For the first time, The Economist is publishing a statistical forecast of an American presidential election, and it created the model in partnership with Andrew Gelman, professor of statistics and political science and member of the Data Science Institute at Columbia University, and Merlin Heidemanns, a doctoral student in Columbia’s political science department.

The model, which has a standing page on the magazine’s site, calculates Biden’s and Trump’s probabilities of winning each state and the national election, with projections updated daily. The Economist has also released the source code for the model, which is based on Stan, an open-source platform for statistical modeling created by a team of core developers, including Gelman and other researchers at Columbia. Stan performs Bayesian inference, a statistical method for combining data from many sources, which makes it ideal for forecasting elections where relevant information comes from an array of polls and economic and political factors.

Gelman and Heidemanns are partnering on the project with G. Elliott Morris, a data journalist at The Economist. Their aim is to predict the two candidates’ share of the vote in each state on Election Day. To do that, they are using state and national polls, along with information that has predicted historical elections. The model accounts for polling errors, along with the possibility of state- and national-level time trends in opinions. The team designed the model to consider several factors: polls, economic conditions, presidential approval ratings, political polarization, and the power of an incumbent president.

You may read discussions of the model on Gelman’s blog.

Four years ago, polling errors in key states led many forecasters and pollsters to be overconfident about Hillary Clinton’s chances of winning the election. The Economist model attempts to avoid this problem by accounting for non-sampling errors in polls in the current election. Earlier research by Gelman and other experts has found that polls are typically off by as much as twice their stated margin of errors. That said, a model is only as good as its assumption, noted Gelman.

“We didn’t model the probability of a meteor strike,” he quipped, “but the model implicitly assumes that both Biden and Trump will make it to the finish line.”

Biden has a large lead over Trump in polls, but with memories of Trump’s victory in 2016 still fresh, pollsters worry that an unexpected actuality might again trip up forecasts. Many of the uncertainties center around how voters perceive the pandemic and how the two candidates contend with it.

The Economist model, which gives Biden a 89 percent chance of winning the Electoral College, is less concerned with the current moment than with the range of possibilities reflected by the historical data. As of this week, the model says Trump has a 11 percent chance of winning more electoral votes than Biden, and that Trump has a 2 percent chance of winning more of the popular vote than Biden. The forecast predicts Biden will win between 223-429 electoral votes, while Trump will get between 109 and 315. 

“Right now, our model thinks Joe Biden is very likely to beat Donald Trump in the Electoral College,” the site reports.

Observers have noted that this election is unusual, given the pandemic, and therefore as hard to predict as the 2016 election. But when it comes to national elections, Gelman said, the unusual is usual.

“People say, ‘Well, 2020 is special,’” he said. “But you go back to every election: 1948 was the first election in a long time without Franklin Roosevelt running; 1960 had Kennedy, the first Catholic to run for president; 1968 and 1992 had strong third-party candidacies; 2008 and 2016 had the first nonwhite and female major-party nominees. You’ll see that in just about every election there has been something unusual to consider. There’s no perfect way to capture the uncertainty in the prediction of a unique event. The best we can do is be open about our assumptions and be clear about what historical data we’re using to make our forecast.”

— Robert Florida