the reaction to biden's cognitive decline (which i don't think is substantively worse than 2020. that is, what i'm saying is that you should have realized he was out of it four years ago, and i wonder what effect consistent republican messaging on the point has actually made. it's subconscious. it gets to you. it would be repeat the lie, if it wasn't actually true, but that latter point is merely a happy coincidence; they'd lie to you if it wasn't. it's just easier to repeat the truth, i guess, but it's the same propaganda model, and i suspect it actually worked.) even indicates as much; people are trying to push biden out because they're afraid trump will win. this isn't some shift towards trump, it's a reaction against him.
in context, however, the culmination of recent events has shifted the race, and it follows that 538's model will shift to adjust...in two months. this is a criticism of 538's model (and of nik nanos' model in canada) that i've been making for years, that it is an inherently conservative model designed to measure brand preferences for soda or cereal or some other purchasable commodity that is being misapplied to politics, which are not conservative, but incredibly dynamic. it works on some level for party loyalty in terms of buying memberships or making donations, because that is kind of like buying soda or cereal. sort of. the core voters can be tracked into terms of consumer preferences, but core voters don't swing elections, independents, new voters (often low information) and young people do. however, when people shift political allegiances, as is currently happening, it tends to happen immediately and all at once. not only is 538's model going to have difficulty picking up sudden shifts in voting preferences in michigan or pennsylvania (or new york or oregon), but it's actually designed to smooth them out, meaning you have to wait for several months after the shift has occurred before the conservative models reflect it.
despite the introduction of advanced voting, elections are still snapshots, they aren't averages. you want to look at snapshot polling to make election predictions, not smoothed out conservative averages. my insistence on doing this is against conventional contemporary wisdom, which is in truth not so wise (they actually should and do know better) but is the reason i consistently beat the models when i sit down and do the math carefully. 538's model consequently isn't wrong, exactly, it just has a tremendous time lag built into it that makes it essentially useless for this application, which it wasn't designed for, and that is 538 and nik's fault: they should know not to use this model for this application.
what the model is useful for is tracking political donations, so the parties like it, but you can't use it to predict elections where voting preference is dynamic, and 538 itself should clarify it, because they know it.
now, the flip side of the argument is that you need to ponder this question: will the debate, and the news conference, and trump getting shot at, and him picking a log cabin republican for vp, and the internal revolt against biden, be things that have long term effects on voting preference, or will it actually smooth out? is 538 arguing it will smooth out? i don't actually think they are, but if it does then their model will make an accurate projection, and what you're seeing right now will be a blip that comes out in the wash.
i don't think that's the case. i think these are real shifts that will last at least a year or two, that biden is permanently done for. 538's model is designed to not pick that up until october, by which point it might have actually flipped back, if it were some other scenario. i wouldn't project that.
in some elections, 538's model will work because you don't have dramatic shifts. in this election, it won't work, because you're going to see massive fluctuations, as everybody realizes these are shit options across the board and nobody knows what to do. 538 should acknowledge that his model can't project an accurate result in the presence of massive fluctuations because it is actually designed to negate them and explain that to people so they get it.
however, the 538 site, which i haven't been to in a while, is useful because it shows it's work. ignore the model, and look at the most recent polling. don't take the grading system overly seriously; try to pick out the surveys that use phones (and therefore have random sampling) while mostly ignoring the online polls. those snapshots will tell you what people want today, not what people will pick in november, but if you follow the snapshot polls, you will get better outcomes than the smoothed out averages, unless the situation stabilizes, which i doubt will happen.