In an election that pollsters, pundits, and the press said was too close to call, our partners at Advanced Symbolics Inc. (ASI) used industry-leading artificial intelligence (AI) to predetermine the federal election seat count with stunning accuracy. Their success will change how elections are covered across the country and around the world.

A day before most Canadians cast their ballots, ASI projections showed the Liberals winning 155 seats – they won 157. The AI model projected the Conservatives would win 118 seats – they won 121. The Bloc Quebecois was projected to win 29 seats – they won 32. The AI model showed the Greens winning 3 seats – they won 3.

The only projection that approached the outer range of probable was for the NDP. The AI model projected 32 seats as compared to the 24 that the party was able to pull off. Like many researchers in-field during the election, the AI model may have overestimated the support for the NDP due to the momentum surrounding its leader Jagmeet Singh.

Seat Data, each party










Elections are funny. In the aftermath of this one, it seems every party leader is claiming victory: Trudeau for pulling off a win, Scheer for growing seats, the Bloc for staging a comeback, Singh for holding the balance of power, and May for winning outside B.C.  But it’s not just the politicians claiming victory – several pollsters have been as well.

Some are congratulating themselves for having had the most accurate projection of the popular vote (based on “vote intention”) while others are rejoicing at having accurate seat projections. The curious thing is that those that got the popular vote estimate right didn’t put a seat projection out because their numbers were, again, “too close to call.”

Meanwhile, those that put out seat projections had estimates of the popular vote that had large total error scores. So, it seems, in this election vote intention data (being an estimate of the popular vote) did not accurately model seat wins. The aggregators, who rely solely on pollsters vote intention data, obviously suffered from this as well.

Overall, though, this was a better election for pollsters. As a 20-year veteran of the research industry, I was impressed by the accuracy of the numbers put out by many of the pollsters who reported during this election. It was a refreshing change from various recent elections when pollsters were, as a group, criticized for being off the mark.

For our part, we at Hill+Knowlton Strategies knew from the outset that this election was going to be close. In order to provide value insights to our clients about the potential impact it would have on their business and government policy, we knew we would need to do more than just the high-level vote intention estimates. Enter Polly ASI’s AI model.

We chose to partner with ASI during the election because of the innovative audience modeling technique it has pioneered. This appealed to us for two reasons; first, because we are fascinated with technical innovations such as ASI’s approach to audience insights, and, second, we wanted to drill our analysis right down to the seat level.

Working with ASI we selected strategically important ridings, built an interactive data dashboard, and we stayed focused on the only election outcome that has any real meaning for our clients – the seat distribution in the house. This culminated in being able to deliver amazingly accurate seat projections before the votes were even counted.

The future is bright for audience modeling and insights – a future that leverages artificial intelligence. We are finally realizing a vision many of us have had for a long time; to use a mix of methods and metrics that accurately peel back the layers of unique but equally important audiences.  It doesn’t take AI to predict that it will be a game-changer.

About ASI’s Methodology. Historical data is collected by ASI from the Canadian Census. Predictive data is from “Polly” an AI that predicts voter intentions based on publicly available social media data. Social media data is gathered from ASI’s online representative sample of over 270,000 Canadians using ASI’s patented CIC algorithm.