Projection post mortem

Over the past few weeks, I engaged in one of my favourite pastimes - predictions. To aid me, I developed a bit of an aid in the form a projection model. Here I try to evaluate how it did and what its strengths and weaknesses were.

First, its biggest flaw is that it is almost wholly dependant on outside factors. It is a poll-driven machine and, if the polls are not accurate, then it is not accurate.

Here is how the polling worked out vs. the actual result (both contrasted against the 2006 results):

	CPC		Lib		BQ		NDP		Grn
	Poll Avg	Actual	Poll Avg	Actual	Poll Avg	Actual	Poll Avg	Actual	Poll Avg	Actual
Atl	-9.2%	-4.9%	-2.0%	-4.5%	-	-	+3.8%	+3.4%	+7.2%	+3.7%
QC	-4.4%	-2.9%	-0.1%	+3.0%	-0.8%	-4.0%	+3.9%	+4.7%	+1.8%	-0.5%
ON	-0.7%	+4.1%	-5.1%	-6.1%	-	-	+1.3%	-1.2%	+4.9%	+3.3%
Pra	-0.6%	+5.4%	-6.4%	-7.2%	-	-	+3.4%	+0.1%	+5.0%	+2.7%
AB	-7.1%	-0.4%	+3.5%	-3.9%	-	-	+2.9%	+1.0%	+1.8%	+2.3%
BC	+1.5%	+7.1%	-5.6%	-8.3%	-	-	-2.7%	-2.5%	+7.5%	+4.1%
Natl*	-1.9%	+1.3%	-3.6%	-4.0%	-	-	+1.6%	+0.7%	+4.7%	+2.3%

*National numbers excluding the BQ were used to project the 3 northern seats

As you can see, the pollsters were a bit off (or at least my average of their polls was). No pollster got it exactly right, though Angus-Reid came very close in both national and regional numbers.

Using the actual regional results, my model gives the following:
CPC 145
Lib 77
BQ 48
NDP 35
Ind 3

Using only Angus-Reid's poll, my model gives the following:
CPC 141
Lib 76
BQ 49
NDP 39
Oth 3

So, I think I am relatively pleased with the model, even with the slightly off polling results, I seem to have done better than most prognosticators for a change.

Here is a list of the gross error for all the predictions listed on Calgary Grit's round-up and myself:

1. Ekos Predictions (18)


2. Calgary Grit's model (23.6)
3. Barry Kay Seat Projections (24)
4. Andrew Coyne (26)
5. nbpolitico (28)
6. UBC Stock Market (30)
7. David Akin (30)


8. Kady O'Malley (35)
9. Democratic Space (36)
10. Election Prediction Project (38)
11. Scott Reid (40)
12. Andrew Steele (44)

If I had had a crystal ball or a better gut, using the real results would have projected a gross error of 8, while using the best poll (Angus-Reid) would have given an error of 6 - obviously showing that the model needs a bit of work!!

But of course I am forgetting my blogging colleague NBT. Who needs fancy projection models when you've got him? His gross error: 22, good enough for second place! As the prominently displayed link on my other site says, a better prognosticator than I!

Projection history: Final :: Eighth :: Seventh :: Sixth :: Fifth :: Fourth :: Third :: Second :: First :: Methodology