Read on Twitter
I am going to use Twitter as a diary of sorts to explain what happened with our presidential election forecast (and why) on most days between now and November 3rd.
A thread: https://projects.economist.com/us-2020-forecast/president
A thread: https://projects.economist.com/us-2020-forecast/president
June 11: The chance of a Biden victory increased from 83 to 85%. Most of this was caused by a Biden+13 & Biden+8 poll nationally (both weighted by past vote), and a D+3 poll in FL.
Trump’s approval ratings also decreased today, dragging down our prior. https://projects.economist.com/us-2020-forecast/president
Trump’s approval ratings also decreased today, dragging down our prior. https://projects.economist.com/us-2020-forecast/president
June 12: Not much movement in the forecast. The topline probability is the same. Biden’s polling margin shifted up 0.1 in FL, AZ, and GA.
The polling average in OH went from blue to red, but the e-day forecast is still red. Worst numbers for Trump yet. https://projects.economist.com/us-2020-forecast/president
The polling average in OH went from blue to red, but the e-day forecast is still red. Worst numbers for Trump yet. https://projects.economist.com/us-2020-forecast/president
June 15: Multiple polls showing Biden up 9 nationally and one showing him down 1 in Iowa nudged the forecast a bit his way. But overall, things haven’t changed that much over the last couple of weeks. State polls have been showing Biden up big for months. https://projects.economist.com/us-2020-forecast/president
June 16: Lots of polls with Biden up nearly 10 points in swing states today. Combined with yet another dip in Trump's approval ratings, the forecast moved ever so slightly toward the Democrat today. A new Civiqs poll pushed AZ to the left of NC (where it belongs, IMO).
June 17: A batch of state-level Change Research polls may have thrown the model for a bone today were it not for several national polls that were clearly good for Biden. The fallout from the unrest continues to push down Trump's chances of re-election. https://projects.economist.com/us-2020-forecast/president
June 18: Trump is in a preeetty bad position if we’re arguing over which of the polls showing him down 8, 9, or 12 points is right. FWIW state polls have show a slightly closer race that has hovered around Biden +8 for a while now. Keep an eye on the gap. https://projects.economist.com/us-2020-forecast/president
June 19: A Trump +3 poll in NC did little to shake things up.
Biden's current polling margin in...
SC: -8.1
AK: -5.6
TX: -3.3
IA: -2.6
GA: -0.5
OH: 0.4
NC: 1.1
AZ: 2.9
FL: 4.6
PA: 4.9
WI: 5.4
NH: 5.7
NV: 6.9
MI: 7.0
MN: 8.3
National: 8.6 https://projects.economist.com/us-2020-forecast/president
Biden's current polling margin in...
SC: -8.1
AK: -5.6
TX: -3.3
IA: -2.6
GA: -0.5
OH: 0.4
NC: 1.1
AZ: 2.9
FL: 4.6
PA: 4.9
WI: 5.4
NH: 5.7
NV: 6.9
MI: 7.0
MN: 8.3
National: 8.6 https://projects.economist.com/us-2020-forecast/president
June 22: We're flying blind over the last week w/o new national polls, so new data in Midwestern states will exert a larger-than-typical sway over the forecast. A Biden +1 poll in MI (probably too low) has pushed Biden down by <1% in WI, OH, PA, for ex. https://projects.economist.com/us-2020-forecast/president
June 23: No huge changes today. We got two new TX & USA PPP polls that showed a similar race to their last polls.
A friend said the other day that it would be fun if I could be more “emo” w these entries, but sometimes there’s just nothing to report! https://projects.economist.com/us-2020-forecast/president
A friend said the other day that it would be fun if I could be more “emo” w these entries, but sometimes there’s just nothing to report! https://projects.economist.com/us-2020-forecast/president
June 24: The new Biden +14 NYT poll is definitely on the higher end of what we've seen, didn't move the model much. Polling in Wisconsin pushed Biden up a bit there, but the average in the state is still around +6-7 (consistent with 8-9 nationally). https://projects.economist.com/us-2020-forecast/president
June 25: Today was probably one of Trump’s worst days in the polls. We got a load of surveys that showed him down ~7-8 points in the swing states & 8-9 nationally. It’s going to take a lot for him to win in Nov. That’s not impossible but it is improbable. https://projects.economist.com/us-2020-forecast/president
June 26: It’s starting to look like the Midwest has shifted left relative to the nation as a whole, with changes in polling numbers outpacing national shifts in states such as MI and WI. That’s a partial explanation for Biden’s improved odds this week. https://projects.economist.com/us-2020-forecast/president
June 29: Some new polls have pushed Trump underwater in GA & OH. I don't think he'll win them, but this is yet another sign of the president's steep uphill climb to re-election. He could still pull off an upset... but his odds are 2-3x worse than in 2016. https://projects.economist.com/us-2020-forecast/president/georgia
(To clarify: I meant I don’t think BIDEN will win GA and OH — our model currently thinks the race will tighten before November and Trump will regain his footing there. But it could go either way! Both are tossups! A Trump win is just the most likely outcome, not the only one...)
June 30: Some movement to Biden from new Pew and Suffolk polls today. Updated Biden margin in our polling averages:
MO: -8
SC: -7
AK: -5
TX: -3
IA: -2
GA: 0
OH: 1
NC: 2
AZ: 3
FL: 5
PA: 6
WI: 6
NH: 6
NV: 7
MI: 8
MN: 10
ME: 12
National: +9 https://projects.economist.com/us-2020-forecast/president
MO: -8
SC: -7
AK: -5
TX: -3
IA: -2
GA: 0
OH: 1
NC: 2
AZ: 3
FL: 5
PA: 6
WI: 6
NH: 6
NV: 7
MI: 8
MN: 10
ME: 12
National: +9 https://projects.economist.com/us-2020-forecast/president
July 1: At this point it's getting hard to imagine Trump winning—IF our data is right. But even though our model simulates a 12pt interval around today's polls (in margin), it's still at 91% Biden today & entering pretty uncomfortably confident territory. https://projects.economist.com/us-2020-forecast/president
I think it's a good model so I'm standing by it... but 91% is pretty high...
July 2: Today’s promising job numbers did little to change our election-day forecast. That’s bc we’re so far beyond “normal” territory for yearly economic change that the model still just sees a dismal outlook for the incumbent. Biden at +8 for e-day vote. https://projects.economist.com/us-2020-forecast/president
July 3: It’s starting to look like Biden may have plateaued in our forecast. National polls are hanging around +9 & the fundamentals forecast is unchanged. That’s not necessarily bad; every day that goes by without a dip increases his chance of winning. https://projects.economist.com/us-2020-forecast/president
July 6: Rumor has it the Trump campaign is going to target New Mexico in order to offset their apparent losses in the Midwest and Florida. Well, NM slipped into the 99%+ “Safe Democratic” category in our forecast today. So maybe find a new way to win.... https://projects.economist.com/us-2020-forecast/president
July 7: At this point in the 2016 campaign, Hillary Clinton was at 52% and on a downward trajectory in our model. Contrast that with Joe Biden holding steady around 54-55%, and with a more robust position across the swing states. https://projects.economist.com/us-2020-forecast/president
July 8: Here are the states our model says will tip the election to either Trump or Biden, and how often that happens:
PA: 20%
FL: 17%
MI: 13%
WI: 10%
NC: 6%
MN: 5%
NH: 4%
AZ: 4%
GA: 4%
TX: 3%
(This is why Trump spending money in Ohio is silly.) https://projects.economist.com/us-2020-forecast/president
PA: 20%
FL: 17%
MI: 13%
WI: 10%
NC: 6%
MN: 5%
NH: 4%
AZ: 4%
GA: 4%
TX: 3%
(This is why Trump spending money in Ohio is silly.) https://projects.economist.com/us-2020-forecast/president
July 9: As of today, the model started incorporating national polls in its "prior" prediction of the popular vote (vs just using "fundamentals"). This will push up Biden's margin a bit in the coming weeks, as polls put him at +10 while fundamental are +7. https://projects.economist.com/us-2020-forecast/president
July 10: It's tempting to doubt the polls, but let me put Biden's lead this way: even if polls were twice as wrong as they were in 2016 (errors that large are very unlikely, to be sure), he would still probably win the election. https://projects.economist.com/us-2020-forecast/president
July 13: Our forecast model for the presidential election suggests that Joe Biden needs to win the popular vote by 2.8 percentage points to win the majority of electoral college votes, a very slight dip in the chamber’s pro-Republican bias since 2016. https://projects.economist.com/us-2020-forecast/president
July 14: Today was the first time our model was above 90% for Biden all day. In recent weeks he has really been propelled forward by great (not good, great -- talking implied shifts of 10+ pts nationally) numbers in red states https://projects.economist.com/us-2020-forecast/president
July 15: Need I say more than the linked tweet? One thing our model does (that the site doesn't show you) is calculate how far left/right of the national margin each state leans. Trump has been imploding in MI, PA & FL recently.
https://projects.economist.com/us-2020-forecast/president https://twitter.com/gelliottmorris/status/1283507997109518341?s=20
https://projects.economist.com/us-2020-forecast/president https://twitter.com/gelliottmorris/status/1283507997109518341?s=20
July 16: Our model is rapidly approaching uncharted waters. At no point in other 21st-century elections did a candidate lead by so large a national margin 110 days out than Biden does today (+10 points). His state-level position is equally unprecedented.
https://projects.economist.com/us-2020-forecast/president
https://projects.economist.com/us-2020-forecast/president
July 17: A new UCLA Nationscape + Democracy Fund poll showed Biden up 9 today. That’s right on the model’s current estimate. Other than that, nothing new to report. Maybe keep an eye on TX? If Biden gets any more favorable polls there it’ll move to tossup. https://projects.economist.com/us-2020-forecast/president
July 20: Every time I say this Biden gets an additional set of promising polls, but nevertheless: Trump's slump has stalled recently. He may be bottoming out. Our forecast appears resistant to hitting Biden +10, probably because of the e-day prior at +7-8. https://projects.economist.com/us-2020-forecast/president
Full disclosure: We are exploring options to increase the robustness of the model when it is forecasting at the tails. Given the small n of the training data we don't really know if the elec is 93 or 95 or 91% Biden, and the forecast should reflect that uncertainty. 1/5
2/5 The way forward is to increase the chance of extreme outcomes by basically hedging toward 50/50 in the cases that they "break" the model. There are a lot of events that are outside the training data. The key is to calibrate the extra uncertainty and we think we are close.
3/5 Let me be clear that we DON'T think that any of the theoretical components of the model are "wrong," it just looks like the model isn't exploring the full range of outcomes when we get to 95-99%. A bit more variance in the polls and more error in the fundamentals will help.
4/5 Anyway, this should have a relatively small impact on the model. We are talking about aggregate probabilistic changes around 5 points at max. We'll disclose everything more when we have an answer. Model-building is hard and just using fat tails isn't a magic fix for error.
5/5 Thanks always to people who have engaged with the model thoughtfully and critically. This is why we put the code up online!
(that code lives here https://github.com/TheEconomist/us-potus-model)
July 21: The average “error” in polls at this point in elections since 1980 is about 8 points. Polls in recent elections have been about twice as accurate. Volatility in 2020 has been closer to the latter, but that’s not a guarantee of a Biden win. https://projects.economist.com/us-2020-forecast/president
July 22: Our model took in that new Quinnipiac poll in TX but it didn't really change its mind about anything. That's partly bc Q polls have tended to be bluer than the average so far (our model adjusts for this) & bc of some slight tightening nationally.
https://projects.economist.com/us-2020-forecast/president/texas
https://projects.economist.com/us-2020-forecast/president/texas
