How Twitter predicted Hillary Clinton’s defeat at the hands of Trump
The polls had her winning. Prediction models had her winning. Financial markets had her winning. But Twitter had Hillary Clinton losing, with a steady downtrend in sentiment that foreshadowed her stunning defeat to Donald Trump on Election Day.
The 2016 campaign featured more than 300 million tweets mentioning Clinton or Trump, with the Republican commanding a dominant two-to-one share of that conversation. Both candidates faced unprecedented unfavourable ratings among voters, but sentiment on Twitter gave Trump an edge. From his 16 June 2015, campaign announcement through Election Day, tweets mentioning Trump were 51% positive (+1). That compares to 51% negative (-1) for Clinton since her announcement on 12 April 2015, according to data from social media analytics firm Brandwatch provided exclusively to Bloomberg.
Clinton’s Twitter sentiment wasn’t just net-negative throughout the campaign. Amid relentless attacks from Trump and repeated blows from leaks of her staff’s e-mails and investigations into her own, it also became more and more negative. Clinton’s best day on Twitter in 2016 was 9 April, when she campaigned for the New York Democratic primary on home turf in Brooklyn, with 79% of mentions rated as positive (+29) by Brandwatch. Her worst was 9 January, with 75% of mentions negative (-25) in the run-up to the campaign’s first votes in the 1 February Iowa caucuses. But a 200-day moving average of those scores—smoothing the volatility of day-to-day swings—reveals a near-constant decline in sentiment toward Clinton.
Social media analysts call real-time Twitter data “the firehose” because it flows so thick and fast. Their challenge is to build algorithms that collect millions of tweets about a topic, and then code those tweets as positive or negative—even to the level of individual phrases—so a tweet like “Hillary Clinton is wonderful but Donald Trump is terrible” is rated as a positive mention for Clinton and a negative one for Trump.
Of the hundreds of millions of tweets mentioning either candidate’s name or a selection of candidate-specific hashtags like #ImWithHer or #TrumpTrain, about 20% passed through Brandwatch’s filter to be coded as positive or negative, while the rest are marked neutral. The trends in those positive and negative mentions are what corporate clients like Wal-Mart and Dell use to gauge the mood of their customers—and what can help reveal the mood of voters toward Clinton and Trump.
Clinton’s long-term, 200-day moving average never recovered after it first turned negative on 1 May. That crossover was driven largely by a string of down days in March and early April, when the eventual Democratic nominee suffered several primary and caucus defeats to Bernie Sanders—including surprise losses in Michigan and Wisconsin, states that would also turn against her in the general election. Even in the month of August, as Clinton soared in the polls following the Democratic National Convention, she enjoyed only three days of net-positive sentiment on Twitter.
Clinton and Trump’s online fortunes reflected their offline styles in the campaign. For Clinton, it was a steady march—although, on Twitter, in the wrong direction—while for Trump it was a volatile battle. Trump’s Twitter sentiment plummeted in March and early April as he suffered primary defeats in states like Utah and Wisconsin. Then in the rest of April and May, it skyrocketed as northeastern states like New York put him back on the path to the GOP nomination. Unlike Clinton, however, Trump’s long-term Twitter sentiment trended slightly upward throughout 2016.
To be sure, some of these tweets come from users that aren’t voters, aren’t Americans, or aren’t even humans at all. The proliferation of “bots,” automated accounts that blast out hundreds of pro- or anti-candidate tweets per day, has muddied the social media waters. According to researchers at the University of Southern California, bots could account for up to a fifth of the 2016 presidential election conversation on Twitter, and it disproportionately came from Trump’s side.
A study by Oxford University’s Project on Computational Propaganda found that pro-Trump bots out-tweeted pro-Clinton bots at a seven-to-one rate during the final presidential debate on 19 October. However, bot activity wasn’t enough to move the data in this analysis—Clinton scored +5 on 19 October, while Trump scored -7—potentially because those automated tweets are less likely to contain the kind of natural human language that sentiment algorithms like Brandwatch’s are looking for.
Overall, seventeen of the 19 fateful days between the final 19 October debate and the 8 November election were net-positive for Trump on Twitter. For Clinton, only five of those 19 days were net-positive, as last-minute revelations about the investigation into her private e-mail server buffeted the Democrat’s campaign.
Pollsters fear that their historic miss in the 2016 election was because some voters refused to admit their allegiance to Trump. For those voters, Twitter may have served as their confessional. Bloomberg