The 2020 US election was the biggest betting market in Pinnacle’s history. While this is interesting for a number of reasons, it’s not necessarily surprising. The US election is arguably the one human centric event that has an impact on the greatest number of people on Earth. And this year’s incumbent is especially good at creating a headline. Add to that we’re in the middle of a pandemic, limiting many people’s choice of entertainment to something from their couch. It’s really no wonder betting interest was so high. As I write, a number of media organisations have called the election in Biden’s favour and bets have been settled by numerous bookmakers to this end. Betfair’s market remains open, having traded almost £600 million.
In the wash-up there’ll be plenty of good analysis looking at the efficiency and accuracy of the betting markets (including this piece from Joseph Buchdahl), and comparing their efficacy to both individual election polls and popular election models based on these polls. I won’t add to that analysis here, but I think it’s worth considering why one predictive platform may be more interesting and informative to follow, as well as more accurate over the long run. That’s what I’ll do in this article.
Both election prediction markets and election polls can be considered level two chaotic systems, meaning predictions about the outcome can influence the outcome. For example, hearing the outcome of a poll or the chances implied by a betting market may inspire a person to vote. Or persuade them not to vote at all.
If a voter favours the candidate ahead in the polls for example, their opinion may be reinforced (they are further encouraged to turn out and vote for them), or they may think that the candidate has an insurmountable ‘lead’ and will likely win anyway (and they don’t vote at all). I’m not sure which human bias wins out most often here - social proof or complacency - though I’d happily bet that polls with an obvious political leaning favour social proof winning out. Manufacture a positive poll for a particular candidate and change some voters’ minds. While there were rumours Mike Bloomberg was betting on himself in the race to be the Democratic candidate, for a similar effect, I just don’t see this as a widespread concern in betting markets.
A betting market is a relatively simple concept. Every participant is attempting to make money from a counterparty by correctly predicting the likelihood of an outcome. Barriers to entry are low, errors are punishable, and the reward for new information is high and universal. Currency. In other words, all participants have skin in the game. A monetary incentive exists for everyone to represent their predictions accurately. The same can’t be said for polls, which can be partisan, black box operations, where neither pollster nor respondent necessarily has an incentive to represent reality truthfully.
This may be especially true for poll respondents in today’s hostile and polarised political environment. It’s not difficult to imagine that a significant proportion of the population would be hesitant to respond or provide an honest answer when contacted by a pollster. This creates problems. Not only do pollsters need to correct for any sampling bias in the responses they receive, they also need to account for sampling bias in the responses they don’t receive. They need to ask: Is there a group with a certain political persuasion who are more likely to not respond? And this is surely an order of magnitude more difficult than extrapolating the responses they do receive.
Two of the most visible poll-based models for the 2020 election came from FiveThirtyEight and The Economist, who gave Biden an 89% and 97% chance of winning the electoral college votes respectively. The odds at Pinnacle and Betfair on election morning were closer to 1.5, implying around a 67% chance. That’s quite a discrepancy between the polls and the betting markets. How can this be explained?
An important thing to recognise is that FiveThirtyEight and The Economist are media companies. FiveThirtyEight is owned by The Walt Disney Company and The Economist is owned by The Economist Group. US elections can mean big money for media companies as they battle for political ad spend that sums into the billions of dollars in election years. Generating political discourse on their platforms is potentially very lucrative for these such organisations. And what better way to generate engagement than to commission seemingly sophisticated models that produce predictions that no one is likely to agree with nor understand. This is the media model in the age of social media. Polarise to engage and enrage. Unfortunately it works.
Polls and poll-based models don’t have any direct incentive to target or value accuracy. They are generally not rewarded or compensated based on the accuracy of their forecasts, which leaves them open to be politicised and/ or optimised as part of a media strategy, as is most certainly the case with FiveThirtyEight. The only election model I would give any credence to is one built by someone with the sole intention of making money via a betting market. If that’s you, fantastic. Good luck to you. But I can guarantee you won’t come across any profitable model in public. Be it in the media or on Twitter. And it’s for the same reason a successful bettor won't publish or sell you their tips. They'll bet them.
In his recent article analysing the efficiency of the 2020 US election Joseph Buchdahl outlined a potential dislocation between state betting markets and the next president market. The weight of money in the headline market likely didn’t allow this relative inefficiency to close, providing an opportunity for a savvy punter who knew where to look. As political betting markets garner more attention such opportunities may dwindle. Despite this, I will always take a signal from a market with thousands of participants - who may possess minimal information individually but nonetheless have skin in the game - before that of a media organisation whose motives aren’t clear, incentives aren’t known, and whose model is a black box. The wisdom of the crowd in election betting markets will trump polling going forward. The apparent accuracy of the state betting markets Joseph referred to is just one piece of evidence to this end.
To be clear, I’m not arguing that polls won’t outperform betting markets in any one election. They probably will. I’m positing that over the next hundred or thousand elections they most certainly won’t outperform betting markets on average. Not just that, by engaging with polls and the commentary that surrounds them, you are participating in a system that is at best trying to extract valuable information from you for free or entertain you, and at worst attempting to manipulate the way you think and vote. Follow betting markets however, and you will learn to think probabilistically, about statistics and markets, as well as observing the countless human biases on display. Win win. And you might even win some money.
The real game changer for election betting may come if and when social media companies such as Google and Facebook, whose data and algorithms I’m certain could predict election outcomes with far greater accuracy than both polls and prediction markets, decide to enter election betting markets. It’s possible they already have. In any case, when the next election rolls around, my advice is to consider following betting markets a little closer and give polls a miss.