The same poll has Biden up only +1 when voters are able to pick other candidates (Kennedy, Stein and West). Now, we don't know how many will actually pick those options. It's probably a low, but still non-zero, number — so you do want to account for this pattern. & it hurts Biden

Empirically speaking, the best thing we can do for forecasting is to root our expectations in historical dynamics. Change in poll margin from May to Nov is 7 points on average since 1948. And even in times of lower volatility (post-2008), four or five points on margin is common.

Here's the Biden v Trump national polling average since last July, using the current 538 methodology applied backwards in time. Not a lot of movement in the national polls! And always inside the model's MOE. Biden has improved since Dec/Jan, but he's been underwater since Oct.

- Biden was +2 on 7/1/2023
- He lost about 2 points from Aug to Oct as the GOP primary helped Trump
- Biden lost a further 2p by Jan. Israel-Gaza + campaign fx?
- Since March he has recovered to -1

Every day since 7/1/2023, Biden's margin has been inside our uncertainty interval

good work in this thread

If you’re nice about it, yes I will absolutely run analyses for you. That’s my job!

not sure how i'm going to be able to finish the forecast model now bloomberg.com/news/articles/…

historically speaking, polls at this point in the campaign do not make for great predictors of November vote outcomes. they are reasonably good at identifying close races and states that have moved more than others, but do not offer much precision beyond that

In terms of revealed issue importance, the anti-Israel protestors on college campuses do not represent the overall population of young voters. Polls show most young people care more about mainstream issues, such as inflation, gun violence and the costs of health care and housing.

i’m sorry to say this, but this is not how the statistics of polling works

Wait until Robert F Kennedy Jr finds out about the sample size of the 1936 Literary Digest poll

They're not really 'priors' it's just literally voter file data matched to polls and election results. I don't get what you think he's doing? Making stuff up?

good thread (not just cause it's nice to me, but that's definitely part of the appeal)

Important points here. Related: AFAIK our 538 polling aggregate is the only one that methodologically adheres to the Total Survey Error Framework — namely bcuz we explicitly encode non-sampling error that dynamically adjusts to the empirical noise in polls projects.fivethirtyeight.com/polls/presiden…

soon

programming challenge of the week... not as easy as it sounds to create an election model that works with a flexible number of electoral jurisdictions over time