In the weeks leading up to November’s US presidential election, American political pollsters, by and large, predicted victory for vice-president Kamala Harris. But they were wrong. President-elect Donald Trump won the election comfortably. Polls have a margin of error of around three percentage points which account for variabilities in their sampling, but this is the third presidential election in a row that pollsters have miscalculated the extent of Trump’s support. Something, evidently, is going wrong. But what exactly?
First, it must be said that polling has never been an easy business or an exact science. Professional national polling was first conducted in the US by George Gallup – the godfather of opinion polling – during the 1936 presidential election. Millions of postcards were mailed out to voters, the returns tallied up. Gallup’s firm called the election for the Republican Alf Landon, but in the end Democrat incumbent Franklin Delano Roosevelt was re-elected with a thumping landslide. Gallup blamed the misjudgement on “participation bias”. Supporters of Landon, he reasoned, were just more enthusiastic about returning their postcards.
These days, polling methods are more sophisticated. Pollsters use data to ensure they sample a broad range of the electorate, whether it is by gender, age or ethnicity. Yet there are factors that may skew their sample of voters one way or another if there are problems with the methodology. A wealthy retiree, for instance, may have more time to respond to a pollster’s questions than someone who works three jobs to make ends meet. For this reason, pollsters may use novel methods to reach voters. If they are worried their methods are unlikely to reach young people, for example, they may poll them on social media rather than by telephone. Nevertheless, the pollsters’ methods – for the most part – underestimated Trump’s support in this election among certain voters. Most polls, for instance, failed to anticipate the massive 25 percent swing toward him among Latino voters. Similarly, in 2020, the polls did not predict a groundswell of support for Trump among black and Latino men. Moreover, in 2016, Trump’s first election victory, the polls did not account for the extent of his support among the white working-classes. The reasons for these errors are not necessarily clear and there may be many factors. But it is probably the case that these groups, on average poorer, are less likely to respond to pollsters.
Pollsters – as we have seen – can and do adjust their methods to find hard-to-reach voters. But in America’s increasingly fragmented media landscape, this is no easy task. Until the 1980s, Americans received their news and information from a handful of sources: radio, newspapers and three evening network news broadcasts. Today, there are a multiplicity of sources, such as social media and podcasts. Announcing the result on election night, a video made by one individual on TikTok, for instance, was viewed by as many people as one major television network’s entire election day broadcast. The new media makes it harder for pollsters to anticipate new trends in politics, whether a grievance against one candidate or a groundswell of support for another.
It should perhaps come as no surprise, then, that pollsters that used very traditional methods to reach voters proved most inaccurate. Selzer & Co, a well-established pollster, predicted Harris would win the state of Iowa by three percentage points. Trump won Iowa by a 13-point margin, a huge discrepancy. There may be many reasons for this. But the British pollster James Johnson – whose firm JL Partners correctly predicted the election result – blamed Selzer & Co’s overreliance on the traditional telephone call. Indeed JL Partners’ success may owe itself to the extra effort it made to reach Trump voters. For instance, the firm was concerned these voters – who are more likely to distrust traditional media and, for that matter, pollsters – were likely to ignore phone surveys so used surveys as in-app games with prizes for completing the poll.
Another factor is what is called “social desirability bias”. In British politics this is also known as the “shy Tory” phenomenon. People, according to this theory, are reluctant to share things they may be ashamed to admit publicly, even in anonymous surveys. For this reason, a French market trader named “Theò” – nicknamed the “Trump Whale” by the press – commissioned his own polling research using what is called a “neighbourhood model”: respondents were asked who their neighbour would likely vote for. According to this model, Trump was the most likely election winner. The trader bet millions on Trump on the cryptocurrency-based prediction market platform Polymarket and won handsomely.
His victory reflected the accuracy of the prediction markets more broadly. Betters on Polymarket, for instance, were more likely to plump for Trump, who on the eve of the election led Harris 58% to 42% on the platform. Again, there are many reasons for this. But financial markets in general do have a good track record predicting US election winners. Before 2024, the performance of the S&P 500 – the benchmark index synonymous with the US stock market – accurately predicted the US election winner 83% of the time. This accuracy can be partly explained by financial self-interest. A trader is unlikely to invest in a company that cannot sell its products in the same way that a punter at the bookies is unlikely to put their money on a lame horse. This in turn affects the share price of a company or the odds of the horse’s victory. The worse a company is performing and the more likely a horse is to lose, the greater the risk and the bigger the pay-out for the investor or punter willing to gamble on a stock or horse. For this reason, markets or odds at the bookies tend to self-correct. For instance, if investors regain confidence in a company, its share price will rise accordingly which is of course very good news for the gambler.
This is not quite the case in the US election. Most punters on Polymarket – the Trump Whale notwithstanding – had only the polls on which to make a judgement about who would win. This was effectively a confidence vote on the polls themselves. Unlike odds and stocks, polls do not self-correct as a reflection of investor confidence. They are predictions made on the basis of often fallible information. As we have learned, the polls underestimated the extent of Trump’s support in 2016 and 2020. It was a reasonable assumption, then, that the pollsters were underestimating the chances of a Trump victory again. So the smart money was to bet against them.
