America's presidential election is less than 100 days away, which means it is finally time to start talking about one of the most contentious subjects in politics – the polls.
Fair warning, this is going to get nerdy.
The short summary here is that the polls are not looking good for President Donald Trump, but the full story is not that simple. Can those polls even be trusted?
We all remember what happened in 2016, when they told us Hillary Clinton would win. That came right after the shock of Brexit, and was followed by an extraordinary polling failure in the Australian federal election last year.
So, we should obviously approach this topic with dollops of salt.
That said, it is becoming increasingly hard to ignore just how consistently terrible the polls are for Trump.
RealClearPolitics' national polling average has the President trailing his opponent, Joe Biden, by 8.3 per cent. Biden leads the average 49.9-41.6.
Trump has not been ahead in a single national poll since February. Out of the last 94 polls, Biden has led in 92, with two ties.
And I should stress this is not merely a case of history repeating, with 2016's numbers happening all over again. It's worse.
At this time four years ago, Clinton only led Trump by about 1 per cent in the polling average, 44.5-43.4. At no point after April of 2016 did she ever enjoy a margin as large as the one Biden has now.
The Trump-Clinton race was much more volatile, with wild swings in the polls as the election approached. Three times, Clinton opened a significant gap over Trump, and all three times he closed it again.
The final polling average before election day had Clinton ahead by 3.2 per cent. She ended up winning the popular vote by a more modest 2.1 per cent, while losing the decisive electoral vote.
'Those are fake polls'
The President himself has a simple explanation for all the polls that show him losing to Biden – they're fake.
"I'm not losing, because those are fake polls," Trump insisted during an interview with Fox News anchor Chris Wallace earlier this month. Wallace had asked him about a Fox News survey which had him trailing by eight points.
"They were fake in 2016 and now they're even more fake," the President said.
"You see what's going on. I have other polls that put me leading, and we have polls where I'm leading. I have a poll where we're leading in every swing state.
"First of all, the Fox polls – whoever does your Fox polls, they're among the worst. They got it all wrong in 2016. They've been wrong on every poll I've ever seen."
It is, perhaps, a little rich for Trump to call the polls "fake", given his peculiar habit of citing polls that don't exist to make his approval ratings sound better than they actually are.
But let's put such quibbles aside.
The more consequential part of Trump's answer was his claim that his re-election campaign's internal polling numbers actually show him winning.
This is the same argument we heard from the New Zealand opposition this week after a particularly bad poll came out – that poll must be wrong, because our internal numbers are way better.
It has become Trump's default response to questions on the subject. He repeated it at a media conference on Monday.
"The poll numbers we have are very good," the President said.
"We're leading in North Carolina. We're leading in Pennsylvania. We're leading in Arizona. Our numbers – we're leading nicely in Florida. I think our poll numbers are very good. We are leading substantially in Georgia."
Take note of the specific states he mentioned. We'll come back to them shortly.
The key to victory
Under America's electoral system, the raw number of votes you receive at a national level is less important than where you win them.
Each state is worth a certain number of electoral votes. Win the popular vote in that state, and you get all its electoral votes. The first candidate to 270 electoral votes becomes president.
So, while 2.9 million more people voted for Clinton than Trump in 2016, he won the election, because his votes were distributed more efficiently between the crucial swing states.
In all, Trump claimed six states that had been held by Barack Obama at the previous election: Florida, Pennsylvania, Ohio, Michigan, Wisconsin and Iowa. Biden's task is to win some of them back. He does not need all of them.
Let's return to the RealClearPolitics polling averages, this time at a state level.
Among the states Trump won last time, Biden is currently ahead in North Carolina (by 3 per cent), Ohio (1.5 per cent), Wisconsin (6.4), Florida (7.8), Michigan (8.4), Pennsylvania (7.4), Arizona (4) and Minnesota (11.4).
If those numbers were reflected on election day, Biden would win easily, with more than 350 electoral votes.
You can see why Republican political operatives are giving panicky quotes to reporters. Not only is Trump at risk of losing five swing states, he's even running behind in states Mitt Romney managed to win in 2012.
On top of that, Democrats are increasingly confident they can win enough Senate seats off the Republicans to take control of both houses of Congress.
"Things are pretty bad," one Republican strategist with a talent for understatement told Vice News this week.
Another put it more bluntly: "It's a whole sh*t sandwich and we all have to take a bite."
"If the election were held today, we're talking about a wipeout. We'd be in landslide territory," said another.
Going back to Trump's answer from earlier, you'll remember he claimed to be leading in Arizona and Georgia.
The fact he even mentioned those states is a bad sign. They shouldn't be part of the conversation. No Democratic candidate has been competitive in Arizona or Georgia, let alone won them, since Bill Clinton in the mid-1990s.
And yet, we know the Trump campaign is running TV ads in both states. Money and effort that could be spent elsewhere are instead being used to defend states that should be easy wins for the President.
If the internal polling were as rosy as Trump claims, that would not be necessary.
But can the polls even be trusted?
We return to the core question. What's the point of talking about the polls if they can't be trusted anyway?
Much of the current suspicion towards public polling is rooted in the 2016 election result. Some of it is fair; some is overblown.
We've gone from trusting the polls as though they're a perfect reflection of reality to discounting them entirely, when the wiser path is likely somewhere in the middle.
First of all, it's important to distinguish between polls and forecasts – two completely different things which are often conflated.
Four years ago, several publications in the United States used polling data to produce a forecast of the election's outcome. These models, almost without exception, declared Clinton the overwhelming favourite.
FiveThirtyEight's final forecast said she was 71.4 per cent certain to win. The New York Times gave her a 91 per cent chance. For much of the campaign, the Princeton Election Consortium had her in the eye-watering 98-99 per cent range.
When Trump won, these predictions all looked very silly indeed.
But as The American Association for Public Opinion Research pointed out in its detailed post-mortem of the election, the spectacular failure of those forecasts is a separate issue from the accuracy of the polling.
"Polls and forecasting models are not one and the same," the report's authors said.
"Pollsters and astute poll reporters are often careful to describe their findings as a snapshot in time, measuring public opinion when they are fielded. Forecasting models do something different – they attempt to predict a future event.
"As the 2016 election proved, this can be a fraught exercise."
Just because the polls gave Clinton a double-digit lead in April of 2016, that doesn't mean they were wildly wrong. A lot happened between April and November, and voters have been known to change their minds – something they could still do this time.
According to the report, the most obvious factor in the final result was the number of people who didn't decide who to vote for until the last week of the campaign, or in some cases, until they were literally standing in the booth.
During that final week, there was "substantial movement towards Trump", particularly in the four decisive states Clinton ended up losing by narrow margins.
Take Florida, for example. Among voters who made up their minds before the last week, Clinton led 49-48. Among the late-deciders, Trump won by a massive margin, 55-38.
How about Wisconsin? Clinton led 49-46 until the last week, when Trump won late-deciders 59-30. He only claimed the state by 23,000 votes.
"It suggests that many polls were probably fairly accurate at the time they were conducted," the report finds (emphasis theirs).
"Both Trump and Clinton had historically poor favourability ratings. One possibility is that these negative evaluations made it difficult for some voters to decide whether to vote and, then, difficult to decide for whom to vote.
"Unhappy with their options, many voters may have waited until the final week or so before deciding, a set of last-minute changes that polls completed a week out from the election would not have detected.
"Perhaps this included those who broke late for Trump, as well as potential Clinton voters who decided not to vote because they concluded she was going to win."
The other major factor was the pollsters' assumptions about voter turnout – how many Clinton and Trump supporters they thought would actually show up to the polls.
Such assumptions are commonplace in polling, and are used to weight the results to reflect (hopefully) an accurate picture of the electorate.
The report concluded the pollsters had underestimated turnout among Trump's supporters; specifically, white voters without a college degree.
"In 2016 the presidential vote was strongly and fairly linearly related to education; the more formal education a voter had, the more likely they were to vote for Clinton. Historically, that has not been the case," it said.
So, were the polls "fake", as Trump claims? "Flawed" would be a better word.
They failed to pick up a very late surge towards Trump, and underestimated his support among white voters without a college degree. In a close election, those mistakes matter.
At the moment, at least, all indications are that the 2020 election is not close.
Joe Biden is not Hillary Clinton
In 2016, large majorities of Americans always had an unfavourable view of both Clinton and Trump.
Plenty of voters were ultimately forced to choose between two candidates they disliked, which is one reason why such an unusually high proportion of them waited until the last minute.
So far, that isn't the case this time around. There is a significant disparity between the candidates – Biden's favourability rating is pretty much split down the middle, while Trump's remains almost as dire now as it was four years ago.
There is more bad news for the President when you delve into the details.
Take the most recent poll by The Economist/YouGov. A total of 25 per cent of respondents said they had a "very favourable" view of Trump. Meanwhile, 46 per cent said they had a "very unfavourable" one.
Almost a majority of the electorate strongly dislikes the President, which means he has to win the votes of pretty much everyone else.
This trend pops up in every poll. For Trump, the problem isn't necessarily his high disapproval rating so much as the intensity of that dislike. It wasn't nearly as much of an issue four years ago, because his opponent was loathed just as intensely.
It's worth noting that among the group of voters who don't much care for either candidate, Biden has a significant lead.
The other danger sign? Far fewer voters are waiting to make up their minds this time.
As pollster Rachel Bitecofer has pointed out, 15 per cent of the electorate left its decision until the final week in 2016.
Compare that to 2020. In the most recent CBS poll, only 4 per cent of people said they weren't sure who they would choose. The Economist also had that number at 4 per cent. Fox News had it at 5 per cent.
The huge mass of undecided voters who broke so decisively towards Trump at the 11th hour four years ago, simply does not appear to exist this time.
Obviously, we won't know whether the polls are accurate or not until the election happens, but writing them off as a meaningless re-run of 2016 would be a mistake.
As things stand, they would need to be wrong on an even more egregious scale than last time for Trump to even have a chance of winning.