I started my this website last October. I created a forecast for the Senate elections, in which it was very accurate. I missed two races, Indiana, which I had as a tossup, and Florida. It had a Brier score of .04843. I then made a few tweaks to the model, but kept the car the same, just adjusting a few tweaks under the hood. I then implemented this model into a presidential model for 2020. It has gotten a little bit of traction and following, but how well will my model translate from a senate forecast to a presidential forecast?
Well the best way to test this is to back test this. I gathered the data I needed from the 2016 and implemented it into my model. I did not change my model at all beside the data. It actually surprised me how well it did. With very little back testing before besides the 2018 senate results, I created “the most accurate forecast” of 2016. I use those parenthesis for a reason. I am a little hesitant to use that title for my model as it was created this year and not in 2016. But what I am getting at is how well the model was built, withstanding itself to 12000 polls not knowing how well it held up. Below is a little overview of the model.
(Shouldn’t be keep the white house, should be win)
It overall was much more bearish on Clinton’s win chance across the board. However it still favored Clinton winning the election. One thing I think my model was good at was getting the separation of the popular vote and electoral vote in its simulation. It gave Trump a 2/3 chance of losing the popular vote if he wins the Presidency. It missed the popular vote by 1.5%, and average a 5.0% miss on margins. It preformed around average on battleground states. Below is a chart of how well it did in states not considered solid for Trump or Clinton.
The last part I am going to get into is comparing it to the other models in 2016. So how I am doing this is using a Brier Score. It treats events a 1 or a 0. It gives props to when confidence is used correctly, but when wrong hurts a lot. The lower the better. The formula is probability minus observed squared. I got these numbers from a huffington post article. Below is a chart of forecasters accuracy.
It predicted the most races and had the lowest Brier Score in 2016. I am still cautious on calling my model the most accurate, but it is up there with the best imo. I hoped this convinced you to keep up with my model as 2020 evolves. Roll Tide!