Trump (“He’s so aggressive”), Kamala Harris (“Fake”). border (“It really is a disgrace”), education (“I don’t want my daughter to worry about sharing a bathroom with a boy”), and economics (“I feel stuck”).

But even though she’s never voted Republican before, when asked to make a choice, Sheree told me she’s leaning toward Donald Trump. Primary reason? Economy: Felt better under Trump than under Joe Biden. He also thought the former president was more original than Harris. “Donald Trump may be a wolf. But he’s honest about it. “Kamala is a wolf in sheep’s clothing,” he said.

Sheree supported Trump on Tuesday, as did a third of her fellow ethnic minority voters; This was the best result a Republican candidate has had among nonwhite voters in decades.

It was just an interview, but the conversation with Sheree felt like a very brief moment to me.

It definitely made me feel better about our late pre-election polls and predictions. Our last poll predicted a three-point lead for Trump in the popular vote; It was the biggest prediction of any pollster, and we were one of the very few to have him win the national vote share. Our final forecast showed a larger margin in the electoral college than any other unit. We’ve been saying Trump would win since September 24th, and for 25 of the last 30 days we had a 312 to 226 electoral vote prediction for Trump. As Harris’ chances improved on Tuesday morning, my pollster suddenly felt like a clear hill to perch on.

The coldness soon subsided. JL Partners It was the most accurate survey and modeling company not only in the country but also globally. Many fell short, refusing to support a position or predicting a Harris victory. This is a testament to the amazing team my co-founder, Dr Tom Lubbock, and I work with every day.

How did we do this? First, instead of using a single method to reach voters, we used a mixed-mode approach. Our research found that telephone-only surveys give too much weight to those willing to give pollsters the time of day over the phone: our statistics show that such a person is more likely to be an older, white, liberal woman. This was a reversal of polling by Ann Selzer that showed Harris winning Iowa by three points (Trump won the state by 13 points).

Online-only surveys also oversample. They’re getting a lot more people who are educated, more engaged, younger, and working from home. All of these groups are more likely to vote Democratic.

Instead, using an in-house algorithm, we blended our methods depending on the audience we were trying to reach. We used a combination of live call to mobile, live call to landline, SMS to web, online panel, and in-app game survey. Each method reached different types of voters; For example, the latter, which incentivizes survey responses by giving in-game points as people play games on their phones, is more likely to recruit young men of color. This in line with the results showed that 30 percent of our non-white sample supported Trump.

We caught up with elusive Trump voters who are less likely to trust polls or have the time to fill out surveys by meeting them on their terms as they go about their daily lives.

We also made bold but firm calls about the composition of our samples. For example, our polls continued to provide us with a significant share of non-voters in 2016 and 2020. In our last survey, the figure was a surprisingly high 17 percent. The sample was predominantly rural and voted heavily for Trump by a 20-point margin.

Other pollsters scaled or weighted these numbers. They assumed they couldn’t be real. But we didn’t herd or guard, and acted with the data even when it challenged our assumptions. We were right to do so, and the increase in turnout, especially in rural areas, helped Trump reach the most favorable demographics.

When Harris entered the race, we made another decision: We would not give Harris the typical incumbency bonus that almost all other forecasting models give. We received criticism for this, but it was the right decision: Including such a bonus would have tipped the national winning odds in Harris’ favor by about five points.

This shouldn’t have been a surprise to modelers; Harris was not the incumbent. In fact, just as incumbents have been ousted one after another around the world since 2020, at many points in the race our polls were showing the incumbent — President Biden — hurting his chances.

We have also learned from our experience in the UK general election earlier this year, where we were one of the most accurate pollsters.

When a voter says he doesn’t know how to vote, most pollsters drop him from the sample entirely. Instead, we used a model trained from demographic data and other answers participants gave in our survey to predict how they would vote. For example, if you said in our survey that you did not know what you would vote for, but that you cared about the economy and rated Trump as better in this regard, we classified you as a Trump voter.

As in the UK general election, where Labor was ahead by a narrow margin compared to other pollsters, this approach got us closer to the final result.

These decisions may feel like subjective judgments. To some extent it is: Polling is a science, but you also need to make calls to decide how to deal with questions like incumbent bias, what your samples look like, and how you’ll treat undecided voters.

The difference for us was that we didn’t make these decisions based on the wisdom of commentators or our own intuition. Because we had conducted hundreds of 90-minute face-to-face interviews with voters over the past two years, we were confident that these decisions were based on the real picture.

These lengthy interviews were also important for understanding people of color who voted for Trump: A 10-minute speech does not reflect their voting intentions. They started the interview by leaning towards Harris, but ended it by saying they would vote for Trump; This is a better indication of where they will go once their worldviews on economics, boundaries, and family values ​​become clear.

Sheree can only be one voter. But our conversations with people like him grounded our analysis in the reality of this election.

I talked to a fortune teller from North Carolina, climbed the dunes of Nevada with an undecided voter, even sat with a nudist from Iowa for two hours. I often questioned what I was doing. But by combining cutting-edge statistical skills with time spent in the field with ordinary Americans, we have managed to succeed where others have fallen short. We weren’t herded or attacked by groupthink, but we stuck to our guns and said the outcome was right.

James Johnson is a New York City-based pollster and was previously an advisor to Theresa May. He is the co-founder of the survey company JL Partners.