Markets Think the Polls are Accurate
Some more thoughts on 2024, and lessons learned from 2020
One Monday, I wrote about a few data points that seemed to be pointing towards polls of the 2024 presidential election being more accurate than they were in either 2016 or 2020. In those races, both national and state polling consistently underestimated Trump’s support, sometimes dramatically so. This fact has stimulated much debate over the years as to whether his idiosyncratic political appeal inherently produces turnout patterns that have proven impossible to detect with conventional polling methods, or if pollsters just got unlucky twice:
Today, I’d like to follow up on that discussion by adding another piece of evidence to the case. Ultimately, the stakes here are largely academic. In slightly more than a month, we’ll all know whether the polls were right or wrong, and by how much.
The first observation is that both prediction markets and oddsmakers have been behaving as if the polls are broadly correct, or at the very least, not systematically biased against Trump. This, in and of itself, is not necessarily a big deal. After all, something similar happened in 2016, and they were wrong, but it’s notable how much it contrasts with the behavior of those same markets in the lead up to the 2020 election.
In 2020, Biden led the FiveThirtyEight national polling average by 8.4% heading into election day, and the RealClearPolitics average by 7.2%. This was a very large lead indeed, and in the context of our recent history of unusually close elections, bordering on landslide territory. It points to a victory at or above the margin by which Barack Obama beat John McCain in 2008.
Biden did, of course, ultimately win the election, but people were much less certain of this eventuality than the polls might have suggested. A Trump victory was trading around the 35-40¢ range throughout most of the campaign on both Polymarket and PredictIt, and bookmakers like Bovada rated his odds similarly. Despite his massive polling lead, sophisticated observers with skin in the game thought that Biden was only around a 60-65% favorite to win the race. They were right, and the polls were wrong.
Let’s compare that to what the markets are saying today. As Rajiv Sethi wrote earlier this week, there has been some interesting variation in how different prediction models and market-based forecasts have been assessing the state of the race. Notably, polls-based models have exhibited much greater variance in their forecasts throughout the 2024 campaign than prediction markets, and the Silver Bulletin model has been considerably more bullish on Trump’s prospects than those published by either the Economist or FiveThirtyEight.
But in the current post-Labor Day iteration of the campaign, things seem to have settled into a more stable equilibrium. Trump’s chances of winning range from a high of 49% (Polymarket) to a low of 44% (FiveThirtyEight), with no significant divergence between polls-based forecasts, oddsmakers, and market-derived probabilities.1
Notably, this is exactly what we would expect to see if the election were being held today and the polls were accurate. Harris has leads in enough states to put her above 270 electoral votes, but many of those leads are narrow, particularly in the case of Pennsylvania, meaning that Trump could easily still win.
The bottom line, though, is that both prediction markets and oddsmakers are behaving as if they expect there to be no systematic polling bias against Trump in 2024, whereas in 2020 their behavior was exactly the opposite. Biden had a big lead in both the polling averages and forecasting models (FiveThirtyEight had him at 89% on election day), but the expectation of another polling error seemed to be priced into the betting and prediction markets.
In 2024, these sources have all basically converged on the same story: Harris is winning, but her lead is a narrow one, and within the margin of error.
Specifically, Trump’s actual or implied predicted probability of winning is 45% at Silver Bulletin, 46% at PredictIt, and 47% on RCP’s betting odds average. The Economist doesn’t publish a numerical forecast, but appears broadly in line with other models.