Which polls specifically are you referring to @js82 ?
I'm referring to the 2016 projections from FiveThirtyEight(which are in turn driven by poll aggregates), which projected Clinton to have a 3.5% national margin in the popular vote(Actual: 2.1%). These same projections generally had Clinton up 5-6% in PA/MI/WI, which Trump ended up winning by <1%.
I don't think that's quite accurate.
1) IIRC, some of those states were at 3% to 4%
2) which as they said repeatedly, was within the margin of uncertainty that frequently occurs due to last minute shifts as well as polling errors
3) In any case, their analysis (not a mere aggregation) of the polls suggested that PA/MI/WI were among the states most likely to swing the election
Here's a link to their election forecast. How does it stack up on these points?
https://projects.fivethirtyeight.com/2016-election-forecast/1) PA 3.7% (not 5-6%), MI 4.2% (not 5-6%), WI 5.3% (yes 5-6%), average 4.4% (not 5-6%).
2) In the section "Who's ahead and by how much", the width of the error bars shows that all three states could go either way; you can see that because the bars' boundaries are on both sides of the center line
3a) In the same section, PA and MI are second and third on the list in terms of how small their margin is, and therefore how likely they are to be a swing state
3b) In the section "The winding path to 270 electoral votes", the winding chart physically shows Pennsylvania right next to the center line, more likely than any state except NH to be the deciding state.
3c) Winding chart also shows Wisconsin and Michigan as 2 of 3 next most likely swing states to flip in the event of a red wave - they're blue, but if the country shifts red, the order of flipping to red was expected as NH, PA, CO, MI, WI.
I agree with you that polls were off by slightly more in those states than others. Then again, the reason I like 538 is that they reason as carefully as possible about whether the details of different polls have predictable flaws, whether other factors are more accurate than mere polls, and adjust their model accordingly wherever they can find data. They're not just aggregating polls, they're analyzing them plus other data to draw inferences that are hopefully more accurate than simply averaging. In 2016 they were correct in identifying the upper Midwest states as swing states when the raw poll numbers weren't as clear.
There's a lot of time between now and November, so polls may be less predictive now anyway than they are in the final weeks. But they were reasonably accurate in 2016's final weeks in the sense that the final outcome was within the known historic typical margin of last minute variance, and the final result was clearly recognized by the model as what would probably happen in the event that the final swing of public sentiment was in Trump's direction. I think the polls (and 538's analysis) are still worth watching this year if you want to mentally prepare.