Sean Trende of RealClearPolitics ran through an experiment on Twitter, meant to demonstrate why big shifts in polling may not actually reflect reality.

Imagine a scenario in which you have one poll a day tracking the presidential election. Let’s say the true state of the race is that it is tied, and that poll has a margin of error of 3 points. On Monday, you might get a result in which Donald Trump leads by 4. That’s within the margin of error: If the poll estimates that support for Hillary Clinton is 3 points lower than it actually is and 1 point higher for Trump, you get a 4-point spread. The next day, the poll says Trump leads by 1 point. The next day, Clinton is up by 2.

To an outside observer, it looks like Clinton has gained 6 points in two days. In reality, though, we know that the state of the race hasn’t changed at all, because that’s the rule we set.