the one thing i seriously got wrong was where the ndp were collapsing *to*. i figured they'd be around 25, but that the swing would go to the bloc. nobody had the liberals at 35 in quebec, but a few had the bloc at 25. so, i pegged the liberals about 160 and the bloc at about 35. that 25 seat difference on the 5 point swing in quebec, which nobody picked up, is what got me.

otherwise, i think my decision to base ontario results on the 2004 federal election rather than any kind of models (which, by the way they work, couldn't pick up the dramatic swings in cambridge or st. catherines or durham or kitchener or barrie - even if the conservatives squeaked a few of those seats out) was a better overall approach. it was pretty accurate in northern ontario. the conservatives were up a little on 2004 (and relative to most of the polling), and the ndp were down a little. if the conservatives had come down just a tad closer to the polling, the liberals would have won some of those seats - which all the models had as safe conservative seats.

i don't think we'll see another election like this for some time. the lesson is in the uniform distribution not applying well to huge swings. but, it's more on the level of requiring the caveat that the models are designed to distribute small changes.

i would suggest maybe pushing pollsters to get better sampling frames, though. if you could split ontario into three or four frames, you'd be better able to apply a uniform swing to those frames. another idea is to use the last five or six elections as a base, rather than just the last one, and then base the distribution on where you've calculated it's most likely to exist based on a broader survey of where it's existed previously. that's why i went looking for a close match in 2004 to compensate for the huge swing; but, what would have been better would have been to integrate new and old data together.

www.tooclosetocall.ca/2015/10/performance-of-polls-for-2015-federal.html3:38