Post-Mortem On My 2020 Presidential "Now-Cast"
Looking Back On How I Did Attempting To Predict The 2020 U.S Presidential Election
From 100 days left until election day to literally an hour before polls closed, I published and updated a “Now-Cast” attempting to predict the 2020 U.S Presidential election. Now that the electoral college has officially cast their votes and every single state’s results have been certified, I can finally take a full in-depth look at how my attempts to predict the election went, and make light commentary on why I believe the result ended up being what it was.
First, how I get my predicted results - I use a simple aggregate of survey data. Like with the UK general election forecasting I try, I take a chance on actually trusting the public polling and public MRP estimates, and average the last three dates of final polling field dates or MRP estimates to each survey available for each race. Whenever data is missing, I use historical results to capture the likely trajectory of a race - which comes in handy with heavily partisan and less polled states. I then count up how many electoral votes a candidate is ahead in the aggregate with and assume the potential for a five point error in my predicted margins either way. Below is the final national polling aggregate I got, this is exactly how I also predict the margins on a statewide (or district for DC, ME, and NE) level.
Nationally, my average predicted a rounded margin for Biden of 8.24. The actual result ended up being a 4.45 national margin of victory for Biden, which is actually a result the Now-Cast thought was possible as a worst case for Biden would have been an even closer 3.24 popular vote victory for him. That said while the Now-Cast thought the actual national margin could happen as it did, that’s still a 3.79 miss in favor of the incumbent President which continues the streak of incumbent Presidents overperforming their national polling in 2004 (Bush +1 poll average to Bush +3 actual), 2012 (Obama +1 poll average to Obama +4 actual), and now 2020.
The strength to using this method for the forecasting is that it catches late swings as the election comes to an end, catching potentially game changing momentum late that other more complex methods would take time to catch themselves. This method did wonders for me in 2016 when it caught a surging Trump as Clinton using this method was predicted to only barely be above the 270 mark and a simple poll average away from losing. It also came in handy this year as Biden slipped into losing territory in his worst case scenario but was still favored better than Clinton was. Perhaps signs of a potential late swing of “stick with the devil you know” voters that helped Trump stay competitive in defeat?
The weakness however is that such a method leaves this sort of forecasting susceptible to noise in the middle of campaigns, and doesn't weight the quality of data so that it can balance out potentially bad polling from actual good polling. The point of such a method though is to catch the likely outcome if the election "were held today" more so than a more complex formula that attempts to anticipate future momentum swings in the race. That’s why I call it a "Now-Cast", or snapshot of the race.
Anyhow below is my FINAL prediction for each state using this method, which each state (or district for DC, ME, and NE) shaded by predicted margins of +<5, 5+, 10+, and 15+. Any state where the margin was under 5 was seen as a state by the Now-Cast where an upset could occur. Based on this the Now-Cast predicted the expected result if every state went as the average showed would be a Biden electoral college win of 351 to Trump’s 187. However, a worst case scenario for Biden would have seen him losing to Trump with 253 electoral votes to Trump’s 285. His best case scenario would have been a 413 electoral college landslide versus Trump’s 125. Keep in mind for a best case scenario to occur, literally every state within 5 points would have to have gone one way and even in 2016 that didn’t occur so Trump overperforming his polling like in 2016 but still losing was something the Now-Cast predicted. For Trump to win, the Now-Cast says he would have had to pull off an even bigger upset than he pulled off in 2016 - a very hard task to complete even in a day and age where polling has had trouble catching low propensity rural Trump voters and an increasingly complex Hispanic vote.
So how did the projections above do versus the actual state by state results? Well there’s two ways one can look at it. First let’s look at the number of states (and districts) called right. The final prediction was that Biden would be expected to win 27 states + DC, every ME district, and NE-2 to Trump winning 23 states. The final actual result was 25 states for each man with Biden also winning DC and NE-2 and Trump once again picking off ME-2 even while losing the state. That’s two states and a congressional district that Trump pulled off an upset in and according to the Now-Cast all three were within the upset range with Biden having just a 2.47 lead in FL, just a 1.09 lead in NC, and a weak 1.67 lead in ME-2. If you only test the forecast on this measurement, it didn’t do half bad - getting 48 out of the 50 states right overall with the predicted overall winner actually winning. Below are the actual results and the margins in each state based on intervals of +<5, 5+, 10+, and 15+.
However there’s also the matter of which was the error went in each state, the Now-Cast thought Trump getting 232 electoral votes was a possibility and obviously it underestimated him in the final expected margins, but how broad was Trump’s overperformance? As you can see from the map down below which shows where each state’s miss erred towards, Trump overperformed the expectations of the Now-Cast in all but 4 states (CO, LA, MD, and VT) and three districts (DC, ME-1, and NE-2). This includes some whopping double digit misses in favor of Trump in 4 states (MT, TN, UT, and WV) compared to Biden’s biggest error in his direction being him overperforming in NE-2 by 2.20. So basically while the Now-Cast can celebrate getting 48 out of 50 states right, Trump significantly overperformed across almost every state with polling, MRP estimates, and even some uses of past electoral data overestimating Biden’s actual winning margin even in the bluest of states.
The Now-Cast assumes errors can happen of course, and I built it so that it assumes it could be off on the final margins by 5 points either way. Therefore it considers any result within 5 points of what it expected to be acceptable and predictable, but anything after that means a considerable forecasting screw-up in that state. The map below shows where states were within the 5 point margins of the final predicted result either way as green, and states that were above that 5 point threshold as red. On a statewide level, 33 states were within the final margin the Now-Cast thought should happen while 17 were outside of that range. While a clear majority of states didn’t see that dramatic a miss, 17 states is still a little more than I’d like and some key states like IA, FL, OH, and WI saw bigger errors than the Now-Cast anticipated could happen. So all in all, I’d say the Now-cast did its job aggregating the data to find plausible paths to victory and getting a good picture of the potential final map, but it did so while dealing with plenty of faulty surveys as Trump over-performed in 46 states and 17 states were outside of the expected possible margins.
Now for some light commentary on why the result we got was well, what we got. As the map below shows, even with Trump overperforming the Now-Cast’s expectations in 46 states, 43 states shifted away from him versus his 2016 margins with Biden outperforming Clinton in those states from just 0.10 (in OH) to an impressive 8.59 (in CO). In comparison among the 7 states Trump had better margins in than 2016 they range from just 0.03 (in NV) to his biggest swing at just 2.72 in HI. Again this is with Trump over-performing in all but 4 states! In other words if you want to look at the results at their face, the reason the sitting President lost was because well…he lost support across almost every state in the union - even in some of the reddest states. And while President Obama saw a near similar amount of states swing away from him in 2012, he had much more insurance from a big win in 2008 than Trump did after a very close win in 2016. The President had little room for error and well he had more than that occur to his coalition as he gained among minority groups but bled suburban white voters.
As to why this occurred, we have various possible answers. The President had the incumbency advantage and good approval ratings on the Economy on his side but he was more unpopular than Biden, had horrible approval ratings on practically every issue outside of the Economy, and the political environment of recession, social upheaval, and pandemic was the perfect recipe for what turned out to be a change election that was only a little bit of the choice election the President tried to turn it into. Furthermore in a high turnout election where BOTH sides show up, independent voters are key and after backing him in 2016 non-partisan voters turned on the President and went with Biden by wide margins based on the preliminary data. But while we can cut and dice those factors (and will in the future when I look back at historical results), it all comes down to the President’s approval rating.
As you’ll see from the table below, a President’s approval rating on election day when seeking re-election is a strong indicator of their fate - especially in more modern and polarizing times. Presidents who are at 50%+ approval ratings on election day win, and those below it (As Trump was) lose. Yes they tend to do a little better in more recent elections than their approval ratings and the President turned out his base just as previous winning incumbents did, but hiss 45% approval rating is the very number that Clinton and Obama had during really rough midterms for them. The President did see his approval rating in a better spot than it was in 2018 when it was then around 42% on average but 45% is a hard approval rating for an incumbent President to survive even with the electoral college advantage Trump had. Furthermore his disapproval rating was averaging at 52% and 53% of the vote didn’t back him, and to add to even that Biden’s favorability was in the low single digits while the President was in the negative range by double digits. The reason Joe Biden defeated the sitting President of the United States? Because that sitting President had not gained the support of a majority of the people he was asking for another 4 years from, either with his job performance or on a personal character level.
In a way the Now-Cast’s prediction of a Biden victory, even with Trump overperforming shows that a Trump loss was expected, maybe I’d even dare call it the boring and predictable outcome. I’m baffled that some of the President’s supporters are having a hard time piecing together why he lost this race. If anything they should at least be happy he stayed competitive and his party limited the damage down-ballot by making gains themselves. Because when you take away all the forecasting and simply look at the mood of the country, Trump’s 2020 loss was probably the most normal thing to happen during his entire era of dominating U.S politics. Congratulations to President-elect Biden and to my 2020 U.S Presidential Now-Cast.