As the world dissects the how and why of Donald Trump’s presidential victory, pollsters are already coming under severe criticism for getting it so wrong. But a tool we created that analysed emotions towards the two candidates on Twitter accurately predicted the result. Unlike other poll predictors that just take a snapshot in time, our tool used real-time data from Twitter which provides a more accurate picture of who people are likely to vote for.
We first developed the tool, called EMOTIVE, in 2013 to predict whether another riot could happen in London following those in 2011. It can analyse thousands of tweets a second to extract from each tweet a direct expression of one of eight basic emotions: anger, disgust, fear, happiness, sadness, surprise, shame and confusion.
We decided to use our tool to track emotions on Twitter around the world about the US presidential election for the three weeks leading up to the vote. During that period, our data showed that Trump continued to lead the race to the White House. Going by this tool, there were very few periods of time when Clinton was leading – contrary to what was being predicted by the polls.
The EMOTIVE system worked by looking at how much fluctuation there was in the number of tweets relating specific emotions towards either Trump and Clinton. The more tweets with reference to an emotion fluctuate – reflecting greater uncertainty towards a candidate – the fewer votes the model predicts a candidate is going to get. Crucially, it’s not the volume of tweets that matter, but the “choppiness” that make the difference. This means that the model does not concern itself with the direction of the emotion or put value on one emotion over another – because an emotion such as anger can be positive or negative.
This first graph shows how emotive tweets towards Trump fluctuated in that final 24-hour period. The y axis here measures what we call an emotional score. The wider the band of colour, the greater the number of emotive tweets related to that particular emotion, which means a greater outpouring of emotion towards Trump. But there was not a large amount of fluctuation overall in how much each emotion was appearing in tweets – meaning more voters were feeling more certain about their support for Trump.
Now look at the second graph, showing the same data in relation to Hillary Clinton over a slightly longer period. This shows greater fluctuation – more spikes and dips – and meant she was predicted to win fewer votes.
A good gauge of the public mood
When we analysed the raw tweets we saw an unprecedented campaign with a race all about personalities, dominated by attacks on Trump and Clinton.
There was very little mention of policy in the tweets we analysed, nor the difference it could make to the US and the world. This campaign saw the working class find their voice via digital communication channels.
This is not the first time the EMOTIVE tool has been used to predict an election result that the pollsters got wrong. In 2015, the system correctly predicted the outcome of the UK general election. And it has been put to other uses too. Another study looked at reactions to the November 2015 Paris terrorist attacks to potentially help survey those who would be at risk of post-traumatic stress.
Being able to monitor millions of digital sensors via Twitter and use them to map emotion provides a way to gauge the mood music of the average person on the street. And it is possible to get a real-life insight into what people really think about two candidates in an election.
In the case of the 2016 presidential elections, this is possibly why the polls were wrong throughout – often they were seen as the establishment trying to interfere in swaying opinion. While polls only provide a snapshot of that moment it time, systems such as EMOTIVE provide ongoing real-time analysis, which tends to provide a much more accurate picture of the landscape – and is proving good way at predicting which way elections may go.