I know this is *veeeerrrry* hard when constructing a model, especially when there's so much indecision about what "path" to take or which factors to consider in, but if I have one Slight suggestion to make -- build the forecast more "blindly" aka not letting yourself view the output until it's completely finished.
Usually when we try to get our numbers to fit and match our preconceptions (often misconceptions) about a race, our unconscious bias seeps into our evaluation of where a contest stands and the impartial neutrality of the numbers themselves gets pushed to the backseat.
What you've managed to done here is SUPER impressive PM! Million pats on the back, what a gem!!
That is definately a very good point.
No statistical model for the 2020 Presidential race is truly going to be non-biased, it's more of an aggregate of a way to interpret the data. Deciding what data to include, how it's weighted, and how it plays into the whole model is inherently biased, and sometimes the data your using itself is biased.
Some of the adjustments I made were thing I knew needed to be changed, but didn't know how for the longest time, such as making it so states with more polls have their polls decrease in weight faster over time than states with few polls. Had to get my math teacher to help me there. Other things were already set up but the data wasn't there yet (early vote numbers, for instance), so when I entered that there was an obvious shift in the model.
I certainly agree with you that not purposely changing the model to match the results you want is a good idea, however, there are some better ways to go about it such as back testing stuff, for example, the current senate model nailed 2018, except for FL, where Nelson had a 52% chance of winning, it never bought into Heitkamp or Bresden winning, though McCatskill and Donnelly were clear underdogs, and that TX-Sen would be close. Sometimes, I have found the things that seem off in the model to end up being correct with time; my model predicted SC would be competative before most people on this forum accepted that it was for instance.
I'm very curious to see how well it holds up on election day, so that I can reflect upon the strengths and weaknesses of the model for a 2022 model. Thanks for all positive reviews!
