Election Models

Created/Modified: 2025-05-03/2025-05-04

This was written in pieces, during and after the Canadian federal election of early 2025, with the early bits edited for clarity and length... which is to say, to make them longer, probably.

As a guide for those who are only interested in some bits, I've broken this down roughly by original post on the topic, which I put on FB as that's where local people live. I'm writing this up here as I'm done with social media, which I've concluded is a fundamentally pernicious force, akin to the Krell Machine that let loose the "Monsters from the Id" in the movie Forbidden Planet. I'll mention--since this has been pointed out to me by a very clever person who thought it important I be informed of this as it was something that I could not possibly be aware of--that I know the "Id" is a concept that has fallen into well-deserved ill-repute. Feel free to write any other thoughtful corrections based on the Platonic assumption that I am an ignoramus living in a cave down on bits of paper and give them to crows to deliver to the wind, eh?

1. Election Models: some common misunderstandings, with explainations
2. Moral Choice in the Real World: we value outcomes, but choose actions
3. Final Pre-Election Thoughts: choosing actions that have the best chance of bringing about the desired outcome
4. Bayesian Updating: how to modify beliefs as evidence accumulates
5. Vote Splitting: the prediction of a vote split and recommendation for a strategic vote does not matter
6. Conclusion: election models provide useful information that is not used by enough people to matter

I live in the electoral district of Nanaimo-Ladysmith, which was hotly contested and--I am told--is very difficult to poll well. It's also not especially well-covered by polling organizations.

The four private political corporations (parties) that had any kind of a shot here were the Greens--who held the seat the last Parliament but one--the NDP--who had the incumbency--the Liberals--thanks to the Carney surge--and the Conservatives, thanks to vote-splitting among the other three.

"Vote splitting" is what happens when a group of people who all agree that they do NOT want candidate A divide their support among two or more other candidates, allowing a minority of voters to elect A despite the majority wanting anyone other than them.

"Strategic voting" is when that group agree to all vote for the alternative candidate who has the best odds of winning.

In Nanaimo-Ladysmith, a clear majority did not want the Con candidate to win, but that vote was split among the Greens, the NDP, and the Liberals.

Strategic voting is a question of who has the highest probability of beating the unwanted candidate, which is where I come in, as probability is something I know a bit about, including how to infer it more-or-less accurately from data, which is uncertain.

My preferred mechanism for identifying the candidate with the best odds of winning is election models, about which there was much disinformation spread this campaign, while at the same time there were a lot of individual, often partisan, polls being shared locally to try to indicate that this or that private political corporation (party) was in the lead and therefore deserving of the non-Con strategic vote.

My primary purpose here is to answer three questions: 1) Did election models get Nanaimo-Ladysmith within their plausible range? 2) Did election models provide useful information about who was likely to win in each of Canada's 343 electoral districts? 3) Does strategic voting plausibly affect election outcomes?

The tl;dr answers are: 1) Yes, but only just. 2) Yes, extremely well. 3) Nope.

Election Models

Canada has at least two good election models available: 338 and SmartVoting, both of which showed the Green candidate--Paul Manly--as by far the most likely contender among the non-Con parties, with at times ten times the odds of winning the seat compared to the Liberal (Michelle Corfield) or the New Democrat (Lisa Marie Barron, our incumbent).

Here were some things I kept hearing about electoral models, with an explanation of why they are either false or misleading:

1) Claim: They [338/SmartVoting outputs] are not polls.

This is true. But the implication that "therefore they do not accurately reflect the underlying reality" is not just false, but backwards.

Raw poll results by themselves are a very poor measure of reality. The general reason for this is "bias" and the bias comes in two different flavours: response bias and sampling bias.

Response bias is the fact that not everyone is equally likely to respond to a poll, and this is a very large effect. Less than 1% of people respond to polls at all, so the ones who do are already very odd in one important respect: they will talk to pollsters or fill out online forms or whatever. People of different age groups, different genders, different socioeconomic backgrounds, different religions, different partisan affiliations, etc, are all going to have different response rates.

A raw poll result does not reflect the underlying voting population very well. It contains some information, which is why polling companies can turn these noisy data into valuable information they can sell to their clients, but they do that via... modelling. Which is what 338 and SmartVoting do to interpret the publicly-available national and local polling data they use.

Sampling bias, on the other hand, is a catch-all term for the biases in the polling organization and the methods they use. This is sometimes called "house bias" or "house effect". Polls commissioned by private political corporations (parties) are not done randomly. The polling company knows full well what the desired result is, which may be different for different polls. A poll that is going to be used internally for targetting campaign materials may be focused on over-sampling "persuadable voters". A poll that is aimed at convincing the public to vote for your candidate over the other non-Con candidates is going to be aimed at over-sampling your competition's voters. Polling companies know how to do this: they know which districts swing, which are solidly green/red/orange, and so on.

One of the most basic disciplines of any experimental science is that if at all possible, you never measure anything just once. Ideally you make multiple measurements with multiple methods. Good polling companies do this, because each method (phone--which includes cell phones and has for decades--online, on the street, and so on) has different biases. All those biases have to be corrected for: you can't just average them and hope they cancel out. Randomness is not in general something that can be achieved.

So a single poll using a single method from a single organization that has a big bias is the least reliable measure of how people are actually likely to vote. This is why relying on local polls commissioned by private political corporations is a very bad idea if you want to know how people are likely to vote: those polls are deliberately constructed to NOT reveal that information.

Good models don't use polls commissioned by political parties for this reason. Good models do correct for biases very effectively. And in particular, no 338 result was in any way biased by any poll produced by any political party--the Greens or anyone else--because 338 does not use any polls produced by any political party.

Let me repeat that: 338 does not use any polls produced by any political party. So saying, "Well, 338 was wrong because it used that bias poll from the Greens"--which unfortunately I continue to hear--translates to, "I do not know what I am talking about."

2) Claim: They [338/SmartVoting] don't use local polling.

This is false. They use any publicly available poll that is produced by a reasonably unbiased organization. That includes local, regional, provincial, and national polls.

One can argue with their choices and ratings of polling organizations, but you can't claim they don't use local polling when it's available and they have good reason to believe it can be corrected for house bias and so on.

3) Claim: They assume proportionality between votes and seats.

This is false.

Among the many factors that feed into these models, there are corrections that are (semi-)proportional to national trends. The use of the word "proportional" in model descriptions is completely unrelated to any assumption about how votes turn into seats, which is not proportional, and the people who write these models know this.

The person responsible for this false claim appears to be a former NDP MP. I have no idea where they got the idea from. It may just be confirmation bias, which is a thing: they wanted to discredit best-in-class election models to salvage their private political corporation's odds of winning, saw the word "proportional" in the description, and made up a story about how that invalidated them, and how people like me are too stupid to notice, and how having an 89% record of accuracy on a district-by-district basis over 2000 races is somehow not very good.

4) Claim: They aren't accurate at the riding level.

This is false.

"The 338Canada / Qc125 model has thus far covered 18 general elections in Canada. In total, 2,039 electoral districts were projected.

So far, the model has correctly identified the winner in 1,821 districts, a success rate of 89.3%." -- 338's "Record So Far" page.

They achieve this accuracy by using local, regional, provincial, and national polling, as well as demographic and other data that allow them to infer and correct for various biases, with voting propensities turned into outcome odds (seats) via Monte Carlo modelling, which samples the probability distributions inferred from polling and other data to generate the range of plausible electoral outcomes on a seat-by-seat basis.

The accuracy of predictions follows the kind of curve of uncertainty you'd expect: seats called as "Safe" have a 98% accuracy record, ones rated as "Toss Up" are 63% accurate, which is about what you'd expect. A literal "Toss Up" would give a 50% call accuracy.

So that is, hopefully, a fairly comprehensive and extremely long-winded summary of why model results like those from 338 and SmartVoting are the best guides to likely outcomes in any given riding:

They use local, regional, provincial, and national polling data from multiple reliable polling organizations--excluding those from known-biased sources like political parties--as well as demographic, historical, and other data to calibrated models that can be sampled using Monte Carlo methods to simulate the election and infer the range of plausible outcomes, where seat counts are not proportional to national vote share.

Rules like "Vote for the incumbent" and "Vote for the Liberal" would make sense in a world where we didn't have reliable information about likely voting patterns, but that is not the world we live in. Thanks to 338 and SmartVoting--which are largely in agreement in this district--we do have reliable information, as as voters and citizens we should act on it.

Is any of this "100% certain for sure"? Of course not. But it's the best guide to the future we've got, and we should use it.

There are people who hate knowledge and value ignorance, and they have been quite strident in their claims that "polls should be banned" because they would prefer people not know anything. As someone who values knowledge above most things, I am obviously not a fan of this approach. Neither was Galileo.

There is a problem that this "ignorance is bliss" attitude avoids, though.

Moral Choice in the Real World

Thinking is hard.

Thinking about stuff we really care about is harder.

Thinking about probability, chance, uncertainty... this is the hardest thinking of all.

Sir Isaac Newton understood the subtle force of gravity and the rise and fall of the tides and laws of moving bodies, but he didn't understand probability: most of the basic ideas for thinking about it weren't worked out until a century after his death, and even now, two hundred years after that, we're still learning how to reason about chance.

Which sucks, because we're called upon to do it with depressing regularity, and it's almost always stuff we really care about. If it wasn't, we'd just shrug and walk away.

But we can't, because we care.

Caring wouldn't make thinking harder if we had bigger brains, but we don't. Our brains can keep three to five things in mind at once. That's all we've got to work with. People who talk about holistic thinking and stuff are asking the brain to something it isn't able to do: attend to more than five things, to treat a complex whole with all kinds of contingent bits and bobs, as if it was a single thing. It feels good, because thinking is hard, and now instead of having this whole complex interacting system to try to deal with our brain has one big thing. The details are kinda blurry, but so what? We're seeing it!

Which, sure. Sometimes taking a step back, grasping the larger context, can be useful. But if we don't dive back in and get our brain dirty with the details we're not going to get very far in the thinking department. Which sometimes matters.

Since we can only keep five (at most!) things in mind at once, and since our attention is naturally and strongly drawn to the things we care about a lot, we're pretty easy to manipulate and mislead. We do it to ourselves all the time. As the physicist Richard Feynman said: "The first principle is that you must not fool yourself, and you are the easiest person to fool."

So this is my brief (?) guide to thinking about uncertain things we care about a lot: which is to say, moral reasoning.

What is right action in an uncertain world?

It turns out it depends on how big the uncertainty is.

We value outcomes. In the present case, a lot of us would like a non-Con party's candidate to win the election here locally.

But we don't get to choose outcomes.

We get to choose actions.

We don't value actions. We value outcomes.

Philosophers didn't understand chance any more than Newton did, so they hacked together two basic approaches. One says we should always act based on the value of the outcome, regardless of principle. These are called "consequentialist" theories. The other says we should act on principle, regardless of outcome. These are called "deontological" theories.

There are completely different approaches to morality--virtue ethics, for example, which says what we "are" rather than what we "do" is what matters--but when focused on a problem in moral reasoning, like "Who to vote for in this damned election?" those are not necessarily helpful.

My take on the consequentialist/deontological split is that which one we should be depends on the degree of uncertainty there is connecting our actions--which is what we have the power to choose--to an outcome--which is what we value. The uncertainty varies by case, but it's always non-zero.

I want to really emphasize this: There is no universe in which we can simply choose an outcome "100% certain for sure". Any moral argument that starts with the idea that we can choose an outcome--rather than an action that will have some chance of bringing about that outcome--is a non-starter, because it's talking about a world we don't live in.

So we've already used up two of the five (or three) things we can keep in mind at once: action, and outcome. And because there are always at least two actions we can take, and at least two possible outcomes of each action, we've exhausted our limited attention before even starting to think. No wonder moral reasoning is hard. So it's best to take it slow. And long, apparently: this is my excuse for verbosity. Reasoning about right action is like solving the Towers of Hanoi. One step at a time.

Here's another world we don't live in: one where our actions don't matter at all.

"The outcome of this action is not 100% certain for sure therefore I can do anything I want," is not an argument, and we never, ever hear anyone say it except when they're pretty sure that moral reasoning is going to come to a conclusion they don't like.

I live on an island, and when I want to get some place else I go down to the ferry... which is not "100% certain for sure" to be on time, running, or even afloat. Admittedly, the latter hasn't happened (yet) but it certainly could. Yet when I want to cross the water, I go down to the dock.

The reason I do that is not just because it makes successfully getting elsewhere more likely, it makes it the most likely out of all the actions that it's in my power to choose, under the constraints I'm living with.

This is the general rule for moral reasoning under uncertainty: when we have some idea of how the odds of the outcome we want will change depending on our action, we should choose the action that has the biggest effect on those odds.

The biggest effect may still be very small. But so what?

But what about principle? Do the deontologists have anything going for them other than a deep sense of moral righteousness while they do something that actually reduces the odds of a good outcome?

Yes they do. Because in a lot of real-world situations we do not have a clue how our actions will affect the outcome. The probability distributions are so broad and flat in a lot of real cases that acting on principle is the best we can do. I don't see this as much different from virtue ethics (pace, Kantians) because we're doing stuff in these cases with no reference to the outcome, so whether we call the principle we're acting on "duty" or "character" doesn't make a big difference, at least to me. But I'm weird, so YMMV, as the kids say.

So this brings me--finally!--to the situation with Nanaimo-Ladysmith and the federal election.

In this case wave have a lot of information about the probable outcomes, and the effect of our vote, depending on where it goes. Assuming you're a non-Con voter, you can vote Green, Liberal, or NDP.

Models like 338 and SmartVoting use local, regional, provincial, and national polling data from multiple reliable polling organizations--excluding those from known-biased sources like political parties--as well as demographic, historical, and other data to calibrated models that can be sampled using Monte Carlo methods to simulate the election and infer the range of plausible outcomes, where seat counts are not proportional to national vote share.

Currently the Green candidate--Paul Manly--is the most plausible non-Con contender for taking Nanaimo-Ladysmith. Because the odds of winning are not proportional to the fraction of the vote in these models--as in reality--this means he has a 27% chance. The Liberal, who is just barely behind in likely vote count, has just a 3% chance. And the New Democrat incumbent less than 1%. The Conservative, who has a small advantage in votes over the Green, has a 68% chance of winning.

Where your vote goes in this mix matters to the outcome.

The graph I've shared here plots each candidate's chance of winning against their likely fraction of the vote, both based on 338 estimates, and it tells the story of how different votes affect the odds. Remember: we can't CHOOSE the outcome, but by our actions we CAN change the PROBABILITY of each different outcome.

People act as if probability isn't real. I sometimes describe us as "probability blind", which is like colour-blindness for chance. But we live in a world that is a constant flux of probability, and that's even before we get to quantum.

The present is made, moment by moment, from the probabilities of the past. And we can change those. So we can change the world. Not by much. Not all the time. But sometimes. And sometimes by enough.

We do that by making choices that make the biggest difference to the odds.

In the case of strategic voting in Nanaimo-Ladysmith, one less vote for the NDP or Liberals doesn't decrease their odds of winning very much, because they're already tiny, and can't actually become negative.

One more vote for the Greens increases their odds of winning by more. It is where our actions can make the biggest difference in the odds of the outcome "The Con does not win."

I've been told "the world doesn't work this way", but I'm reminded that "BC United" pulled out of the recent provincial election so as not to split the vote. So in fact the world does work this way. We know this because it has worked this way, and there is a well-know law in economics that says: "If something has happened, it is possible."

There is something I do routinely when I reach a conclusion--which in this case is, "A vote for the Greens is the best way to stop a Con from winning in Nanaimo-Ladysmith"--which is to ask myself, "What am I missing?"

I've only got a small brain, after all. Three to five things, with a lot of concerns vying for my incredibly limited attention. And I'm a natural Green voter, so advocating for a Green vote is not a big deal for me. I am, however, based on past form, confident that if the numbers showed the Liberal or New Democrat in the same position as the Green I'd be making the same case for them. People who don't know me probably don't believe that, and I can hardly blame them.

So what am I missing?

There's a bunch of stuff I've not mentioned but that I'm taking into account, because I'm not an idiot, so I don't need to bring it out explicitly or be reminded of it in the comments by people who've not bothered to read what I've written here. Things like:

1) All the data could be wrong. Sure. 338 has 90% accuracy at the riding level. That's a bad call 10% of the time. They're calling Nanaimo-Ladysmith as a toss-up, but who knows, maybe it's actually safe NDP and all this effort to switch votes to the Greens is going to put that at risk. Could happen. Prediction is hard, especially about the future. I know this. There are long tails, fundamental uncertainties, stuff I can't do a damned thing about. That is always the case, and this "argument" can be turned equally well against every single position. If anyone says it to me, I'll simply ask it back to them. Consider it done.

2) Things could change. Again, this affects every position that makes any reference to the data, and if someone's position doesn't make reference to the data I don't care about it. Maybe I should. Maybe that makes me a bad person. I've been told that often enough. Too bad. This is who I am. Don't like it? Stop reading. Go about your life. Be happy.

3) ???

4) I dunno. The problem with things I've not thought of is that it's often hard to think of them. I'll think on this more.

To bring this section to a long-overdue close: despite the uncertainties, we know enough to say with considerable confidence that switching a vote from the NDP or Liberals to the Greens is how to make the biggest difference in the odds of a Con win.

In 2021 there were about 69000 votes cast in Nanaimo-Ladysmith. That's consistent with the 62% turnout across Canada in that election, and there is a reasonable expectation we could break 70% this year. That means about 78,000 votes cast, and at the current fractions predicted by 338 that comes out to a little more than 2000 votes separating the Green candidate from the Con.

Switching 2500 votes from the NDP would reduce their odds from almost nothing to almost nothing. Switching 2500 votes from the Liberals would reduce their odds from 4% to 2%.

Switching 2500 votes to the Greens would change their loss into an almost certain victory.

Final Pre-Election Thoughts

We choose who we vote for.

We don't get to choose who our vote benefits.

This is worth thinking about. "Who benefits from this?" is a worthwhile question, to my mind. But my mind is a little different, so I've been told a time or two or three.

In both my professional and personal life, I'm much concerned with what the WWI poet Wilfred Owen called "chance's strange arithmetic": probability.

The ontology of chance is tricky. When we take an action, we make some futures more probable and other futures less probable.

Not certain, of course, because nothing is. But we can make some things more likely, other things less likely, by our actions.

In the case of voting in Nanaimo-Ladysmith, we can make a Conservative win more likely, or less likely, depending on how we vote.

The data tell a pretty clear and consistent story, as shown in the graph from 338 I've attached here: the Liberals and the NDP are competing for each other's vote, both trying to claw their way into third place. A vote for either of them does not materially change the probability of the Conservative winning.

You can choose who you vote for.

You cannot choose whose odds of winning are meaningfully changed by your vote.

If you vote Conservative in Nanaimo-Ladysmith, you increase the odds of the Conservative winning.

If you vote Green--Paul Manly--you materially decrease the odds of the Conservative winning, and increase the odds of the Green winning.

If you vote Liberal or NDP, you change the odds of who comes in third. Because neither one has a significant chance of winning, the impact of this vote on who wins is negligible.

This is an example of how the arithmetic of chance is strange. But being strange doesn't mean that it's wrong.

All of that is based on what we know today, using Monte Carlo models calibrated from local, regional, provincial, and national polls, and that use demographic and other data to predict riding-level outcomes with 89% accuracy over almost 2000 past races.

Could these models be wrong?

Do the people who wrote these models breathe air?

Those aren't really interesting or insightful questions, are they?

When you think in terms of chance's strange arithmetic, the interesting questions are:

Are these models more likely to be right than anything else?

Are these models less likely to be biased than anything else?

In both cases the answer is yes.

So vote! Vote for whoever! Vote your conscience! Vote your chakras!

But do so with this uncertain knowledge of how your vote will affect various people's odds of winning, which is given to us by the data and the strange arithmetic of chance, where a vote for Manly significantly changes the odds of who wins, and a vote for Barron or Corfield does not.

Really finally, written on the day before the election: I generated this histogram from the final 338 data and some assumptions about Green/Con and Liberal/NDP vote anti-correlations. It gives the Greens about a 14% chance of winning, which is a bit higher than the 338 value of 8%.

Since I'm in the prediction business, and I'm not a coward, I like to make crisp, definite, predictions based on clearly stated reasons so I can use the outcome to update the plausibility of those reasons, although in this case the outcome is so up in the air that the update will be modest, one way or the other.

That said, my prediction for Nanaimo-Ladysmith is that the finishing order will be:

Con, Green, Liberal/NDP, Volk und Reich (PPC).

The Liberal/NDP split is too close to call, with about a 60/40 split between them for third/fourth place, which may as well be 50/50. It would be silly to distinguish these outcomes based on the data as it stands today.

My reasoning is: 338 and SmartVoting models are a reasonable reflection of reality.

Ergo: if this call is correct, the plausibility of the proposition that 338 and SmartVoting are good models goes up.

If this call is wrong, the plausibility of the proposition that 338 and SmartVoting are good models goes down.

I tend to believe the models are good--thus my prediction--so if they get it wrong I'll be somewhat surprised, and if they get it right I'll be impressed but not amazed.

How big my surprise is in each case is a measure of how much I'll update my beliefs. My analysis gives the Liberals about a 25% chance of coming second when the Cons win, and then NDP about half that, so I won't be utterly shocked at any particular finish order, just a bit bemused. That means a modest update, a bit more or less plausible, depending on outcome.

Finishing order is a harder standard than calling the winner, but I think it's worth using as a test for model accuracy. I'll look at the whole country in the fullness of time, and update my beliefs about these models accordingly.

Bayesian Updating

This is sometimes called the "post-mortem" phase of data analysis, which is needlessly macabre.

It's Bayesian updating, which is the only way of consistently altering our beliefs in the face of new data.

Traditional views of knowledge were obsessed with certainty, much the way traditional views of matter were obsessed with incorruptibility. Gold was viewed as "pure" and "noble" and "incorruptible", and producing it from "base" substances like lead was the goal of students of matter--alchemists--for centuries.

It didn't work, because their whole understanding and conceptual system was based on nothing but human hubris and arrogance. The limited human mind naturally turns to notions of purity and corruption because disgust is one of our most basic evolved emotional responses to the world. Presumably it had some evolutionary benefit for us to not want to get near things that were perceived as marred or flawed: rotting organic matter, faecal matter, and so on.

When that purely human response was applied to metal it made about as much sense as talking about how sexy or tuneful it was.

Likewise, people have to take action or die, and choosing one action out of many possible ones is harder if we acknowledge uncertainty. "Confidence" is a trait people admire, and lacking the ability to reason about probability--which is fantastically hard--the simplest way to achieve confidence was to plow forward regardless. Philosophers noticed that confident action often came to a bad end, but still maintained that certainty was the appropriate goal for anyone who wanted to know what to do.

They were wrong.

"Certain knowledge" is as impossible as a "circular triangle" or a "square with only three corners."

To be knowledge, a belief must be supported by evidence, and as such it can always be updated in the face of new evidence.

A belief that can't be changed by new evidence is not knowledge, but faith, which is an epistemic error.

Bayesian updating is provably the only way of updating our knowledge in response to new evidence that is independent of the order in which we learn things. This is a desirable feature: it would be fairly odd to have two people who had exactly the same evidence in hand but drew different conclusions depending on what they learned first! I mean, we do have a tendency to privilege things we learn first--the phenomenon of "anchoring", which scumbag salespeople use to sell over-priced items--depends on this. But anchoring and similar biases are bugs, not features, when it comes to reasoning about objective reality.

As shown in the graph above, the model predictions were neither completely wrong nor quite as close as one might like.

The call I made was only correct for the winner. The Greens performed on the low end of the predicted range, and the Liberals on the high end. The NDP result was pretty much dead on the peak, though.

The first update I make looking at this is: I am too peak-biased. That is, I should have been even more emphatic about the range of possible results. I'm not astonished by the Liberals coming in second, because I said they had about a 25% chance of doing so, and 25% chances come up one time in four, but given the ranges and lack of separation of the peaks, I probably should have just said, "The other three party's finishing order is pretty much random, given what we know."

I'm not a big fan of those long lists of cognitive biases, as they are too unwieldy to be of any practical use. I find it more useful to focus on the One Big Bias, which is motivated reasoning. In this case, my motive was that I wanted there to be a sufficiently clear signal for action in the data, and a week before the election there was: the Greens were a clear stand-out from the other two. But that lead decayed, and I didn't update as aggressively as I probably should have.

Another interesting way to view the outcome is to look at the final result in the context of the predictive graph:

As can be seen from the graph, all parties came in within error of their modal result, but the Liberals, Greens, and Cons were all at or close to the extrema of their predicted range, which is not what you want, really.

It's interesting that both the Liberals and the Conservatives performed better than their polling suggested.

One claim that every single private political corporation always makes is "Our internal polling shows us doing better than the public pollsters say." This means that whenever a party does better than public polling suggests, their supports say, "See! We were right! Those dirty pollsters!"

But since everyone is claiming they're doing better than the polls show, the odds are as close to 100% as they can possibly be that someone will be able to make this claim.

And there is no way before the election is held to know who that will be.

This is one of those other problems of probability that people struggle with: an outcome can be both practically certain (for someone) and have almost zero probability (for any particular person or group) at the same time.

In the case of a party out-performing their polling in a given district, the odds are 50/50 if the data are accurate, although it's rarer for the difference to be as been as seen in Nanaimo-Ladysmith this time around.

How rare?

For that I had to look at how election models did across the country.

The top-line figures are shown in the table below, which demonstrates the actual seat count was less than one standard deviation away from the modal result for every party.

So on this basis we can say: election models did a good job of prediction given the data available. If, as an experimental physicist, I got a result that agreed with theory as well as this, I'd say the theory was pretty good, although the experiment could have been better. The error bounds are wide, but that's the nature of knowledge: we can't know things with more certainty than the underlying data allows us to. There is no magic.

I'm not going to go into any deep statistics here, but if you don't trust your eyeball, the Spearman rank, which is my preferred test of correlation as it's non-parametric, tells us the predictions correlate with the actual results very well. Anyone claiming that predictions provide no information--which is a claim I've seen--is going to have to explain these p-values:

Party	Spearman r	p-value
CPC	0.93	1.4e-153
LPC	0.92	3.0e-141
GPC	0.84	2.8e-93
NDP	0.81	1.8e-80
BQ	1.00	0.0

Now in fairness, extremely low p-values can still mean lousy practical predictivity. I should really generated a confusion matrix or something, but at this point I really can't be arsed, because anyone not convinced by looking at the graphs below is not going to be convinced by more formal analyses.

If we look at the Liberal and Con results vs predictions across all 343 districts, we get the graph shown here:

Fitting lines to the CPC and Liberal data gives very good fits with slopes near to unity.

CPCActual = 2.8 + 0.997*CPCPrediction

LPCActual = 2.8 + 0.951*LPCPrediction

Both fits have RMS error of 5.5, consistent with error claims of ~5%

With regard to the Green and NDP vote, as shown here, the SmartVoting model consistently over-predicted support in this case.

The party that the model actually did best on was the BQ, as suggested by the table above.

Vote Splitting

There were 78 districts ranked "Toss Up" or "Con Likely" by SmartVoting, 30 of which had an active vote split, and 48 were not affected by vote splitting.

Of the 30 with splitting, 23 went to the Cons, which is just over three quarters.

Of the 48 without splitting, 21 went to the Cons, or less than half.

Ergo: all this talk about strategic voting is meaningless when it comes to actually influencing the results of an election.

We may as well talk about people flapping their arms to fly to the Moon: we can imagine it, but it will not happen.

Conclusion

Election models provide useful information, and could be used to guide strategic voting, but there is no evidence that they do. Of the 30 toss up or Con likely races where there was vote splitting, the Cons won three quarters, as opposed to less than half when there was no vote splitting.

This tells us that people do not vote strategically, even when the future and possibly the very existence of the country is at stake.

So let's stop talking about things that have no chance of happening. There is nothing more pointless than saying, "People need to do X" when we know with as close to a certainty as can be that people will never, ever do X, regardless of how extreme the situation is.