Model update -- @bot_2024 V3.0 is out! @notkavi and I have completely rebuilt it to use more generic data across elections and better model partisanship and turnout on a demographic level, which should give us more interesting and realistic coalitions for 2024. Details in thread! 
This model projects 2024 demographics from the available 2012, 2016, and 2020 demographic data. It then takes in county-level demographic data for 2012, 2016, and 2020 and uses it to construct underlying demographic features for the electorate.
We use pretty interesting demographic features commonly available across the 2012/2016/2020 electorates, such as CVAP with racial breakdown, poverty, income, religion, and education. That also helps us get more information about what's happening with the underlying electorate.
The model uses a (simple!) neural network to carve the electorate up into a latent "demographic space" on a per-county basis before predicting the turnout and partisanship per demographic and uses that to compute county partisanship.
The model is trained on the swings in each county from 2008->2012, 2012->2016, and 2016->2020. We perturb the model's predictions to get different maps for 2024, showcasing a variety of unique and plausible coalitions.
It is important to note that in predicting 2024, we center the model around the partisanship per "hidden" demographic, but we use a weighted average of the 2012/2016/2020 turnouts, because we do not know what turnout will look like and have no reason to prefer any scenario.
Here's how the model predicted 2012 based on underlying demographic data. In the residuals, you'll see Obama's incredibly significant overperformances in Ohio and Iowa -- this is the type of stuff legends are made of.
This is how it thinks 2016 should have gone. The midwest isn't all that surprising, actually, in hindsight. We just lost it badly in Michigan, where Democrats didn't turn out at all. I'd ignore Miami-Dade listed as a Trump overperformance -- we don't have a "Cubans" variable
Lastly, here is 2020, with Biden's overperformances in New England, upper Michigan, and several sunbelt suburbs being the most significant things for Democrats. Trump, meanwhile, smashed through expectations in Ohio, Iowa, and Florida.
We also added a legend and state lines, per request. Anyways, hope you like this! @notkavi has restructured the codebase three times by now in the course of writing this, and I am a bit sick of looking at census data and writing scripts to aggregate and parse it by this point.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
