markjrieke / 2024-potus

repo for constructing a 2024 presidential election forecast

2024-potus

repo for constructing a 2024 presidential election forecast

loose methodology

See associated notion timeline for more detail, but model is loosely the following:

$$ \begin{align*} \text{Y}_i &\sim \text{Binomial}(\text{K}i,\ \theta_i) \ \text{logit}(\theta) &= \beta_s + \beta{s,d} + \beta_p + \beta_m + \beta_g + \beta_c + \beta_n \ \end{align*} $$

$\beta_s$ and $\beta_{s,d}$ comprise the “true” latent voting intention in each state. $\beta_s$ is the time-invariant component, set by a Gaussian process over the euclidean distance between states in some normalized feature space and $\beta_{s,d}$ is a the daily time-varying offset from each state’s time-invariant component, set by a AR(1) / Gaussian random walk process over time in each state. The predicted voteshare in state $s$ on day $d=E$ (election day) is $\text{expit}(\beta_s + \beta_{s,E})$. The state-level prior is either over the time-invariant parameter or the predicted voteshare (TBD).

The remaining parameters account for bias in the individual polls:

• $\beta_p$: pollster
• $\beta_m$: poll mode (online, RDD, etc.)
• $\beta_g$: poll group (RV, LV, adults, etc.)
• $\beta_c$: candidate/party sponsor (D, R, or none)
• $\beta_n$: noise (per poll!)

Most of these have sufficient groups to be modeled hierarchically. I may model $\beta_g$ and $\beta_c$ with fixed effects, given the small number of groups in these parameters.

resources

• Models & methodology
  • Linzer 2013 paper
  • Pierre Kemp 2016 model
  • Economist 2020 model
  • Abramovitz time-for-change
  • FTE 2020
  • FTE 2016
  • DDHQ — ensemble of ridge, random forest, elastic net, and gradient boosts
  • Race2WH — normal approximation of candidate voteshare
  • JHKForecasts — simulation methods based on a normal approximation under the central limit theorem
  • Cory McCartran, Data for Progress — Bayesian model with a student-t response
  • Gelman/Microsoft
  • FTE — dig into once not on the MH network
  • NYT — dig into once not on the MH network
• Data
  • FRED
  • FTE Polls
  • Urban Stats
  • Cook
• Misc
  • FTE poll inclusion policy
  • FTE weighting methodology
  • Notes on copulas (used for generating distances in some feature space)

banned pollsters

• Center Street PAC
• Traflagar
• Rasmussen

loose workflow

• derived data (constant)
  • distance matrices
  • cpvi
• (approval model?) [may not actually do, we’ll see…]
  • approval data
  • e-day approval model
  • write results
  • write diagnostics
• prior model
  • economic data
  • approval data (or model)
  • fit
  • state-level priors
  • write results
  • write diagnostics
• poll model
  • polling data
  • prior data
  • fit
  • write results
  • write diagnostics
• reporting
  • update site
  • blastula email diagnostics

implementation notes

• See here for an example of y-axis text inlay using the ggh4x package
• See here for a basic overview of automating scripts on a schedule with github actions.
• See here for an example of installing a package from github as a part of a github action

misc notes

• Colors for display!
  • Safe D (>99): 3579AC
  • Very Likely D (99 >= x > 85): 7CB0D7
  • Likely D (85 >= x > 65): D3E5F2
  • Uncertain (65 >= x <= 65): F2F2F2
  • Likely R (65 < x <= 85): F2D5D5
  • Very Likely R (85 < x <= 99): D78080
  • Safe R (>99): B13737
• abramovitz data notes
  • Incumbent net approval pulled from FiveThirtyEight’s averages on the day before the presidential election. If the exact date is not available due to data resolution, (these are manually pulled) the net approval from the closest day prior to election day is used instead.
  • Third party flag is set to 1 whenever an individual third party candidate garners more than 5% of the national popular vote.
  • For Biden’s net approval, pulling the All Polls variant of FiveThirtyEight’s presidential approval tracker (this is consistent with what’s displayed for the previous presidents).