1) Randomness. Never rule out randomness.
2) Movement in and out of the "undecided" column, which is different from crossing from one candidate to the other.
3) The Clinton Foundation.
4) Trump hasn't had a Trump moment for a few weeks.
Those are the easy ones, but why can't political science explain this? Here's what we can do. We can look at well-designed surveys, and tell you what distinguishes those who support Clinton from those who support Trump, and what distinguishes those who vote from those who don't. Right now, we don't have that. We have, instead, a slew of weakly-designed polls that can only be taken in aggregate, partly because of weak design, and partly because the full data are proprietary, and not released publicly.
Some surveys allow us to run a few cross-tabulations, confirming basic facts, such as the fact that whites lean Trump, and minorities lean Clinton, etc. But we knew that. Explaining movement is harder.
Consider, for example, Trump's recent visit to an African-American church. Trump won't win black voters. The general interpretation was that it was an appeal to some category of white voters who are nervous about his rhetoric on race. Suppose that we could construct a demographic profile of that hypothetical nervous white voter. If so, then with full access to well-designed survey data, we could test whether these voters have become more supportive of Trump over the last couple of weeks, explaining the closing of the Clinton-Trump gap, if we have similarly constructed surveys done over a period of several weeks.
Right now, we just don't have the data, so it's all speculation.
Who will have the data, and when? To do that kind of analysis, only one survey: The Annenberg Center, out of the University of Pennsylvania. Will that hypothesis pan out? I doubt it. My reference point for the general election is still John Sides & Lynn Vavreck's The Gamble. Following their model of a primary, Trump should have gone the way of Herman Cain, in my opinion, but I suspect that general election poll fluctuations will match campaign events this year to the same degree they did last time. We'll see, though, and eventually, the Annenberg people will put out a book throwing massive amounts of data at every conceivable hypothesis.
Right now, though, the polls are tightening. What's going on? Anyone who claims to have an answer is either overconfident, or just bullshitting you. Clinton is ahead. Her lead has slipped a little bit. That's what we know.