Did social media call Brexit?

On 22 June, one day before Britons went to the polls to cast their votes in the EU Referendum, a press release arrived. Nothing unusual in that, but this one stood out. It claimed that an analysis of social media posts for ten of England’s most populous counties revealed they were keener to leave than to remain in the EU, and that Leave hashtags dominated across the UK with the exception of Scotland. The split, it claimed, was 54.6 per cent Leave and 45.4 per cent Remain.

With the official pollsters and the bookies indicating that Remain would sweep to victory, with 55 per cent of the votes cast in its favour, the release from social media analytics company Talkwalker was not picked up by a single news organisation.

Perhaps it is unsurprising that the analysis was ignored. With the exception of highlighting pollsters’ findings and exit polls, respected news outlets are loathe to make major predictions just as results are about to be released. And, more importantly, social media has got public voting patterns wrong in the past.

The results of last year’s General Election, which returned a small Conservative majority government, astounded many Labour supporters who were convinced that everybody on Facebook and Twitter shared their view. The realisation that they had been communicating in a giant echo chamber with like-minded friends and associates had come as a shock.

But Talkwalker was not alone in its analysis. Brighton-based Brandwatch had analysed more than 6.2 million social media mentions of the EU Referendum in the month preceding polling day and found #VoteLeave was mentioned more than 1.4 million times, compared to #Remain’s 800,000 mentions over the same period. However, the social media monitoring company had cautioned against reading too much into the results. People tend to tweet against issues more often than they tweet in favour, which can skew attempts to measure opinion, while using volume or frequency of tweets to gauge sentiment is fraught with difficulty, it warned.

Robert Glaesener, chief executive of Talkwalker, believes that, this time around, social media was telling the world, loud and clear, what the electorate was thinking. Several leading hedge funds agreed, reportedly building artificial intelligence, which gauges sentiment on social media, into their trading algorithms, and ignoring betting odds and public surveys. Each bucked the trend and made money in the aftermath of the result when financial markets initially plummeted, and their rivals lost billions.

Glaesener adds: ‘Throughout the whole campaign, what we were seeing was stronger opinions on Leave rather than Remain. It was striking and we were quite concerned. It was only in the last week before the Referendum, when people began to notice that this really was going to be close, that Remain people started getting onto Twitter and Facebook. But by then, it was too little too late. A lot of work had been done for the Leave campaign.’

This time around, social media was telling the world, loud and clear, what the electorate was thinking

(The late arrival of the Remain campaigners was a trend also noticed by Twitter, which claimed that the first and only time that the Remain camp led the conversation over the six week period was on polling day itself.)

Glaesener believes that social media was a better indicator of the final result this time around because ‘the Referendum was much less structured than normal elections’, adding: ‘Within the Conservative party, you had two different opinions. While the Labour Party was officially more for Remain, it did not advocate this position as much as would have been expected. There was no official party machine pushing out messages from spin doctors.

‘People looked to get opinions from other things. Social media could expand much more than before and influence opinion more than it used to. The less you have structure, like coherent political party strategies to win votes, the more social media will automatically fill that void.’ When Talkwalker analysed the chatter further, it found that the two main topics being discussed in relation to the EU were the economy and immigration, although the former greatly exceeded the latter.

But rather than entering into discussions on the topics, people were instead setting out their views.  ‘How can you make a counter-argument about immigration with an economic dialogue?’ asks Glaesener. ‘You can try to fight argument a with point b, but a and b were not really related, so on social media people didn’t get into a dialogue.’

Richard Sunley, Talkwalker’s content marketing manager, adds: ‘We focused on the Leave and Remain hashtags, and one thing we noticed was how quickly these would shift and react to news. When The Sun came out in favour of Leave, on Merseyside, where there is a negative view of the newspaper [following its comments after the Hillsborough tragedy], Remain hashtags quickly picked up.

‘As a general rule, after each televised debate, particularly those involving Michael Gove, or Question Time, there was a high spike in discussions but the spike in the Leave hashtag was always higher than in the Remain one. It was interesting to see.’

This resonates with claims by the Leave campaign that whenever public figures, such as US President Barack Obama or billionaire Sir Richard Branson, urged the electorate to Remain part of the EU, they added supporters to the Brexit cause.

Since the results of the EU Referendum were announced, several reports have been published suggesting that Twitter was right all along. The Kantar UK Political Pulse tracked tweets from the date the campaign started on 15 April until polling day.

It found that more than 9.3 million politically themed tweets were sent, with the hashtag #EUref used more than 236,000 times, picking up pace as 23 June approached, but the second most popular hashtag was #VoteLeave, used in 199,000 tweets.

As a general rule, after each televised debate, there was a high spike in discussions but the spike in the Leave hashtag was always higher than in the Remain one

But Claire Davies, editor and head of content at Kantar Media, is less keen to suggest that Twitter predicted the outcome, and suggests the popularity of #VoteLeave might be explained another way. It seems the most prolific tweeter of the entire campaign was  @Vote_leave, with more than 73,800 followers, which sent 108,000 tweets since April.

If each of these included #VoteLeave, it might explain the hashtag’s popularity. Similarly, the Remain campaign used a variety of hashtags, including #StrongerIn, #Remain, #VoteRemain and #LabourInForBritain. In total, these hashtags were used 246,000 times.

‘You can’t compare this with a general election. It is very different,’ explains Davies. ‘But it is certainly true that, at first, the EU Referendum did not capture the public’s imagination; they preferred initially to talk about the junior doctors’ strike and other issues that seemed to more directly affect them.

‘But as it got closer to the time, people started to get more and more engaged on social media. There were an awful lot of retweets, rather than people sharing their own thoughts.’

Andrew Williams, senior director at FTI Consulting, adds: ‘With the exclusion of Jeremy Corbyn, left wing politicians failed to penetrate Twitter conversations [in the four days prior to the EU Referendum]. But his handle @JeremyCorbyn was mentioned in just 14,725 tweets, significantly less than the handles of Boris Johnson, David Cameron and Nigel Farage.’

However, just like Twitter, Kantar Media found that, on the polling day itself, the social chatter was dominated by Remain posts. ‘If you looked at a Word Cloud on that day, you would have believed that Remain would succeed. But if you looked back over the entire campaign, then you would have thought Leave had it,’ says Davies.

But Hugo Zaragoza, chief executive of Barcelona-based Websays, a company dedicated to analysing online conversations and opinions, disagrees. At 4.30pm on polling day, the SENSEI project, with which he is associated, announced that the result would be 48 per cent Remain, 52 per cent Leave. The final result was 48.11 per cent to 51.89 per cent.

‘We actually saw mostly Leave comments, but that the impact of the Remain comments was stronger [in the run up to the Referendum]. We actually thought Remain was winning, it was not until two days before the election that we saw a switch but it was still so tight [one or two per cent] that we thought Who knows? It’s within a margin of error,’ he explains.

Indeed, as the polling booths opened across the UK, Websays’ analysis was pointing to a Remain win. Social media chatter was indicating 49.63 per cent in factor of Leave and 50.37 per cent for Remain, with no undecideds involved in the chatter. But as the undecideds arrived to vote, there was a dramatic swing to Leave.

The SENSEI Project, which is backed by the European Commission, is developing a natural language processing and analytics system to review online comments. It monitors hundreds of sources, including Twitter, Reddit, Google + and Instagram, along with dozens of news sources, including the BBC, The Guardian, The Daily Mail and other newspapers. It also monitors coverage beyond the UK and Europe websites, in countries such as the US, Israel, Japan, Taiwan and New Zealand.

At the start of the campaign, SENSEI’s geographical ‘radar chart’ revealed that the dominant UK sentiment, based on 1.5 million posts, was indignation, followed by amusement. At that stage, its prediction was that Remain would achieve 53 per cent of the vote. The indignation was attributed to the misleading claims and dubious statistics put forward by both sides.

Over the course of the campaign, the SENSEI project analysed more than six million social media conversations relating to the Brexit vote. But where SENSEI is different, according to Zaragoza, is that it considers the impact of these rather than just the chatter. ‘You can look at social media in a very naïve way. If you take a sample of 1,000 tweets, then it did seem that these were dominated by Leave. But doing so grossly over-estimated the Leave campaign,’ he says.

‘We collect all sorts of data, such as comments on newspapers online, and try to immediately classify it as a ‘Leave’ or ‘Stay’. We use natural language analytics to decide what the exact message is. We then look at the relative impact of the data: a retweet doesn’t count as much as a newspaper article, for example.

‘All these parameters are fed into our algorithm, and they are optimised by machines. At times the social media chatter was intense and passionate. Since we started this project, we have listened to and analysed 300,000 social media conversations on Brexit every day.’

The methodology, according to Zaragoza, draws from the marketing world, where it is standard to analyse how a particular message is landing on social media, but before now had not been used for political analysis. Since 2012, Websays has been ‘approaching elections like a marketing problem’ to assess whether social media analysis is just as good or even better than official polls. ‘Polls are very expensive procedures, which rely on small samples,’ he says. ‘And relatively cheap social analytics are a little biased and certainly not an oracle.’

Zaragoza has analysed elections in Spain, Italy and South America. ‘We don’t always get it right,’ he admits. But Websays’ analysis of the Spanish election was more accurate than official polls which ‘gave the third party much more votes than it got’ whereas the social analytics group found that Brexit discussions were impacting the party’s support.

‘Echo chambers were very important,’ he explains. ‘These changed perspectives. We were listening to all the conversations to get a realistic view.’

The prevailing view in Spain was that older people did not use social media and so this channel was mainly ignored by the political parties seeking to influence that demographic, but Websays found older people were engaging, particularly on Facebook. But, in general, Zaragoza concedes that national elections, which are historically fought by many parties and involve different political agenda, are difficult to predict solely using social media analysis.

The EU Referendum was less complex because it was ‘a binary decision’, but nonetheless the traditional pollsters once again got it wrong.

‘For years, people have blindly trusted the polls but the game is balancing,’ he says. There is still a role for polls but with much of the debate now conducted online, on social media, in public forums and comment boards on national media, there is also increasingly a role for social media analytics.

‘It should be a combination of the two,’ concludes Zaragoza.

This article first appeared in issue 107