J.L. PARTNERS’ US Presidential Model

16th September 2024

50% Chance of A Harris Win

Methodology

Our model of the Presidential election is a two stage, Bayesian process that begins by examining historical data in both the long term and short term

  • We analyze elections from the past 80 years and use economic data, approval ratings as well as incumbency effects and term effects – this forms the basis of our Long-Term Fundamentals model

  • We then analyze the election since 1996 using similar sorts of economic data, but also include information on Consumer Confidence and Congressional Approval – this forms the basis of the Short-Term Fundamentals model

  • These two models are then mixed together to yield our Hybrid-Fundamentals model that gives us our initial Presidential Prior – the vote share each candidate would be predicted to win if we had no polling data.

Then we gather polls from the state and national level and filter out any data that only polled questions involving direct head-to-heads – this helps account for some of the third-party effects in the model

  • Taking our prior for the Presidential vote share we produce priors for the states using historical data and then run an autoregressive Bayesian algorithm to fit a function to each of the states and the entire country

  • Some states aren’t polled as frequently as others, so we calculate how similar states are based off their 2020 vote, demographics, house prices, mortgage payments as well as other factors. This allows us to use data from one state to infer how other states will perform.

  • For example, Pennsylvania and Illinois are quite similar based on our measures so the more frequently polled Pennsylvania can be used to inform the state level polling of Illinois. 

  • These state level estimates are then combined to give an initial country-wide vote share

  • Note that we include adjustments to account for who is on the ballot in each state in order to properly account for voter flows. As such, we only use polls that have some form of full ballot testing.

We then add national polling data which mixes with this state level data and updates it. After this we produce the final estimates and probabilities