This diary is an analysis of the model I developed to predict the seat-by-seat outcome of the Australian federal election held in September.
Unfortunately late polling released the morning of the election was somewhat worse for Labor than the consensus prior to that. My last full update used only the data available the day before the election. I mentioned in a last minute update that the model's final prediction was 91 seats to the Coalition and 56 to Labor but wasn't able to provide final seat by seat predictions due to time constraints (I work at Australian elections and the election day polling was released literally minutes before I was legally obliged to abjure posting on the internet). So this analysis post refers to the model's final prediction rather than my final posted prediction.
In reality the Coalition has won 90 seats and Labor won 55. However the Coalition also lost Indi to a conservative local independent and Fairfax to the conservative eccentric billionaire Clive Palmer (of Titanic II and world's-largest-dinosaur-park fame) so as far is the model is concerned the Coalition won 92 seats (the model can only actually predict if a seat will be won by the left or the right sides of politics). Therefore the model was, overall, off by only a single seat with Labor losing one more seat to the right then was predicted.
However the model purported to be able to predict the results and final two-party preferred percentages in each individual seat rather than merely an overall seat prediction. Therefore here are tables detailing the predicted Coalition winning percentage in each seat, the actual Coalition vote, and the model's predicted Coalition vote.
You might notice that eleven seats are missing from the tables. These seats had a "non-classic" two party preferred vote so Labor/Greens or Liberal/National or Coalition/Right-wing minor party were the top two vote-getters. Therefore I don't know the actual left-right two-party preferred breakdown it these seats, yet.
Labor had some bad luck in Tasmania. They won a majority of the two-party preferred vote but hold just one seat.
Overall the model was pretty accurate. Every seat ranked at least "likely" to be won by either party was (except Lingiari in the Northern Territory but I said not to trust that one) even though only 18 of 150 seats were regarded as a "lean" or "tossup". 30% of all seats were predicted within 1 percent of the actual result and a further 32% were within 2 percent (16% 2-3 percent miss, 14% 3-4 percent miss, 6.5% 4-6 percent miss, only Canning and Lingiari outside that).
Over the next six months or so I'll update the model for the next election cycle. There are data availability problems at present due to the probable fresh election for the Western Australian Senate.