Previous attempts to predict elections using social media data have attempted to emulate traditional polling by predicting the share of votes received by major parties. However, in parliamentary elections, such as those held in Canada, the party who wins the most seats in parliament forms government (which may not be the party with the most votes nationally). In this paper, a method is presented which predicts seat counts using supervised learning with Twitter, polling, and historical election data. The model was trained on the 2015 Canadian federal election and was able to accurately predict the outcome of the 2019 Canadian federal election (a Liberal minority government, despite the Conservative Party winning the plurality of votes nationally). The model was designed before the 2019 election, and predictions were made public before election day. It is demonstrated that Twitter data about local candidates is more predictive than incumbency.
展开▼
