Mired in a growing frustration with how political polls were being reported on, a Quebec astrophysicist tried his hand at aggregating polls and projecting the 2018 Quebec election. Three provincial elections later, Philippe Fournier is hoping to correctly predict 90 per cent of the winning candidates of the Oct. 21 vote.

From coast-to-coast-to-coastfrom Nunavut to Skeena-Bulkley Valley, B.C., to Avalon, N.L.338Canada dives into individual races, as the websites name suggest, across all of Canadas 338 ridings.

I was looking at some Quebec polling before the [2018] election, and I noticed that many articles in newspapers were just badly written. Some journalists are told to write about polls when they dont know much about polls and statistics, Prof. Fournier told The Hill Times in a phone interview, a little more than a week before the Oct. 21 election.

I told my students I could do much better than that, he said.

Prof. Fournier teaches astrophysics at Cgep de Saint-Laurent in Montreal, where he is currently teaching only part timeas he currently spends 70 to 80 hours per week on 338Canada.

He first became involved in polling aggregation in the Quebec provincial election in 2018, when he started writing about polling projections on his website qc125.com.

It took about three months and then La Presse in Montreal contacted me asking me questions and political parties contacted me asking me questions. So it kind of became viral, Prof. Fournier said. It went so well that after the Quebec election, I figured, well, why not do a Canada-wide system.

The distinctive feature, Paul Adams, a journalism professor at Carleton University and former EKOS pollster, said of 338Canada, is the individual riding projections.

To get insight on the individual ridings, an aggregator takes the regional and subregional polling results, and apply them to historical patterns, Prof. Adams said.

As of Oct. 15, 338Canada is projecting a close race between the Conservatives and Liberals. The Hill Times photograph by Andrew Meade

Where aggregators can miss in its projections is where there is no historical baseline.

In this election, the obvious one would be Maxime Berniers party [the Peoples Party], Prof. Adams said. We dont actually know where we would expect them to run stronger.

How the Peoples Partys 2.8 per cent national support, according to 338Canada, will be distributed in certain individual ridings remains an unknown, Prof. Adams said.

So far, Prof. Fournier has worked on three provincial elections where he correctly projected more than 90 per cent of the winning candidates over the three votes. In the 2018 Ontario election, which resulted in a Progressive Conservative government, he identified 111 of the 124 winning candidates; 11 of the 13 misses were within the margin of error. In the 2018 Quebec election, he identified 112 of 125 correct candidates; of the 13 misses, four of them were in the margin of error. His most recent projections during the 2019 Alberta provincial election were his most successful, identifying 94 per cent of the successful candidates. He projected 82 of the 87 winning politicians. Of the five that he missed, three were within the margin of error.

I dont aim for perfection because its statistics and I know its impossible, Prof. Fournier said about projecting the Oct. 21 federal vote. I have this threshold that I want to reach: 90 per cent. But 90 per cent means I will miss about 35 [ridings].

The election right now is so close that I might miss more. So it might be 85 per cent. But my desire would be 90 per cent. Above that is unrealistic, he said.

As of Oct. 14, 338Canada was projecting the Conservatives to win 135.8 seats, and the Liberals to take 134.9 seats, both with massive margins of error. As he explained in Macleans, Prof. Fournier said his model has no fewer than 139 of 338 electoral districts labeled as either toss up or leaningmeaning more than a third of all ridings remain too close to call. A week before the election, the Bloc was projected to win 32.4 seats, the NDP 30.1 seats, the Green Party at 3.8 seats, and Independents and the Peoples Party at less than one seat apiece.

Prof. Fourniers model takes into account demographics of ridings, as well as historical performance and the effect of star candidates.

Pollster Greg Lyle, president of Innovative Research, said in cases where there is a strong candidate challenging an incumbent its hard to quantify what will tip the scales.

Mr. Lyle pointed to the current campaign in Kamloops-Thompson-Cariboo, B.C., which has been held by Conservative MP Cathy McLeod since 2006. But her current Liberal challenger is Terry Lake, a former Kamloops mayor and B.C. health minister.

A lot of Conservatives have been used to voting for Terry Lake, because provincially they vote B.C. Liberal, he said. And Terry Lake was very well regarded, so he has a strong personal brand.

In that seat, would you put your finger on the scale for the Liberals if it was close, Mr. Lyle said. But once you do that, now you are not betting on the methodology anymore, youre betting on the grey area.

As of Oct. 15, 338Canada is projecting that Ms. McLeod is likely will win re-election, with Mr. Lake finishing a distant second place.

Prof. Fournier said initially some pollsters werent all receptive to his modelling.

There are amateurs online and Im sure those amateurs really grind the gears of many pollsters. I know some pollsters really dont like aggregators, he said.

When he first started Qc125.com, Prof. Fournier said pollster Jean-Marc Lger was initially not pleased with him.

[Mr. Lger said,] I didnt know you were a scientist, I thought you were just some guy on the internet pulling numbers. Like everything on the internet, there are amateurs that just do anything and there are the serious people, Prof. Fournier said.

In addition to the website, Prof. Fournier also writes about his polling aggregate projections for Lactualit and Macleans.

When looking at historical examples, Prof. Fournier takes into account the 2011 and 2015 elections.

I look further in the past for some districts, he said, but the thing is the demographics in some parts of the country, especially the urban areas changed so fast [that] they would not be very useful to use the 2004 [or] 2006 numbers.

Prof. Adams said, generally speaking, aggregators are better than individual pollsters in predicting outcomes.

Pollster Frank Graves, pictured at the Green Partys 2016 convention in Ottawa, says aggregators can have the corrosive impact on a voters decision to not cast a ballot. The Hill Times photograph by Sam Garcia

Thats not to say that in any given election, there may be an individual pollster that comes out better than the aggregations, but the problem is you cant predict in advance which pollsters that is going to be, he said. Youre safer to stick with the aggregators.

There are two key concerns that aggregation has to focus on to influence the success of the model, Prof. Adams said. Aggregators have to decide the way each individual will be weighted and when to discard old polls for the aggregation model.

For Prof. Fournier, he will weigh a poll down in his model if the results are too out of line with the overall average. But, he said, if subsequent polling shows that the poll wasnt an outlier and was a precursor, it will regain its weight.

Outside of campaigns, Prof. Fournier said he can keep a poll in the model for around two months given how slowly the numbers move. As the poll gets older, the less weight it will have in the model.

During the election, Prof. Fournier said, he will only keep a poll in the model for a week.

Prof. Adams said if a poll is kept in the model for too long, it can miss the rapidly changing mood of an electorate.

If you take a weeks worth of results, you have a larger total number of cases, like you have tens of thousands of cases potentially adding up all the different polls, in a more stable environment that would give you a better result than just one or two polls at the end of the campaign. But if theres movement at the end of the campaign then a failure to decay older polls quickly enough will not serve you as well, he said.

EKOS Research president Frank Graves said he didnt think the aggregators are adding a service that cant be done elsewhere, as many pollsters put out their own seat projections.

Over the years, Mr. Graves said EKOS has performed better than the aggregators in projecting seats. In the 2006 federal election, EKOS projected the Conservatives would win 125 seats, plus or minus five seats. In the end, the party won 124 seats.

Mr. Graves said a seat projection should be within 20 seats of the winning party.

He added that if the aggregators are inaccurate or overstating their precision, there can be a corrosive impact on the voter decision making, as some people may decide not to vote if they look at a projection and see one candidate projected to win easily.

I would not rule out the fact that the aggregators contributions in the U.S. presidential election couldve been the victory for Donald Trump because a lot of disaffected, weakly-engaged Hilary [Clinton] voters were told, This is mailed in. Its done. Dont worry. And they stayed home [on election day], Mr. Graves said.

nmoss@hilltimes.com

The Hill Times

Neil Moss is a reporter at The Hill Times covering federal politics, foreign policy, and defence.- nmoss@hilltimes.com

View original post here:

The astrophysicist whose polling aggregator is projecting the election - The Hill Times