Raphael Nishimura

Congrats!!

@badsurveyq.bsky.social

CNBC and Kalshi Strike Exclusive Partnership

I remember that when I would urge media to talk more about winning probabilities derived by forecasters/aggregators in their polling coverage, people would tell me the general public would not understand it.
I guess the media decided that this is not a problem anymore?
www.cnbc.com/2025/12/04/c...

What would once be the dream of any subjectivity Bayesian -- the popularization of using wagering odds as probabilities, is now becoming a real dystopia... 😫

A IA está causando um enorme estrago. Eu vejo um post desses hoje em dia e logo penso: "Isso só pode ser IA, não é possível que isso seja real" 😅

Come join us at the Strength in Numbers Discord, Marcelo!

yeah yeah, but how cool is going to be seeing this in live action 😛

Haha I was wondering about that after reading your new handle.
I wouldn't call myself a hater, I just don't think they use sound methodology and they are super sketchy about that, while at the same time being very vocal about being the best it ever existed :P

The world is finally healing! 🙌

Lol yeah
Intuition might be deceiving me, but I think for this sort of thing explain this difference, there should be a substantial amount of clustering happening.
In another matter, will I see you at MAPOR next week?

"Damnit Georg, at some point you have to start asking yourself why all these Republicans love you so much!"

This is some sort of measurement error, just don't ask me which 😅
www.nbcnews.com/news/amp/rcn...

This is an incredible opportunity if you want to work in polling and surveys with some of the best in the field! 😊

For reference, it took Selzer 12 days after the election last year to produce and release a full post-mortem of her infamous poll.
Given their background on methodological transparency, I'd be very surprise if we see anything remotely similar from this so-called "America's most accurate pollster".

Did you say "representative sample"?

A year ago today, AtlasIntel said they would deposit at the Roper Center the raw data from their final polls. It's still not there.
Something tells me they will not be depositing that data file, neither the ones from this week's polls...

That's similar to the shy Trump voter hypothesis. However, we never found strong evidences to support it, either in 2016, 2020 or 2024.

I was just saying this in the other site: Mamdani's share should probably have a heavier weight in the post-election polling analysis than Cuomo's share, or at least it should be put together with Sliwa's, given the likely late shift.
And looks like somebody already did it.

Believe or not, the New Jersey poll wasn't the lowest point in US elections for the (former?) "American most accurate pollster", AtlasIntel
Take a look at their final poll in the PNP Primaries in Puerto Rico last year:

Poll: Pierluisi +14
Result: González +9

That's a 23-points difference!

Yup, same here!

What an exciting future! 😭

Yeah, I'm not at all surprised. They always pull the same kind of BS in every other country they have a bad performance. Either that, or they just never talk about it, like in the primary of the gubernatorial election in Puerto Rico or the presidential election in Mexico last year.

l guess NJ is a small demographic subgroup for AtlasIntel standards?

Mood da noite

Also, can we all put to bed now the myth of "the most accurate pollster in the US"?
I know I've been saying this since after the 2020 election, but I think we can finally agree on that, right?! RIGHT?!?

Wow, turns out weighting the hell out of your sample recruited through river sampling out of ads in social media doesn't work after all, who knew?! 🤭

I've been calling AtlasIntel's poor (to say the least) methodology out since after the 2020 elections, when it was not a popular thing to say. I'm just glad you all caught up and we are all on the same page now!

Based on their performance in São Paulo mayor election last year, they will come up with the lamest excuse ever and then never mention it again lol

ANES Data Release! electionstudies.org/data-center/...

The 3-wave ANES panel is now available. It merges data from 3 election studies (2016-2020-2024), the first time the ANES has collected interviews of the same respondents across 3 presidential elections.

How something like ~ 1.96*se(estimate) is incorporating bias from sampling and non-sampling error?