Note on Failed Projection Model

My projection for the election was wrong. Like off by 100+ seats wrong if you add up all the differences in seat counts. I'll start by saying that. But I don't actually dismiss my model. Here's why:

First, I'm mostly wrong in Quebec where I had the Bloc holding on and the NDP surge far less overwhelming. This is somewhat predictable, in fact, I basically said I thought I was wrong there when I put out my first projection for the province. My model is based on previous results so when the new result or poll is completely divorced from the previous reality it's not likely that the change will show up in large scale. It is inherently conservative, in the small c meaning of the word and most nights in Canadian politics that's not a bad thing. This year, it was a millstone.

The other reason for underestimating the size of the NDP surge is the lateness of the surge and how that screws up my polling aggregate. My aggregate is designed to give the most weight to recent polls but it doesn't exclude old ones. I actually played around with my aggregate on the weekend to try get something that would look close to what was going to happen. The closest I got adding weight to new polls and weakening old ones without completely fudging the numbers was CPC 37, NDP 28.5, LPC 22, BQ 6.5, GPC 5. Low for the Tories and Dippers, high for everybody else. Probably this weekend I'm going to use the real election results as a dummy poll in my model and see what would happen. That will be a better test of my model as a tool for extrapolating poll results which is what it is designed to do. If I get closer, especially outside of Quebec which might be a write-off, I think I'll have a little more faith in my methodology. It's all kind of moot as seat projection on a riding-by-riding basis like I have becomes a lot more difficult once the House gets reconfigured to match the new census as it should be the next time Canadians vote in 2015. I'm not sure how I can incorporate it into the model. Suggestions always welcome.