A data scientist who correctly predicted Donald Trump's shock victory over Hillary Clinton in 2016 says the US President is currently on track to win again.
Professor Bela Stantic is the founder and director of Griffith University's Big Data and Smart Analytics Lab, where he analyses social media data and sentiment to predict voters' behaviour.
In the past, those predictions have been extraordinarily accurate.
Four years ago, Stantic successfully picked the winner in 49 of the 50 American states. His lab also nailed the result of both the 2016 Brexit referendum and the Australian federal election last year.
In all three cases, public opinion polling pointed to the opposite result.
At the moment, the polls show Trump trailing his opponent, Joe Biden, by an average of 6.2 per cent at the national level. They're a bit closer in the key battleground states, where Biden leads by 3.9 per cent.
It looks like a comfortable lead for the Democratic Party's nominee. But, just like Clinton's lead four years ago, it could be a mirage.
Stantic recently conducted a preliminary, draft analysis of the upcoming US election. His lab's complete analysis, along with a final prediction of the result, will come closer to polling day on November 3.
"It is obvious again that Trump will lose the popular vote," he told news.com.au.
"However, he's tracking really well in the crucial states. Florida is a coin toss, but he's slightly ahead for me. And Minnesota and Pennsylvania as well. And then Texas, he will win easily.
"So then that gives him an edge to get about 270, 280 electoral votes.
"It is maybe early, but I can tell you that the trend we identified in advance last time is holding."
So, according to that analysis, the US is heading for an electoral map which looks something like this.
In other words, the race is close – pretty much neck-and-neck – but Trump is once again on course to lose the popular vote while winning the decisive electoral vote.
About 2.9 million more Americans cast ballots for Clinton than for Trump in 2016. However, Trump's support was distributed more efficiently.
While the Democrat racked up huge margins in populous but uncompetitive states like California and New York, Trump managed to scrape to relatively narrow victories in the states that actually mattered, such as Michigan, Wisconsin and Pennsylvania.
That gave him a 304-227 edge in the Electoral College, comfortably above the winning threshold of 270.
Stantic said the 2020 election was, broadly speaking, the same sort of race.
"It's really a coin toss. I think Florida, at the moment, is a coin toss, but Trump is just ahead," he said.
But his draft analysis dug up one particularly important – and perhaps surprising – difference between 2016 and 2020.
"I find that this time it is more polarised than last time," Stantic said.
He reached that conclusion by analysing the comments on Biden's social media posts.
"People reacted so harshly against Biden. It was 30,000-something comments, and all strongly against him," he said.
"They are saying that he cannot be trusted, that he doesn't know what he's talking about. 'At least Trump, what he says, he thinks.' Comments along these lines.
"There was not much support for Biden."
This is an interesting wrinkle, because the conventional wisdom you often hear from political experts – and occasionally, from self-important journalists – is that Clinton was a more polarising figure than Biden is.
Four years ago, Trump and Clinton both had unusually high disapproval ratings in the polls. This time, Biden's favourability rating is pretty much split down the middle.
"I feel they hate him more than Hillary," Stantic said.
He pointed out that, as potentially the first female president, Clinton could at least rely on enthusiastic support from women.
"Why would you vote for someone who has issues expressing himself, and doesn't know what he's saying? That is what people – when I looked at the topics of discussion, what is dominating," he said.
"If it were a better candidate (than Biden), Trump would lose easily."
Stantic's method is not infallible. It did, for instance, get the result of Australia's same-sex marriage plebiscite wrong, for reasons he explained in detail afterwards.
But he says his lab's analysis is more reliable than opinion polling, because it involves a significantly larger sample size.
"I think the polls are volatile because their sample size is very small. They have a thousand people, and it depends on who you interview," he said.
"I'm talking about millions of posts. Last month, I think I had 800,000 posts in one day.
"It's not just about these 800,000, but it's also that some posts have 20,000-30,000 likes."
People also tend to be more honest when expressing their opinions on social media than when a pollster quizzes them.
And the peculiar nature of America's system, which hinges on the candidates winning states (rather than, say, seats like in Australia), helps make Stantic's job simpler, because the data allows him to pinpoint exactly which state people will be voting in.
"Australia, it's a bit hard because of seats and their locations. The US election, it's easier to predict," he said.
"Users put the state in their location when they open their accounts.
"It's much easier to identify a location, and then everything else is pure mathematics."