Monday, October 19, 2015

Projection Roundup

If you are a Canadian citizen over 18 and haven't voted yet, VOTE. You have until (all times local):
- 7pm in the Pacific time zone
- 7:30pm in the Mountain time zone
- 8:30pm in the Central, Atlantic and Newfoundland time zones
- 9:30pm in the Eastern time zone
Find your polling station at Elections Canada's website.

Click here for a summary of the projection.
Click here for the projected winner by riding.
Click here for the complete trends since early September.

Below are the projections based on polling aggregates - there are links to all of these on the left. They are listed in increasing magnitude of Liberal victory (difference between Liberal and Conservative seat counts).

L-C-N-B-G-Other (Vote L, C, N, B, G)
137-120-75-  8-1-0 (36.8%, 32.5%, 21.4%, 4.6%, 4.1%) Too Close to Call
138-120-75-  1-1-0 Election Almanac*
138-117-76-  6-1-0 (36.8%, 31.8%, 22.7%, 4.2%, 4.5%) Le calcul électoral
142-119-66-10-1-0 (37.3%, 32.4%, 20.1%, 4.9%, 4.3%) Canadian Election Watch
140-115-79-  3-1-0 LISPOP*
142-116-68-11-1-0 Election Atlas
146-118-66-  7-1-0 (37.2%, 30.9%, 21.7%, 4.9%, 4.4%) ThreeHundredEight
149-105-81-  2-1-0 David Akin's Predictionator**
160-120-50-  7-1-0 (36.7%, 32.0%, 20.4%, 5.2%, 4.0%) The Signal
178-115-44-  1-0-0 (37%, 31%, 22%, 4%, 4%) CVM Election Model
177-  95-53-11-1-1 (38.0%, 30.9%, 21.3%, 5.7%, 3.4%) Teddy on Politics
*Do not appear to include late polls from yesterday.
**Significant use of information other than polls.
***It's really 137.5-117-76.5-6-1 since there's a tie in Davenport. Pierre Martin has clarified that the Liberal percentage is slightly higher in Davenport; the Liberal number above for Le calcul électoral has been updated.

Also, the Globe's election model provides similar probabilities as I gave here.

Average (may not add up due to rounding):
150-115-67-  6-1-0 (all 11 projections)
143-117-71-  6-1-0 (9 Liberal minority projections)

The first thing to notice is that two models are projecting a Liberal majority. But both of them are quite puzzling: they only have the Liberals at 37-38%! I'm almost certain that if these models are right and a Liberal majority occurs, it would be due to offsetting mistakes: bad vote projection (too low for the Liberals) and bad model (too many seats for Liberals for a given vote projection). The CVM model is also an outlier in predicting that Elizabeth May will lose her seat, while Teddy and Politics is the only one forecasting a seat for Strength and Democracy. The Signal also looks weird: it is projecting a much bigger Liberal seat count than me on similar popular vote. My sense is that 160 seats is plausible for the Liberals, but seems quite unlikely with less than 37% popular support.

The other eight more "mainstream" projections are all within very tight ranges for the major parties:
LIB 137-149
CON 105-120 (115-120 excluding David Akin)
NDP 66-81

A few observations about comparing these eight projections:

- The divergence between proportional and uniform swing models for the Liberals and Conservatives has mostly disappeared. This is because, when a party (LPC) starts very low and starts getting competitive, proportional swing concentrates the rise in support in ridings where that party has the best chances of winning. However, if that party increases further, the increase in support in marginal ridings becomes similar in the two models. The latter is what happened with the Liberals in the past week. (Theoretically, if the party increases even more and becomes very far ahead of the other parties, uniform swing would give it more seats.)

- The divergence is instead showing up for the NDP. Excluding my model, the three uniform swing models have the NDP at 75-79 seats, while the two proportional swing models having incorporated the latest data have the NDP at 66-68 seats. This is due to the NDP decrease in Québec: it is now only holding on to relatively good seats, and proportional swing makes the drop in those seats bigger than uniform swing.

- So why is my uniform swing model more like the proportional swing models for the NDP? The answer is found in the popular vote: I have the lowest NDP popular vote of all projectors. This is due to two factors:
(i) I applied a mild change to some polls from earlier this week in order to reflect the change in voting intention since then; this cost the NDP a few tenths; and
(ii) the pro-CPC turnout adjustment.

- I also have a relatively high CPC popular vote, obviously due to the turnout adjustment. Too Close to Call also has a turnout adjustment, but it is smaller than mine; on the other hand, that model did not apply the change described in (i) above. These two differences appear to cancel out. The Signal also has a similar CPC vote, presumably because it missed some late polls.

- I have a relatively high LPC popular vote despite the turnout adjustment because of the change described in (i) above. Note that without the turnout adjustment, my projection of 150 seats (on 38.3%) for the Liberals is higher than all non-outliers.

- Finally, the Bloc seat count is incredibly hard to pin down - projections range from 1 to 11... Honestly, for the Bloc, it felt like throwing a dart with a blindfold since no pollster has come out with a regional breakdown of Québec since the NDP's plunge there.


Anonymous said...

The CVM Election Model is projecting that the NDP will win zero seats in Ontario (with the Liberals winning 90 seats), and that the Liberals will win all 32 seats in the Atlantic region. While this is not necessarily impossible, it does seem unlikely.

Bryan Breguet said...

I swear, if some of these crazy models are right because they'd have badly estimated the popular vote and are just lucky (Like Liberals get 39% tonight and the Signal is right...) I'm gonna lose it lol

Awesome post by the way! I used to write stuff like this but writing in newspapers made me give up a lot of these technical posts. I should do it more often.

Election Watcher said...

@Anonymous: That's batsh*t crazy!

@Bryan: I'm kind of afraid that might happen with The Signal... I'm glad you liked the post, and I think there's enough geeks online for this technical stuff :)

Anonymous said...

And now it appears that the Liberals have indeed won all the seats in Atlantic Canada... Crazy like a fox?

Election Watcher said...

Ha, yeah. Though they projected the Atlantic sweep on L53, N22, C21 vote. Actual vote was L59, N18, C19, meaning L-C and L-N gap 8 and 10 points larger than they projected. A few races were decided by less than that.

And the NDP zero seats in Ontario sure didn't come to pass.

Bryan: Are you gonna lose it now?

Bryan Breguet said...

lol yeah I'm not super happy. The Signal got it more right by having the NDP much more behind the Liberals in Quebec in votes, although I don't really get how they got the Liberals at 33% given the polls. Especially since they don't do like 308 and give the latest poll a much higher weight.

As for CVM, well honestly, that doesn't make sense. Great, they got the seats very close... I still maintain this is most likely luck. I'm guessing that with the actual %, they'd have the Liberals above 200.

I maintain that a Liberal minority was the correct "call" given the info we had. Probably higher than what I had (should have trusted these last minute polls more) but not at 170.

Election Watcher said...

Wait, isn't The Signal's Liberal vote share for QC 29.4%?

Yeah, CVM is indeed crazy.

For what it's worth, using uniform swing on the actual results, my adjusted polling averages should still have given the Liberals 170 seats (ATL 29, QC 31, ON 80, MB/SK 8, AB 5, BC 14, Terr 3) instead of the 142 (146 without riding adjustments) I projected. That's a big miss. The Liberals didn't just win more ridings than expected, but won many of the unexpected ridings by large margins!

However, using actual results, my model (stripped of riding adjustments) would have given 173 to the Liberals, which is not too bad. It would have given Liberals, by slim margins, many ridings that they actually won by a lot.