Thread by @gelliottmorris

Here's what our election model would have predicted for final outcomes on this day in...

2008: Obama +6 in the natl pop vote, median EV outcome of 319, 78% to win
2012: Obama +5, 323 EVs, 89%
2016: Clinton +0, 271 EVs, 50%

2020: Biden +9, 349 EVs, 91%

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

Note that our confidence in Obama 2012 was a lot higher than others' at this point back then, and that in 2016 we were lower than the consensus. That's bc we put a lot of weight on econ & political "fundamentals" which we find are less noisy than polls at this pt in the campaign.

Our model also thought (beforehand) that 2012 would be easier to predict than other elections. That's bc the relationship between POTUS approval and POTUS vote share is a *LOT* higher when an incumbent is running for re-election, and also bc of good-not-great econ growth in 2012.

This also helps to explain why our model is more bullish on Biden than most other analysts seem comfortable with. Instead of assuming the state of the race is a random walk to election day, our model drifts toward what we expect based on the economy and POTUS approval ratings.

And what we expect based on those indicators for 2020 is a pretty broad victory for the Democratic candidate. Economic stagnation + unpopularity + polarization is not Trump's friend. (And yes, remember that our model increases uncertainty when the economy is volatile.)

In other words, our model frames elecs a bit differently. Others say "Oh, it's early, anything could happen." We say, "Oh, it's early, lots of things could happen BUT we expect from underlying data that X is still more likely than Y, even acknowledging variance in the polls."

Anyway, we explain this in our methodology so I'll let you read more about it there. The important things are:

(a) Biden's lead in the polls right now is larger than any candidate's since Clinton 1996
(b) The fundamentals suggest some tightening, but still a rather large victory

Oh, a final thought. The revised version of our model shows some relatively large differences than the old version, particularly in 2016 when our correction for biases from weighting and nonresponse give about a pt back to Trump nationally. So the 2020 model will prob change too.