This was a triumph. I'm making a note here: huge success.
Well, sort of. With only two races to check, having my model get 2/2 races correct could have been just chance. Let's take a look.
Yesterday my prediction was for Petros to beat Saadiq by a margin of 39.6%-35.1%, and William to beat Kevin by a margin of 47.9%-35.5%. In reality, Petros won by a margin of 41.9%-38.8% and William beat Kevin by a margin of 53.0%-36.5%. On average, there was a 3.0% difference between what I predicted and what happened.
[Note: I realize of course that it's possible for someone to edit a blogger post. To cover my bases, I uploaded a screenshot of the post at 5:00 pm yesterday, and you can check the date the photo was uploaded.]
Part of the reason for the differences in the race for President was that I considered Anthony Goertz as a legitimate candidate. This was pretty much just a judgement call. My model typically lumps joke candidates and None of the Above into the same category which often works reasonably well. If I had put Goertz in that category, the race for President would have been predicted to be 43.9%-38.8% - only an average of 1.4% difference from the actual results.
To see how this compares with the results from before, take a look at this graph:
previous iteration of this graph, it looks like the data points fit quite nicely. How cool is that?
At the request of some people, I've added this little excel web app. It will let you pitch any of the candidates of this election against each other in a fierce battle. You can have all eight competing if you really want!
1) Please insert their name exactly as it was on the ballot. For instance, use "Josh Le" instead of "Joshua".
2) This used the budget values that they used for their actual election, so it may be a teensy bit unfair to pitch a candidate from an uncontested race against one that was contested.
Back to the original post:
Another thing I tried this year was to project the voter turnout before the election was finished. This graph shows the actual voter turnout and my projection of the actual voter turnout for each hour throughout the election:
The method itself of using the previous trends to project voter turnout may not have been super accurate (my mid-day Wednesday projection was off by 1.4%), but it's possible that it may get better as it becomes more refined.
Anyway the end there got kinda rambly - sorry about that. Look at me still walking when there's science to do!