listen. it's not a philosophical discussion. it's an empirical reality. ivr has a proven, successful track record of being accurate to the definition of what accurate means, in context. i'll give the panels some breathing room because it's still novel, but their track record to this point has been very poor. you can construct whatever rhetorical house of cards you want to describe how you think things ought to be, but the measure of how things actually are does not have to agree, and must take precedence. further, there is a good explanation: randomness is important.

the riding projections are really an expression of artistic freedom, based on a lot of very weak inferences. it's one thing to collect the data. it's another to really more or less make guesses on how it's being distributed. it's the exact wrong way to work if you're trying to build a model. you want specific, local data that builds - not broad, vague data that you have to guess how to categorize.

if you compare the leadnow polls to the riding polls, the thing that often comes up is that it can't predict swings that increase the vote share of the party that finished third in 2011. it will instead consistently increase the share of the second place party; the polling is telling us that this is too simplistic. the issue is probably simply ignored altogether in the algorithm. you could conceivably improve the model by adding in more complicated logic, but it's only likely to be a marginal improvement.

i can't be sure, but it seems to want to weight the data relative to only the last election. if it were to average out the last five elections (weighted) and fit it to the result, it would do a little more to find the center of gravity in the riding. that will probably be more predictive when you have sudden reversals of the previous election or several previous elections (as i think may be the case; we may be heading towards 2004 results, except with a weaker liberal party and a stronger ndp, as well as no bloc - or with the roles of the ndp and liberals reversed, but with liberals stronger than the ndp ever were in such a scenario), but won't help with unpredictable situations, which are unpredictable, of course. i think that this is likely the right way to weight the data in ontario, at least.

another option for this election specifically - i'm not claiming this will be true in general, or even often - may be to look at weighting the ontario data relative to the last provincial election. you'd have to promote such a thing as speculative, of course. but i think it may be illuminating both in finding methods that are more generally applicable and in predicting the outcome of this specific election.

http://ipolitics.ca/2015/09/23/strategic-voting-group-releases-new-riding-polls-but-hard-data-remains-elusive/