The breakdown of this poll is 40% D / 32% R / 27% I. 35% Liberal, 35% Conservative, 27% Moderate. That is quite a bit more Dem and liberal than the electorates that showed up in 2016 and 2018. I reweighted it to the 2018 electorates for both party ID and ideology, and here's the result:
Party ID
Biden: 46.6%
Trump: 45.0%
Ideology
Biden: 46.5%
Trump: 45.4%
Even the 2012 electorate ended up only being a 3.3% Biden lead. If the subsamples are correct, this is much closer than Biden +6. The good news for Biden is that undecided and supposed third-party voters skew towards Demographics favorable to him.
Unskewing polls by party ID and/or ideology is a terrible idea (see: 2012). These are self-reported identifications that can change for an individual voter, unlike invariant characteristics such as race, gender, and birth year. Weighting by the latter kind of demographics is appropriate and necessary, but guess what: the pollster has already done that for this poll. From the methodology statement:
This sample was weighted according to gender, age, race and education based on the American Community Survey, conducted by the U.S. Bureau of the Census, as well as 2016 presidential vote and registration status.
I'm wondering if our Atlas ElectionsGuy was the creator of the infamous 2016 LongRoom.com website, that was "unskewing" and “removing the bias in the polls."