Predicting the Canadian federal election using search trends
- Fred Maude
- Reading time: 10 minutes
8 years ago, Alex Haynes and I published a piece of work correctly predicting a hung parliament for the 2017 UK general election. Whilst it has since been removed from Incubeta / NMPi’s site (ouch), the company we were previously colleagues at, remnants still remain dotted across the internet (The Drum covered it here).
That project was devised from a shared belief that the world of political polling and forecasting was heading in a difficult direction. Accuracy levels were falling as the online experience for voters became more complex, with polling organisations struggling to understand how this would impact their modelling.
Almost a decade later, the problem not only persists, but it has evolved in its complexity. The levels of investment political parties throw into social media platforms is now astronomical. Misinformation and disinformation runs amok. Hype bubbles and echo chambers look to throw analysts off course at every turn. The flow of information is constant. The experience, scarily intimate and personalised.
Alex and I continue to believe that more needs to be done to incorporate behavioural insights from online platforms if greater accuracy is to be restored. Some of this data, as we’re about to show you, is at our fingertips already. Other forms will rely on the likes of Meta and X giving greater access to their data architecture – a topic for another day.
For now, we’re returning to our trusty Google search trends as the key source of information – with some sprinkles on top. Our hope is that it provides a starting point to build from, whilst revisiting a conversation about how this data can be better used in the polling industry.
We’d also like to be right.
Why Canada?
In 2017, Alex & I were only just emerging from behind our sofas after the combined shocks of Donald J’s first election victory, and it-that-shall-not-be-named (Brexit). We were on the losing side in both, despite most pollsters communicating safety and warmth in the lead up to the events.
The extent of Donald’s latest victory, a similarly surprising outcome, has reignited our flames for understanding why modern forecasting methods get it so wrong. Thankfully things have settled down slightly in the UK, so we turn to our Canadian cousins for their moment of destiny on Monday 28th April.
Whilst proximity to the event of course plays a part, there are many reasons why this is a fascinating election to focus our methodology on. The majority of them come in the form of key recent events that could lead to huge shifts in voter opinion. Canada-US trade relations as a result of Trump’s recent tariff experiments. Conservative leader Pierre Poilievre’s MAGA-mirroring tactics now backfiring, having previously had a clear lead. The emergence of Liberal leader Mark Carney as ‘Captain Canada’, and his charge up the betting tables.
A voting population primed for short term decisions based on a rapidly evolving political landscape. The perfect cauldron for the forms of online influence and coercion that has stumped traditional pollsters.
Add all this together with the majority online population Canada holds, the significant investment both main parties have pushed into digital campaigns, and the fact it’s the tenth largest economy in the world, and it was an obvious place to focus.
So, here goes. Stick with us, it’s about to start reading like board game instructions.
The Methodology
Based on an initial search term volume check of their leaders, we narrowed the analysis down to 5 main parties – the Liberals, Conservatives, Bloc Québécois, New Democratic Party (NDP) and the Greens. We then took the 13 different defined regions in Canada and weighted them by the electoral ridings, assessing the search volumes for each candidate within these regions. This gave us a basis for understanding interest levels in each leader.
From there, we turned to ChatGPT (a not so surprising sprinkle, we know) to establish the key electoral themes being discussed. These were categorised as;
– Affordability & Cost of Living
– Canada-US relations
– Healthcare
– Housing
– The Economy (jobs, taxes, inflation)
ChatGPT then supported us for the second (and last) occasion in establishing how much each party was talking about each issue as a percentage of their total time.
Next, we used Google’s keyword planner to check relevant search terms for each of the five key issues, selecting keyword groupings that represented the most highly searched terms for each topic. This is where our first key insight arrived. Despite being a list of five, only two of the topics dominated online search trends – Affordability & Cost of Living and Healthcare. All of the media noise around the relationship with the US and it barely registers in what matters to Canadian voters. There’s a hype bubble for you.
We then weighted the key issues by party and gave the top party a weighting of 3, and the bottom a rating of 1. These weightings were applied to the search volumes by region to give a prediction of the number of ridings per region.
At this point, a slight adjustment to our process had to be made that took into account the unique political and cultural context of Quebec within Canada. Important issues here are seen much more through a regional lens, rather than national. To address this, we repeated the same steps but only on the Quebec region. Ultimately this helped us generate a much more accurate picture of Bloc Québécois’ performance in the region
The final act (well done for sticking with us if you’ve made it this far), was to balance how much interest in the party leaders mattered against interest in the policy issues. In our previous modelling, 80% of people decide who they’re voting for based on the candidate, not the issues. We therefore applied this weighting to arrive at the final results.
Our Conclusion
And so, our final prediction: HUNG PARLIAMENT (again), with the Liberals needing to form a coalition with one of the smaller parties.
Final standings (170 is needed to form a majority):
Liberals – 166
Conservatives – 118
Bloc Québécois – 16
NDP – 27
Green Party – 9
It’s challenging to define the political influence that mobile devices, and the platforms they connect us to, have on individuals. What we do know is that there is a clear impact, and that impact isn’t recognised in modern polling methods.
However, this exercise isn’t about saying that political forecasting needs to shift entirely to the digital. Instead, we’re trying to demonstrate that it needs to play a bigger part than it does currently. If nothing else, please take that away with you.
Also, we’d like to be right.
That project was devised from a shared belief that the world of political polling and forecasting was heading in a difficult direction. Accuracy levels were falling as the online experience for voters became more complex, with polling organisations struggling to understand how this would impact their modelling.
Almost a decade later, the problem not only persists, but it has evolved in its complexity. The levels of investment political parties throw into social media platforms is now astronomical. Misinformation and disinformation runs amok. Hype bubbles and echo chambers look to throw analysts off course at every turn. The flow of information is constant. The experience, scarily intimate and personalised.
Alex and I continue to believe that more needs to be done to incorporate behavioural insights from online platforms if greater accuracy is to be restored. Some of this data, as we’re about to show you, is at our fingertips already. Other forms will rely on the likes of Meta and X giving greater access to their data architecture – a topic for another day.
For now, we’re returning to our trusty Google search trends as the key source of information – with some sprinkles on top. Our hope is that it provides a starting point to build from, whilst revisiting a conversation about how this data can be better used in the polling industry.
We’d also like to be right.
Why Canada?
In 2017, Alex & I were only just emerging from behind our sofas after the combined shocks of Donald J’s first election victory, and it-that-shall-not-be-named (Brexit). We were on the losing side in both, despite most pollsters communicating safety and warmth in the lead up to the events.
The extent of Donald’s latest victory, a similarly surprising outcome, has reignited our flames for understanding why modern forecasting methods get it so wrong. Thankfully things have settled down slightly in the UK, so we turn to our Canadian cousins for their moment of destiny on Monday 28th April.
Whilst proximity to the event of course plays a part, there are many reasons why this is a fascinating election to focus our methodology on. The majority of them come in the form of key recent events that could lead to huge shifts in voter opinion. Canada-US trade relations as a result of Trump’s recent tariff experiments. Conservative leader Pierre Poilievre’s MAGA-mirroring tactics now backfiring, having previously had a clear lead. The emergence of Liberal leader Mark Carney as ‘Captain Canada’, and his charge up the betting tables.
A voting population primed for short term decisions based on a rapidly evolving political landscape. The perfect cauldron for the forms of online influence and coercion that has stumped traditional pollsters.
Add all this together with the majority online population Canada holds, the significant investment both main parties have pushed into digital campaigns, and the fact it’s the tenth largest economy in the world, and it was an obvious place to focus.
So, here goes. Stick with us, it’s about to start reading like board game instructions.
The Methodology
Based on an initial search term volume check of their leaders, we narrowed the analysis down to 5 main parties – the Liberals, Conservatives, Bloc Québécois, New Democratic Party (NDP) and the Greens. We then took the 13 different defined regions in Canada and weighted them by the electoral ridings, assessing the search volumes for each candidate within these regions. This gave us a basis for understanding interest levels in each leader.
From there, we turned to ChatGPT (a not so surprising sprinkle, we know) to establish the key electoral themes being discussed. These were categorised as;
– Affordability & Cost of Living
– Canada-US relations
– Healthcare
– Housing
– The Economy (jobs, taxes, inflation)
ChatGPT then supported us for the second (and last) occasion in establishing how much each party was talking about each issue as a percentage of their total time.
Next, we used Google’s keyword planner to check relevant search terms for each of the five key issues, selecting keyword groupings that represented the most highly searched terms for each topic. This is where our first key insight arrived. Despite being a list of five, only two of the topics dominated online search trends – Affordability & Cost of Living and Healthcare. All of the media noise around the relationship with the US and it barely registers in what matters to Canadian voters. There’s a hype bubble for you.
We then weighted the key issues by party and gave the top party a weighting of 3, and the bottom a rating of 1. These weightings were applied to the search volumes by region to give a prediction of the number of ridings per region.
At this point, a slight adjustment to our process had to be made that took into account the unique political and cultural context of Quebec within Canada. Important issues here are seen much more through a regional lens, rather than national. To address this, we repeated the same steps but only on the Quebec region. Ultimately this helped us generate a much more accurate picture of Bloc Québécois’ performance in the region
The final act (well done for sticking with us if you’ve made it this far), was to balance how much interest in the party leaders mattered against interest in the policy issues. In our previous modelling, 80% of people decide who they’re voting for based on the candidate, not the issues. We therefore applied this weighting to arrive at the final results.
Our Conclusion
And so, our final prediction: HUNG PARLIAMENT (again), with the Liberals needing to form a coalition with one of the smaller parties.
Final standings (170 is needed to form a majority):
Liberals – 166
Conservatives – 118
Bloc Québécois – 16
NDP – 27
Green Party – 9
It’s challenging to define the political influence that mobile devices, and the platforms they connect us to, have on individuals. What we do know is that there is a clear impact, and that impact isn’t recognised in modern polling methods.
However, this exercise isn’t about saying that political forecasting needs to shift entirely to the digital. Instead, we’re trying to demonstrate that it needs to play a bigger part than it does currently. If nothing else, please take that away with you.
Also, we’d like to be right.
