For Trump, No Comfort in Forecasts or Twitter in Final Stretch of 2020 US Presidential Election
Trump’s down in the polls, and forecasts by experts point to a decisive loss on November 3. He doesn’t even seem to be doing as well on Twitter as he did in 2016. Is it game over for the White House incumbent?
With about two weeks to go before the 2020 US Presidential Election, the statistics on multiple fronts are looking rather grim for White House incumbent Donald Trump.
Daily forecasts from data analysis outfit FiveThirtyEight and weekly magazine The Economist point to a resounding defeat for him on November 3. Trump even appears to be underperforming on Twitter upon closer examination of the metrics.
Is it game over for Trump? As FiveThirtyEight’s editor-in-chief Nate Silver has pointed out on numerous occasions, having a low chance is not the same as having no chance of winning. A 9-to-10% chance — Trump’s current estimated probability of winning the White House — of a crash is still a terrifying prospect if you are on a plane.
This article, my second in a series on forecasting the White House race, will take a closer look at the latest electoral projections, as well as Trump and his challenger Joe Biden’s performance on Twitter in the past month. My first post can be found here.
1. JUDGE ME BY MY SIZE, DO YOU?
Trump has seen his number of Twitter followers grow by 6.8 times in the last four years — from about 12.8 million in October 2016 to the current figure of 87.3 million. On average, total daily interactions for his tweets have gone up about 5.3 times, based on my analysis of two month-long samples of his tweets from September 20 to October 19 2016, and the same period in 2020.
At this point in 2016, his tweets received an average of 613,042 interactions (sum of the number of retweets, “likes” and replies) per day. This figure has since shot up to an average of around 3.26 million per day between September 20 to October 19 2020.
The red and pink-dotted lines in the chart below highlight Trump’s lead in terms of absolute number of interactions on Twitter. But on closer examination, his “Twitter lead” is not as impressive as it looks at first glance.
For one, Biden is holding his own against Trump pretty well on Twitter with “just” 11.3 million followers. The Democratic challenger averages 1.99 million interactions on Twitter a day in my 2020 sample, more than three times what Trump managed in 2016 with a comparable followers-base.
The two candidates’ performance gap on Twitter might have been even closer, if Trump had not contracted Covid-19, a major news event which focused a lot of additional attention on the White House incumbent’s social media comments.
However, a very different picture emerges when I compare the two candidates’ interaction rates on Twitter, which account for the changes in the size of their followers-base. Interaction rates are better for direct comparisons of how two rival accounts performed over time.
2. LESS THAN MEETS THE EYE
The interaction rate for tweets is easily calculated: Divide the total number of interactions (retweets, likes, and replies) by the account size at the time of posting (a more detailed explanation is available at the end of this article).
In my samples, I calculated the interaction rates for each tweet in 2016 and 2020 before grouping them by day and taking a daily average. The chart below shows how Trump’s tweets are in fact less engaging than those by Biden, or even by Trump himself in 2016.
In my sample, Biden’s Twitter interaction rate comprehensively beats Trump’s across the entire home stretch of the campaign.
The interaction rates for Trump’s tweets are in fact consistently lower in 2020 than in 2016, with the exception of October 2 to October 6 when news of his Covid-19 infection appeared to have given his social metrics a boost. I observed a similar trend on Facebook, but could not find comparable 2016 data for Trump’s FB posts at the time of writing.
Another way to look at how Trump is underperforming on social media is via the dashboards available from social listening service provider Socialbakers. The two screen shots below show how Trump is leading in absolute interaction numbers (due to his higher number of followers/FB fans) but trailing badly in engagement rates, such as “average interactions per 1000 followers per tweet”.
These numbers, however, are not as dismal for Trump as those in the forecasts for the election results on November 3.
3. STEADY AS SHE GOES?
When Trump announced his Covid-19 infection on October 2, political pundits braced themselves for potentially huge changes to the polls and election forecasts. But surprisingly, the “October surprise” for the 2020 White House race did not shake up the polls or forecasts by much.
My latest projections, based on forecasts by FiveThirtyEight and The Economist, barely budged from the initial version. With two weeks to go, two separate algorithms arrived at the near-consensus forecast of about 338–342 electoral votes (EV) for Biden, and between 195 to 196 electoral votes for Trump on Nov 3.
My previous forecast on October 1 had anticipated Biden to win with 344 EVs and Trump to lose with 190 EVs.
To recap: The White House race is not decided by the popular vote but rather by the electoral college, a unique system where the successful candidate has to stitch together a winning coalition from various states that would give him or her at least 270 EVs out of a possible total of 538. Not all states are equally influential in the outcome, with the election results often coming down to how Americans vote in a few critical “battleground states” like Florida, Pennsylvania, Wisconsin and Ohio.
For this new iteration of the forecasts, I included the projections from a second model using XGBoost, a popular machine learning algorithm. My initial forecast used Prophet, a forecasting tool developed by Facebook. A detailed explanation of my methodology is available at the end of this article.
I also made use of additional data in FiveThirtyEight and The Economist’s forecasts to project Biden and Trump’s chances of winning the electoral college:
Again, the two models don’t deviate significantly in their final projections — both give Biden about a 9-in-10 chance of winning the election. Trump, meanwhile, has seen his chance of winning the election slide steadily from around 20% to about 10%.
4. GAME OVER?
Political pundits are still recovering from the shock of the 2016 US election, and many have pointed out that Hillary Clinton also enjoyed a clear lead in the polls at this point in the election four years ago, as this chart from RealClearPolitics show:
But there are clear differences between 2020 and 2016. Biden’s polling lead over Trump, currently at about 8.6% compared to Hillary’s 5.6% in 2016, has been more consistent. More importantly, Biden has not once trailed Trump in the national polling average over the past year, whereas in 2016, Hillary was overtaken by Trump on two occasions (indicated by the small red patches in the chart on the bottom).
Still, many continue to wonder: Could the polls, forecasts and social media metrics be disastrously wrong yet again?
Signs are that the pollsters and the Democrats have learnt from their mistakes in 2016. But no one is prepared to rule anything out at the moment, especially in a year that has been rife with black swan events.
DATA, METHODOLOGY AND REPO
This article uses data from four primary sources — Twitter, Facebook and the respective daily forecasts by FiveThirtyEight and The Economist.
The tweets examined include:
- 431 tweets by Trump sent between September 20 and October 19 2016.
- 681 tweets by Trump and 385 tweets by Biden sent between September 20 and October 19 2020.
The FB posts examined include:
- 674 posts by Trump and 276 posts by Biden sent between September 20 and October 19 2020.
The total interaction per post is calculated by summing up the number of likes, replies and retweets (or in the case of FB posts, by adding up the number of likes, shares, comments and reactions).
The interaction rate per post is calculated by dividing the total interaction score by the size of the fan base at the time of posting. No weight was assigned to the various forms of interaction, meaning retweets were treated as equally important as replies.
I also aggregated two separate sets of forecasts, made between June 1 and Oct 19 2020, by FiveThirtyEight and The Economist — for electoral vote count and probability of winning the electoral college:
The aggregated predictions are then used to generate time-series forecasts up to November 4 using Prophet and XGBoost.
The Github repo for this post, containing the data and notebooks for the Prophet and XGB forecasts, can be found here. I’ve not included the datasets for the social media data out of copyright concerns.
If you spot mistakes in this or any of my earlier posts, ping me at:
- Twitter: Chua Chin Hon
- LinkedIn: www.linkedin.com/in/chuachinhon
