If you’re a Donald Trump supporter, this week’s CNN poll showing him leading Hillary Clinton gave some joy. But if you’re a Clinton supporter, you just had to wait five minutes for another poll to show her with a six-point lead.

In keeping with a political season that’s kept professional politicos on their toes and found them often wrong, the polls have swung wildly between Trump and Clinton as they fight for the White House. When the best pollsters in the country disagree, how can people who aren’t in politics read the polls?

RCP v. Huffington

Yesterday, The Stream interviewed a top polling analyst and a professor of polling about how to read the wild swings in the polling. Both Sean Trende of RealClear Politics (RCP) and Philip Harold of Robert Morris University told The Stream that a poll’s accuracy depends on the way it collects its information — which is why different polls give different results, and confound pundits and other observers.

So is it safe to average out the polls to get the collective best guess, the way RCP does? The media frequently cite RCP’s poll averages, but Natalie Jackson of The Huffington Post (HuffPost) said there’s a problem with averaging the polls that way. “The HuffPost Pollster charts calculate trend lines rather than traditional arithmetic mean, so the numbers will differ from simple averages used by sites like RealClearPolitics,” she said. “The reason for this is that a simple average will be very susceptible to individual polls’ deviations — one outlier poll could pull the average in a completely different direction than the rest of the polls indicate.”

In deciding what counts as a real trend, HuffPost looks at the dates the different polls were conducted, the days between the polls and the difference in what they found. In general, Jackson said, 4 or 5 polls have to change within a week to change the HP results.

Jackson said that HuffPost’s way of reading the different polls avoids RCP‘s problem by requiring that the polls in general change before it revises the direction the collected results are moving. HuffPost “looks for trends in the polls and downplays data that deviates significantly from what most data show.” The technical term for their technique is “a Bayesian model called a Kalman filter,” a technique engineers developed to distinguish between actual “signals” (real data) and “noise.”

Jackson argues that polling averages that don’t take trends into account over-estimate temporary results. “Currently, we show a wider race between Trump and Clinton than simple averages because the trend has favored Clinton, and the polls showing Trump ahead — or within a couple of points — appear to be deviations from the trend for now. However, we do have the gap closing, since it’s become clear that Trump is gaining support.”

HuffPost‘s poll tracker finds Clinton’s lead is more than two points larger than RCP‘s 3.3 points.

RCP‘s Trende said that both approaches have advantages and disadvantages. The HuffPost method “does filter out short-term aberrations, but can make it slower to pick up ‘real’ shifts when we have infrequent polling.” Without what he called HuffPost’s “smoothing function,” he said, “there isn’t a lot of difference between their findings and ours.”

Misleading Assumptions

As the argument between Trende and Jackson shows, accurate polling is hard to do. Conservatives found this out the hard way in 2012.

As Hot Air blogger and Going Red author Ed Morrissey told The Stream earlier this year, conservative pollsters — and the Romney campaign’s pollsters as well — thought the electorate had changed and that the polls showing Obama’s comfortable lead were based on “an outdated model.”

“We weren’t trying to be dishonest, we weren’t lying about this, we were just really wrong in our assumptions,” he said. “We weren’t taking the time to find out who these voters were, where they were at, what Barack Obama was doing with his organizing and media strategy in 2012. And those errors compounded on each other to give a completely false impression of what we were going to see that night.”

The same problem affected pollsters in the 2016 presidential primaries. In the Democratic primary, no one gave Bernie Sanders a chance of winning as many votes as he did and staying in the race so long.

Likewise no one expected Donald Trump to do so well, much less win the nomination. As Harold told The Stream yesterday, “Trump won the Republican primaries by bringing in new voters to the ballot box. That makes it very, very difficult for pollsters, who are relying on the statistics from previous elections, to filter out those who do not plan to vote and to weigh their samples.”

The Polls This Week

The objective versus subjective argument came up again this week with CNN’s poll that showed Trump in the lead. MSNBC determined that CNN oversampled lower-educated whites, and using 2012 voting data determined that the poll should have shown Clinton leading by four points. MSNBC assumed minorities will vote for Clinton at the same rates they backed Obama in 2012 and that lower-educated whites don’t come out in unforeseen numbers for Trump. The network could be right — but they could be wrong.

By the way: this article was drafted on September 6. Clinton had a 3.3-point lead, according to the RCP head-to-head matchup. Today, it’s just three points, according to RCP, (and just 2.1 points when the Libertarian and Green Party candidates are counted). HuffPost’s tracker likewise shows Clinton’s lead has shrunk. Who knows what it’ll look like tomorrow?