Did betting markets predict the 2020 outcome better than the polls?
Betting markets appeared to anticipate that Trump was underestimated in the polls. But a more detailed analysis of the odds suggests a persistent pro-Trump bias.
One hot take that has emerged from the 2020 election is that the betting markets were pretty good at predicting the election, especially compared to polls-based forecasters. Here’s an example from Betfair’s live blog:
Here is another example from Twitter:
The aggregated betting predictions from Election Betting Odds prior to the election do indeed pretty closely reflect the final map.
To assess the performance of betting markets, let’s compare them to FiveThirtyEight’s poll-based forecast, which predicted a Biden win with 90% probability.
In this piece I want to address three questions:
1. How good were betting markets at predicting the result in each state?
2. On election day, were Biden’s chances to win closer to 65% (as the betting markets suggested)? Or 90% (as FiveThirtyEight predicted)?
3. Why did betting markets differ from the polls?
My position is that FiveThirtyEight’s prediction was better than the betting markets. While betting markets were a little better in predicting the results of individual states, I will argue that they gave Trump too good a chance of winning overall. I will further argue that the reason betting markets differed from what the polls suggested was largely due to a persistent pro-Trump betting bias, rather than any particular polling insight.
Were betting markets better at predicting who would win each state?
We have already seen that betting markets favoured the winning candidate in all states except Nevada and Georgia. How does this compare to FiveThirtyEight’s predictions? Below is their snake chart, which orders the states by their likelihood to go for each candidate.
FiveThirtyEight did predict Georgia and Arizona correctly, unlike the betting markets. However, North Carolina and Florida were predicted incorrectly, so that looks like a 48–48 draw on correct calls (ignoring Maine and Nebraska’s congressional districts).
However, this kind of binary analysis is a bit misleading. We should note that states like Ohio, Iowa and Texas were predicted to be quite competitive by FiveThirtyEight (a Biden victory was predicted with 38–45% probability), but Trump appears to have won these states pretty handily by 6–8 points. While this lead could narrow a bit when all the votes are counted, the betting markets might deserve some credit for identifying that these states were quite unlikely to go blue in 2020.
A full analysis of how good each state prediction was is tricky (and not really feasible until all votes are counted), but the betting markets look pretty good at a glance on this metric — although not hugely better than FiveThirtyEight.
Was Biden 65% likely to win on election night? Or 90% likely to win?
If two forecasts make the same prediction, and that prediction is correct, but they offer different probabilities, whose projection was better?
Without more information it is ambiguous, but in this case, I think the 90% probability is more well justified.
Firstly, this was not a particularly close election. I know it FEELS like one, because of blue shifts/red mirages or whatever you want to call them, but the final results show a pretty convincing win for Biden.
Comparing to elections in the last two decades, this is obviously less close than 2000 and 2004 — both of which were decided by the result in a single state (Florida and Ohio respectively). With his (likely) 306 electoral college votes, Biden could have lost at least two of the swing states he carried and still won the election. It is also less close than 2016, where Trump won the decisive blue wall states of Pennsylvania, Michigan, and Wisconsin by less than one percent. While Biden’s margins in Wisconsin and Georgia are similarly tight, he could have lost both of those states and still won the electoral college (Biden is likely to win by around 2 points in Pennsylvania when all votes are counted).
If all votes had been counted on election night, this election would probably be viewed as only slightly less close than 2012, typically viewed as a decisive win for Obama. In other words, the election result is perfectly consistent with a 90% forecast.
On the other hand, this election was decidedly not a landslide; the result is perfectly consistent with the 65% forecast of betting markets too — intuitively perhaps more consistent. To make a proper evaluation, however, it is important to understand why the betting market predictions differ from the FiveThirtyEight forecast.
Why did FiveThirtyEight see Biden as a significant favourite?
Simple: the polls showed Biden had sizeable leads in a number of key states, which gave him many more viable paths to winning the electoral college than Trump.
“But the polls were wrong again!” Sure, but the whole reason FiveThirtyEight’s forecast had Biden as a 90% favourite was because even in the event of a sizeable polling error (such as that seen in 2016), Biden would still win enough states to win the presidency. While a Trump win was still a live possibility, this polling cushion gave Biden much better chances of winning the presidency than Clinton had in 2016 (when FiveThirtyEight gave her a 71% chance).
In other words, the main reason to be bullish on Biden was that his polling lead was robust to a large polling error in Trump’s favour, and this was precisely why Biden was a significant favourite in the race.
Why was the betting market more bullish on Trump than FiveThirtyEight?
This could have been due to many factors. betting markets are, after all, an aggregation of different people’s opinions, but broadly I see two possible explanations. The first is that bettors shrewdly deduced that the polls were very likely to be underestimating Trump. Perhaps they realised that likely Trump voters were being consistently under sampled, and as a result, the votes in key states were likely to come down to narrow margins.
This explanation is pretty implausible. Pollsters did change their methods after 2016 to try and avoid similar errors, most notably by weighting responses by education. It was anyone’s guess whether pollsters had over or under adjusted their methods, or whether they were missing something new about the 2020 electorate. I have no idea how the market could have been reasonably confident about a Trump polling miss.
The second possible explanation is that bettors were simply overconfident about Trump. In 2016, Trump massively defied our expectations of who wins elections. He not only won the Republican primary and the 2016 election, but he did so while spouting nonsense and lies that would have ended the career of many politicians in earlier years. Trump has an aura of being able to defy expectations, so perhaps people believed (and bet on) Trump doing it again.
If the second explanation is dominant, then the high-quality appearance of the betting market prediction is the result of a dumb luck guess. It is also plausible; a majority of Americans appeared to believe that Trump was going to win the election, despite most of them planning to vote for Biden.
But punters are not necessarily most Americans (most Americans cannot even place a bet on the election), and perhaps collectively, they did know something fishy was going on with the polls. I doubt it however, because the way the market moved on election night and in the following days suggests a consistent pro-Trump market bias as a more likely explanation.
Below is one betting market forecast for the week before the election was called. Notice how on election night, Trump was, for a few hours, seen as a 60–70% favourite to win the election. This was around when Trump was clearly going to win Florida.
What is odd about this moment is that the betting markets had already predicted that Florida would go red, so why did Trump’s chances go so high? Even as I watched on election night, in an anxious sweat as my Biden bets temporarily plummeted in value, I struggled to understand why Trump’s apparent win probability had increased so much. Florida was a must-win state for the president, but even if Florida’s polling error were replicated across the entire country (around 4.5 points), that would still have implied that Trump would probably just lose Pennsylvania, and with it the election. At best, I think it could be argued that the race was a tossup at that point, I see virtually no argument that Trump was priced correctly as a 2 in 3 favourite at that point. That suggests a Trump bias in the market.
If you are not convinced, one final argument. On Betfair, the 2020 presidential election betting has not closed at the time of writing, despite having been projected by every major news network. As of time of writing, punters give Biden only a 91% chance of winning, days after every major news network projected that he had already won! The idea that Trump has a 8–10% chance of changing the election through legal challenges is completely farcical.
It may be possible that the idiots still betting against Biden are very different from the punters earlier in the election, but to me the picture is clear. Betting markets have consistently been overconfident in Trump’s ability to win re-election at every point, and their apparent wisdom was little more than a statistical fluke. When it comes to predicting elections, I will continue to stick my hat in with the folks at FiveThirtyEight.
