Thread by @gelliottmorris

🥁
🥁
🥁
🥁
🥁

The Economist's forecasting model for the US presidential election is now live! We think Joe Biden has around a 5-in-6 shot at winning the presidency.

projects.economist.com/us-2020-foreca…

Three links to be aware of:

1. Here is a piece explaining what the model says today: economist.com/united-states/…

2. Here is the methodology: projects.economist.com/us-2020-foreca…

3. We have open-sourced the code for the model. You can download it here: github.com/TheEconomist/u…

Here's the electoral map as we start the year in earnest. The model thinks Joe Biden is favored to win about 331 electoral votes come November.

There is a lot that makes our model special, but let me highlight a few points for you here:

First, the forecasts uses presidential approval ratings and an index of 8 economic indicators to create an expected range of outcomes for November. This "prior" model tells us how well/poorly Trump should do based on the how his predecessors with similar numbers have performed.

We call this our "prior" forecast for the race and it helps us calibrate the model early. If polls are diverging wildly from the prior, it will split the difference. Today the prior is that Joe Biden will win 53.5% of the two-party vote. By election day it will be mostly polls.

Second, our model treats polls conservatively. Election-day polls (at the state level) have had a margin of error of about 5 points in elections going back to 2000. That's a bit wider than many forecasters assumed in 2016, possibly explaining why they got the election so "wrong."

Every time our model runs it "simulates" 20k different elections where polls are wrong 0 to 5 percentage points in the favor of either candidate. A technical note: these errors are correlated across states & drawn from a t-distribution w/ fatter tails than the normal bell curve.

Third, and in a first among forecasters (from what we can tell), our model includes a correction for partisan non-response bias in polls. In effect it draws two sets of averages each day: one for polls that don't weight by party/past vote, and one that does.

Then it pushes the numbers from unadjusted polls back towards the trend lines in polls that are weighted for partisanship. This doesn't make so much a difference now, but come the conventions and debates our model will probably be much more stable than others.

A technical note: the trend line we're drawing here is for data that has already been adjusted for pollster, mode and population effects. So we're not conflating the fact that most "adjusted" data comes from online pollsters.

The adjustment for partisan non-response bias also provides a neat little trick for adjusting other polls for not weighting by education; all the pollsters in our "unadjusted" set do and many of the ones that aren't don't. So the differential in trend lines will pick that up too.

We are aware that this adjustment might be controversial. If you don't like it, change the code and run the model for yourself! github.com/TheEconomist/u…

Some other notes:

- Our "prior" model inflates the amount of error in its predictions during uncertain economic times (like these)
- We include most pollsters' data in our forecast, but not from those who have a history of making things up or who field their surveys on MTurk.

I'm sure there's other stuff I'm missing.

Although I think of it as my baby, the model is was very much a group effort. @StatModeling, @MHeidemanns and @DanRosenheck wrote/reviewed much of the code for the model & @martgnz and @futuraprime did the site.

A final note (for now): there is a long time between now and the election. You can see that early in the year our model would have given the candidates a much closer to a 50-50 shot. When the facts change, our model will change its mind too.

Oh! Another thing:

It's clear that political polarization has fundamentally changed US politics. Above all, partisans don't react to good/bad economic news like they used to, dampening the impacts that growth should have on our prior forecast. The model takes that into account.