Biden Remains Strong Favorite with One Month to Go

Despite the most news-filled week in recent memory, the big picture of the 2020 US Presidential Election remains clear. Joe Biden is a strong favorite to deny President Donald Trump a second term. In 10,000 simulations of my polls-based model Biden won 9,039 simulations (90.39%), Trump won 914 (9.14%) and 47 (0.47%) were tied.

Figure 1: Race Ratings/Biden Win Percentage by State. Ratings Legend: Solid >95%, Likely 75–95%, Leans 60–75%, Toss-up <60%
Figure 2: States/Districts Not Pictured Above

Although Trump’s electoral position is dire, his outlook has improved since my last update in August due largely to improvements in two key states, Florida and North Carolina. These improvements have nearly doubled Trump’s win percentage from a meager ~5% to a somewhat less meager ~9% over the past two months.

Figure 3: Candidate Win Probability by Month

It should be noted that since this model is polls-based, it hasn’t fully priced in the impact of the Trump tax release, Trump’s poorly received debate performance or his hospitalization for COVID-19. Based on initial polling these events seem to have had a negative impact on Trump’s standing, at least in the short-run.

Key Takeaway #1: A Biden Landslide Remains A Strong Possibility

Figure 4: Number of Simulations by Biden Electoral Votes. Color Legend: Red — Trump Win, Gray — Tie, Blue — Biden Win (<400 EV), Dark Blue — Biden Landslide (>400 EV)

Though Trump has made some modest gains over the past two months (at least prior to last week’s news dump), a Biden landslide (>400 electoral votes) is about 2.5x more likely (24.73%) than a Trump win (9.14%). Strikingly, the three most likely individual electoral outcomes are all Biden landslides (Biden wins 413, 416 and 422 electoral votes respectively).

Key Takeaway #2: Trump Likely Won’t Be Saved By The Polls Being Wrong, He Probably Needs Significant Poll Tightening As Well

Many point to the polls being wrong in 2016 for why Trump will win again in 2020 despite poor polling. However, the polls being wrong again will likely only get Trump so far. My model agrees with the consensus that Pennsylvania is the most likely tipping state. Biden leads high quality PA polling over the past month by ~6.8%. Looking at the 66 races that Real Clear Politics has polling averages for going back to 2004, there have been only two differences greater than 6.8% between the final RCP margin and the actual result:

Figure 5: The difference between the RCP Final Presidential Polling Margin and the Actual Result 2004–2016

Only one difference (notably Wisconsin in 2016) benefitted the Republican by more than Biden’s current 6.8% lead in PA. Let that sink in: the polls were off by more than Biden’s current lead in PA in favor of the Republican a single time out of 66 (1.5%). As such, Trump likely needs some significant tightening in the most likely tipping point states (e.g. Pennsylvania) to have a material chance on election day. This is especially the case given polling adjustments such as educational weighting that pollsters have made since 2016 in an attempt to fix former D bias in the polls.

Deep Dive — Advertisting and Strategy

Campaign ad spending offers intriguing insight into campaign strategy. I pulled Advertising Analytics data on traditional media (tv/radio/print) for both the Biden campaign+D groups and the Trump campaign+R groups to see who has booked more spending in each state through election day. Overall, D groups have a 27% advantage over R groups in bookings over this time period ($170 million to $134 million).

Figure 6: Which campaign+aligned groups booked more spending in traditional media for October and November (source: Advertising Analytics as of end of September). Note: bookings are subject to change and this map doesn’t include the $22 million dollar advantage Biden has in nat’l advertising over this time period.

With the caveat that these campaigns have private data we don’t have access to, both sides have some head stratchers in their spending. For Biden, he’s spending money in Virginia, Colorado and Oregon! and getting badly outspent in Florida. The first three states are all Solid D while Florida is the second most likely tipping point state. Perhaps those Solid D state buys were a hedge against the race tightening to the point where those states were competitive (as evidenced by Biden’s recent cancellation of some spending in CO and VA). As for Florida, Biden could be holding back spending in anticipation of yet-to-be-booked Bloomberg spending in that state as part of his $100 million dollar Florida spending pledge.

For Trump, his spending is a mixed bag. My model thinks the three most likely Trump electoral vote win totals are 274, 279 and 285. These maps correspond to Trump winning all tossup states (Iowa, Ohio, Georgia and North Carolina) as well as Florida, Pennsylvania and at least one of Nevada and Arizona while limiting Biden’s flips to Michigan, Wisconsin and Nebraska-02/Maine-01. Trump wins the election a mere 1% of the time when he loses Florida and 4% of the time when he loses Pennsylvania. Trump seems to be driven to flip at least one Clinton-won state (hence Rs outspending Ds in Minnesota as well as Massachusetts to persumably get into the New Hampshire market). Given the overall spending disadvantage of Republican groups, it seems prudent for R groups to redirect spending from long-shots like Minnesota to the states closer to the national margin (namely Florida, Pennsylvania, Arizona and Nevada). If Trump’s set on trying to flip a Clinton-won state, Nevada is his best bet (and he’s already outspending Biden there). Trump not only has the best chance of winning Nevada (18%) of the Clinton-won states but also Nevada crucially lies on one of Trump’s best paths to 270.

Model Methodology

Political models are often a black box. Of late, there’s been increased competition to create the most complicated, often inscrutable model. With additional complexity, there’s additional opportunity to inject one’s biases into models. As such, I aim to keep my models as transparent as possible with no more complexity than necessary to keep myself honest.

My model is polls-based and focuses on keeping data inputs as high quality, recent and specific as possible for each state. All polls used were completed between September 1 and October 1, were of likely voters (preferred) or registered voters and had a 538 pollster rating of “B” or better for national polls and “B/C” or better for state polls. If a pollster (e.g. Yougov) had multiple polls over that time period, only the most recent poll was included.

For each state, if there were no state polls meeting the criteria, I took a uniform swing from 2016 election results using national polls meeting the criteria to estimate the margin in that state. If there was one state poll meeting the above criteria, I averaged that state poll with the uniform swing margin to get that state’s estimated margin. If there were two or more state polls meeting the above criteria, I averaged those state polls to get the estimated margin in that state.

Next, based on historical data on poll predictiveness leading up to an election and state variability, I built distributions of expected results for each state. For the first part, the spread of the model’s distributions slowly decreases as the calendar approaches election day. Moreover, for states with a significant percentage of voting already complete via early or absentee voting (>10% of 2016 total vote), I slightly reduced the distribution sizes to reflect reduced uncertainty. For the second, certain states have higher varability than others. For instance, states with less diversity and more elastic voting populations (e.g. Iowa, Montana) have a wider ranges of possible outcomes than racially polarized states in the south (e.g. Mississippi, Alabama).

After finding the expected margin and distribution for each state, I divided the states into different regions based on their demographics. Using these regions, I built a correlation matrix with states in the same region (e.g. Michigan and Ohio) strongly correlated and states in different regions (e.g. Michigan and Arizona) moderately correlated with each other. Finally, I ran 10,000 Monte Carlo simulations to get the above results.

Notes: 538 grade “B-” and “B/C” state-specific polling has been incorporated into this update where it was not in previous updates. Additionally, the model now makes adjustments based on the quantity of high quality state-specific polling. For example, if a state has 2 distinct high quality polls over the past month, that state’s outcome will be somewhat more uncertain than a state with 7 distinct high quality polls over the past month all else held equal.

A special thanks to 538 for the robust, easy-to-download data.

I’m a Kenyon & U Michigan alum who writes about analytics and elections.

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store