As the party conventions are set to kick off, we are presenting the 2020 Decision Desk HQ Election Forecasting Model, developed in conjunction with 0ptimus Analytics for the second straight cycle. Election models are no longer unique qualitative takes on a political elections. Rather, they have become somewhat commonplace analysis pieces that permeate the discussions among political enthusiasts and novices alike. Our hope is that our model services as a net positive addition to the quantitative discourse around U.S political analysis and thus we have designed our forecast to be robust and transparent. New for this year, we are submitting our forecast to the Harvard Data Science Review where the methodology will be reviewed by academics.
You can view the model here.
In 2018, we released our first forecast model, predicting the probability of victory in each of the 435 House races and 35 Senate races, as well as simulating the 2018 election 14,000,605 times to produce a probability of chamber control for the Republicans and Democrats. It performed very well- predicting 233 Democratic House Seats (actual: 235) and 52 GOP Senate seats (actual: 53). Now in the home stretch of the 2020 election, we are ready to reveal the continuation of that project.
2020 Model Highlights
For 2020, we’ve taken the same model that was predictive in 2018 and added new features, such as additional engineered fundraising metrics, political betting markets data, and more demographics (Asian and Cuban-American population, population density). We have also made additions to the Decision Desk HQ Poll Average, incorporating the recency and probabilistic modeling elements that produced good polling estimates in 2018, while adding pollster ratings into the equation. We also experimented with Google Trends, further modifications to Partisan Voter Index, and wage inequality by state- before concluding that- while cool- did not help improve the model.
The Presidential model is new, though built similarly to our tested formula for Congress. Less important in the Presidential model are candidate financing variables- these are less predictive for Presidential elections, and more important are demographics and political environment factors. For the national simulations, we incorporate the potential for strong correlated errors, similar to what we saw in the Midwest in 2016. If in a given simulation, Trump wins Pennsylvania, he becomes a favorite to win Michigan and Wisconsin as well. Conversely, in simulations where Biden wins Texas, his odds of winning Georgia, Florida, and Arizona grow significantly. Given the additional level of complications with Electoral College simulations, we have to settle for 140,605 simulations here- though that is still plenty.
Why Another Model?
Different people will engage with our model, and those of our colleagues, in different ways. We can’t say we know what all of those ways will be or that they will be considered good by all. We do know based on our own experience with providing election results and as engaged members of society, new ideas, information, and opportunities to access data will lead down paths we can’t imagine. That’s the story of human advancement. Technology has enabled collaborations in ways never before possible. We hope our contribution to this field plays some small role as a building block for something useful and of value that we can’t even imagine today.
So with that hope, and a sense of responsibility to be as thorough and transparent as possible, we present the 2020 Decision Desk HQ Election Forecasting Model, developed in conjunction with 0ptimus Analytics.