Too soon to tell?

 Donald Trump currently has a narrow lead in the polls (about 1% in the 538 average), so all you can say now is that it could go either way.  But what if one of the candidates opened opened up a substantial lead--how much would that tell us about what is likely to happen in November?  I looked up surveys from late May/early June in presidential elections from 1948-2020 and compared the margin in the survey to the actual popular vote margin in November.*

The regression is y=.44x, where y is the election margin and x is the May survey margin.**  

There was a large residual in 1948, which was a notorious failure of the polls, followed by pretty good predictions in 1952-1968.  Then there were large errors in 1972, 1980, 1984, 1988, and 1992--in the first four of those, the Republicans did substantially better in November than the May surveys suggested, and in 1992 the Democrats did substantially better.  Since then, the predictions have been pretty good, except for 2008, when there was an obvious reason for a late shift towards the Democrats.  A graph of the absolute value of the residuals:

The sample is small, so you could plausibly claim the apparent pattern is just a matter of chance.  However, I think that there probably was a real change.  In the earlier years, there was fairly strong party identification based on ethnicity, region, religion, and family tradition.  Although candidate qualities and current conditions made a difference, the main thing that happened during the campaign is that some of the people who had been thinking of voting against their normal party would drift back "home."   Then there was a stretch when party identification was weaker, so that events during the campaign could make more difference.  That was followed by a period in which party identification became stronger again, but now based more on ideology.  

Note:  Through the 1960s, only one organization (Gallup) regularly did election polling.  The number has grown since then, and in recent years you can calculate an average based on large numbers of polls.  In order to make things comparable, I just selected one survey for each election, based partly on the date (whatever was closest to May 24, when I compiled the data) and partly on my judgement about the general reputation of different organizations.  

*Of course, the candidate who leads in the popular vote may not win the Electoral College, but that's a different issue.  

**If you include an intercept, the estimate is -1.7 with a standard error of about 1.7; the estimated coefficient for x is still .44.  A non-zero intercept could mean either a consistent bias in the polls or a tendency for the vote to shift in favor of a particular party during the campaign, neither of which seemed likely in principle.