Forecasting and Understanding the 2021 Canadian Federal Election Using Twitter Conversations

Abstract

Previous studies using Twitter data to predict election outcomes have focused on sentiment and/or volume of tweets as predictive features. However, it is difficult to determine whether a tweet is expressing views in favour or counter to a political entity. By contrast, in this study deep learning is used to cluster related tweets together into “conversations”, and then those conversations are classified. Classification is achieved by identifying who is involved in the conversation: if a majority of partisans of one political party are present, then the conversation may be classified as being in favour of that party. The conversation is then predicted to change public opinion, relative to size of the conversation. This method has the added explanatory benefit of linking moments of large public opinion change directly to the source conversation, answering “why” a change happened. Time-series public polling data from the 2021 Canadian federal election is used to validate the model, and the final result of the election was forecast with the model the day before the election. It is observed that changes in public opinion seen in polling data are seen earlier in time in Twitter “conversations”