Donald Trump has a 53% chance of winning the most popular presidential election Election forecast from the Silver Bulletin declared midweek.

Fifty-one percent, to counter its main competitor, 538.

True obsessives scrutinize the forecasts daily, looking for any sign that their favorite candidate is getting closer and closer to the White House.

But even casual consumers of political news come across the latest from the Silver Bulletin and 538 on cable news, in the newspaper or on social media.

They’ve become a mainstay of public life—orderly, if often disturbing, estimates of who’s up and who’s down in the Battle for America’s Soul.

But the models have had some rough times over the years, most notably in 2016 when they appeared to underestimate — and in some cases grossly underestimate — Trump’s chances of winning his first presidential run.

And more recently, as these predictions have come to occupy an ever-increasing space in our political imaginations, some bigger questions have arisen.

Are they really the advance in the predictive arts their owners claim?

And perhaps more importantly, should we be concerned about what they are doing to our political culture?

Nate Silver has always been obsessed with numbers.

“When I took him to kindergarten one time, I dropped him off and he announced, ‘Today, I’m a number machine,’ and he started counting,” his father, a professor of political science at the University of Michigan State Rep. Brian Silver. “When I picked it up two and a half hours later, it was ‘two thousand one hundred and twenty-two, two thousand one hundred and twenty-three.’ . . ‘”

By age 11, he was building a sophisticated analysis to determine whether baseball stadium size affects attendance (it doesn’t). And after college, he sold a model predicting player performance to the statistical office Baseball Prospectus.

Then, in 2007, his career took a turn.

Watching cable news about the Democratic presidential primary, he couldn’t believe how thin everything seemed to him.

So much blovia. So little data.

And when the pundits bothered to cite the polls, they often got it wrong—they latched onto an outlier poll here or there to herald a new “momentum” for a candidate who may be going nowhere.

So on his website FiveThirtyEight — named after the number of votes in the Electoral College — he built something different: a prediction machine that aggregated polls, weighed state demographics and assessed voting patterns going back decades.

And he achieved a small measure of fame when he correctly predicted how 49 of the 50 states would vote in the 2008 presidential election—and then surpassed 50 for 50 in 2012.

The day after the election, the Christian Science Monitor went so far as to ask if Silver had “destroyed survey.”

By 2016, plenty of news outlets were publishing election forecasts and generally rated Hillary Clinton as the favorite.

But Silver had some cover when the shocking result came out.

He had given Trump a 29% chance of victory, even though other forecasters gave him a 1 or 2 percent shot. “Those are really different answers,” Silver later he begged The Washington Post. “One says, ‘Look, Trump is going to win the election every time a good baseball player gets a base hit.’ And one says, “This is a once-in-a-blue-moon scenario.”

It was a reasonable point.

But an uncomfortable question lingered: Can a model that gives a candidate a 29 percent chance of victory really claim to be “right,” in a way, when the heaviest underdog wins by more than 70 votes electoral?

It would have been “right” at 20 percent? At 15?

And how reliable were the metrics at the heart of those predictions, anyway?

Read Silver long description of the 2020 presidential election model he built for FiveThirtyEight — a model he moved, mostly intact, to the Silver Bulletin last year – and you’re impressed not just by how carefully thought out the algorithm is, but also by how many judgment calls it’s built into.

Silver had to estimate how COVID-19 would affect the election; decided to split a “home state adjustment” for Trump between the state in which he built his business (New York) and the state of his official residence (Florida); and created an elaborate “uncertainty index” that measured, among other things, “major news volume, as measured by the number of full-width New York Times headlines within the last 500 days, with more recent days weighted more.”

One could be forgiven for trusting the assumptions of a model that seems to work pretty well most of the time; Silver’s final forecast for the 2020 election gave Joe Biden an 89 percent chance of defeating Trump.

But a recently published paper by Justin Grimmer of Stanford University; Dean Knox, University of Pennsylvania; and Dartmouth College’s Sean Westwood argues that there is really no way to judge the effectiveness of an election forecasting model in the short or medium term.

If you want to measure the effectiveness of a financial or weather forecasting model, you can test it against millions of real-world observations.

But there is no comparable data set for presidential elections, which are relatively rare events; now we are nearing the end only the 60th presidential contest ever.

Indeed, the researchers found that it would take at least 24 presidential elections (over 96 years) and as many as 559 presidential elections (over 2,236 years) to confidently say that an electoral model is better at predicting winners than a simple coin toss.

Differentiating between one sophisticated model and another could take even longer.

The electoral forecast cannot be wished away.

People will always try to see into the future.

And Andrew Gelman, a Columbia University statistician who consulted on the development The Economist the magazine’s electoral forecasting modelsays you should do it as best you can.

Better to aggregate the polls, pull out the economic data that seems to move voters, and use what we know about past voting patterns than issue ill-informed summaries of the latest single poll in Michigan or Arizona.

“As the great baseball analyst Bill James once said, ‘The alternative to good stats is no stats — it’s bad stats,'” Gelman says.

But good stats don’t always land as they should.

Electoral models are built on probability: Candidate A has a 55% chance of winning. But in a laboratory experiment published 2020, Dartmouth’s Westwood; Solomon Messing, then of data outfit Acronym; and Yphtach Lelkes of the University of Pennsylvania found that more than a third of people who see these forecasts confuse probability with vote share.

In other words, if they read that Candidate A has a 55 percent chance of winning, they think that means he has the support of 55 percent of voters.

They are very different proposals. A candidate with a 55 percent chance of victory is in a virtual election. But when polls show a candidate has the support of 55 percent of voters — to just 45 percent for the opponent — she has a very solid lead.

However, this is not the most worrying finding in the study.

It also suggests that election forecasts could reduce voter turnout. Potential voters who learn that a candidate has a high probability of winning or losing are more likely to jump to the polls—thinking their vote won’t make any difference to the outcome.

A critical finding: probabilities, however unfathomable to some voters, seem to carry special weight.

When potential voters in the experiment saw something like the projected vote share instead—in other words, when they saw something like simple poll results rather than a probability score—they were just as likely to vote as before. .

Of course, this is just a study, done in a lab. But it suggests a way forward.

Instead of producing opaque and hard-sounding probabilities, it may be better to focus on simply aggregating polls, as publications like The New York Times do.

It’s easier for voters to understand what a data journalist means when he says he combined all the polls and found that, say, 51% of voters support Trump and 49% support Kamala Harris.

Grimmer, the Stanford political scientist, says focusing only on polls also creates opportunities for greater transparency, making it easier to reveal the decisions forecasters make about which data to prioritize.

“You could have a little widget on the site,” he suggests, allowing readers to “weight the polls differently and get a different result.”

Whether a news channel decides to show poll aggregation or a more complicated forecast, says Jay Rosen, a journalism professor at New York University, it’s important to keep all forecasts in proper perspective.

Stores should make it available, but “it shouldn’t be the focus” of their campaign coverage, he says.

Instead, reporters should cover the issues voters care about.

The focus, as Rosen likes to say, should be “not the odds, but the stakes.”

Of course, most major media organizations can’t resist taking a chance. In who’s up and who’s down.

But that doesn’t mean you have to.

The next time you come across a forecast, linger a moment if you must. But recognize what he can tell you and what he can’t.

Then it’s on to something meatier.

David Scharfenberg can be reached at [email protected]. Follow L @dscharfGlobe.