US Election 2024 - What do all the models say [LIVE]?

Continuously updated summary of the major election modellers at the national and state level

Every four years, we data nerds get a treat, while everyone else melts down. Yes, we’re only a couple of months from a US Presidential election on November 5th. And we want to know who’s most likely to win.

Lots of people want to tell you.

Oh, they’re all the same …

Everyone pretty much everyone agrees on what data to use.¹ National polls, approval ratings, state-level polling, economics fundamentals. All these need to be brought together through some model in order to create the best possible picture of a jigsaw puzzle on November 5th.

On that day, 50 states, each assigned a share of 538 electoral votes will vote on who they want to be President. There is a “winner takes all” system for each state (with some subtlety for Nebraska and Maine), and the national winner is the candidate who wins more than 270 EV. The models come up with some probability weighted model of what this process is likely to look like,² and then simulates thousands of possible “election days”, weighted by the probabilities given to each outcome, and their correlations to one another.³

So, what do they say this time around?

As far as I know there are six major polling aggregates running with live forecasts at the moment,⁴ plus a bonus source, with a very different approach.

• FiveThirtyEight: One of the oldest, and certainly the best-known. Now run by G. Elliott Morris after founder Nate Silver was forced out in an ABC/Disney takeover. Appears to be taking a fairly “activist” approach to selecting and adjusting polling inputs. They also appear to have made fairly major tweaks to their methodology in the interregnum between Biden and Harris. Explained here.

• The Economist: Applying an evolution of Drew Linzer’s original “Votamatic” methodology. In contrast to previous years (and to other projects) they have gone closed-source, though they still give a fairly detailed overview of their methodology.

• Princeton Electoral Consortium: Based on state-level polling. Notorious for their 2016 correlation-between-state-polls-error, which I think they have not fundamentally addressed. They are - however, fully open-sourced. They also do not give a overall probability - but it is obtainable by summing up some numbers buried in their Github repo.

• The Silver Bulletin: Nate Silver’s new outfit, running on Substack, operating what appears to be a very similar model to the old FiveThirtyEight approach - which uses relative weighting in averages as the fundamental mechanism to adjust polls and signals (i.e., he tends to shun the fancier Bayesian methodology)

• Data Diary: Designed by Mark Reike, a data developer, who has taken the 2020 Economist model and adapted it to to a fully Bayesian methodology (it was already close). It is fully open source - you can run you own copy on a laptop with R and Stan, and you can delve into the code to discover what it’s up to.

• Decision Desk HQ / The Hill: Their methodology statement name-checks every machine-learning approach out there (XGBoost, LASSO, elastic-net), but they have a good pedigree in calling elections as results are coming in (projecting from available information), so they are likely using the same techniques they have developed there.

• Betting Markets: Various betting markets: Betfair, Polymarket, etc run markets on the election. Their implied odds - implied from the value of thousands of bets, give an alternative way to “aggregating the information available”, and directly gives an implicit probability from “the crowd” that can be compared to the models above.

Enough talk, what are the results?

Two important points all these methodologies currently agree on:

1. Kamala Harris has improved a great deal on Joe Biden’s chances

2. Pennsylvania is pretty much the ballgame - whoever wins that state, is very likely to win overall (though several Georgia, Nevada, Arizona combinations can provide alternative, lower probability scenarios).

And an important point that they do not agree on.

1. Who is winning.

Here are the latest numbers: these charts will be updated daily or even more frequently - they are a fully-updated one-stop look at each aggregator/modeller’s probability of each main party candidate winning, alongside an average of what the betting markets say. (On mobile, please click through the charts to ensure that you are not viewing an old version).