In June 2020, the New Statesman’s data unit and our international team decided to produce a forecast model for the US presidential election, informing readers of the likelihood of a Joe Biden or Donald Trump win.
Two months in the making, the model was revealed in late August, and on election day gave Joe Biden a 90.3 per cent chance of winning the presidency. Had that model been employed in 2016, the numbers would have put Hillary Clinton at a 70.7 per cent chance, and Donald Trump at a 28.9 per cent chance.
With almost all the state results in – and Biden on course to win Georgia, while Trump takes North Carolina – we can begin to assess how well the model has performed.
Early figures suggest that, in the case of the electoral college, our model did well – and was the second most accurate of all popular forecasters. On the day of the election we projected Biden would win 339 electoral college votes. As things stand, Biden is on course to win 306, 33 short of our model’s projection.
Had Biden won Florida, a state with 29 electoral college votes, our model would have been the closest of all popular models – and only four votes out.
Our forecast vote shares were substantially out in a number of states, and less so in others. We were right to give Trump the edge in Ohio and Iowa, and to give Biden the edge in Arizona. However, we were wrong to give Biden a better shot in North Carolina than in Georgia. Our model suggested Trump had a one in three chance at winning Florida, and win it he did.
It appears one of the main problems was that US pollsters were still unable to properly sample voters more likely to favour Trump. That said, it is worth emphasising the fact that polling in the last decade has become more accurate, not less. This year’s result is well within the margin of error of what polls and models had predicted, and both were right about one thing: the election of Joe Biden.