As I was going back through my Kalman Filter code, I realized that I left in a hacky shortcut that I meant to go back and fix but never did. That shortcut was that I said that on any day where a poll showed 0% for a candidate, all polls from the same day showed 0% for that candidate, which is obviously wrong. I have since fixed it so that it is now a full weighted sum (where 0% is said to have the same variance as 3%), however when fixing this problem, I noticed another, much larger problem.
In the Kalman Filter, there are two terms, Var(et), which is the sampling variance from the polls. Then there is σu2, which I assumed to be the same thing. This is because I read the u in the subscript as a t. This led to a pretty large flaw in the code1, which honestly only translated into a small change in the chart. Nonetheless, read your subscripts, and here are your new charts!
Footnotes [ + ]
|1.||↵||As always, you can see the change on our Bitbucket.|