2025 Election Predictions
Here are the election predictions my Gaussian Process model has made. The model shows high accuracy with previous elections
Charts may be added later. A post election recap site at the bottom of the page.
For these elections, polling data was collected from wikipedia sections of these elections. Win% was made from a simulation of 10,000 elections.
What happened?
All the win%'s were correct in the statistical sense. Every race that flipped beyond what was expected were around the 20-25% chance range, a doable probability.
Now most specific vote percentage predictions upper bound's were slightly below the actual result. This is likely from two main factors:
1. This model split undecideds between all option. In reality it is clear that those polled as "undecided" leaned towards Democrats. Many polls are conducted that specific ask for a head-to-head or what candidates people lean towards. In future models that information can be used to probably handles those polled as "undecided".
2. Polls likely undershot. This model uses a complex form of aggregation and projection which is more accurate than traditional moving weighted averages, but is ultimately still based on polling data and previous election results. If polling data was poor, it will be reflected. In the time it took since I got this model working (months of on and off side work) to the election day, I did not have time to look for poll biases or reliability. That error may have also dragged the predictions down.
​
Something to also note is shrinkages. It is clear that in NYC many people polled for Sliwa strategically voted for Cuomo. The "other" category of vote was already shrunk to adjust for historical patterns between polling and voting, which proved right. I did not assume the same level of strategic voting would occur with Sliwa as there was no patterns to look for. That call was wrong.
What's next?
I will hope to get up a midterm model, in at most, 6 months. This model, instead of being run separately for each race like this one, will be run hierarchically to further strengthen results and give projections for every House and Senate race, whether they have individual polls or not. Fixes will be made to how pollsters and undecideds will be interpreted. I will also attempt to streamline the process of making projections to allow for more constant updates, while making on-site charts so they will work better on other aspect ratios. I will also likely give the option to replace the Monte Carlo win% with a more accurate cumulative distribution value.