Prediction markets are powerful, but they're bad at elections

The most prominent application for prediction markets is also its worst: election outcomes. As these markets have grown, it’s become harder to ignore that they consistently underperform non-market prediction methods.

The empirically observed inefficiency and underperformance of election betting markets can partly be attributed to tribal dumb money and participation limits for smart money. But more fundamentally, there is reason to think that markets are the wrong tool for predicting elections - that prediction markets will always underperform. That’s because, unlike in other markets, election betters have little useful private information to contribute to the market.

Prediction markets are a very promising tool - in other applications. But advocates of prediction markets are not putting their best foot forward when they emphasize election markets. To preserve the credibility of prediction markets, advocates should recognize that prediction markets perform better at some prediction problems than others.

The promise of prediction markets

Markets are a powerful tool. There is no better way than a market to determine how much of a good should be produced, who should produce it, and who should consume it. It’s very lucky that society so often finds optimal arrangements of production and consumption when individual actors follow locally rational behavior. The reason this works is explained by Hayek’s short and compellingly titled essay “The Use of Knowledge in Society”.

Markets aggregate the private information of every individual actor, without anyone needing to have all the information. No individual actor needs to know why the price changed. They need only to consider their personal circumstances when reacting to price changes. Their reaction - to consume or produce (or to buy or sell) more or less - in turn impacts prices for the whole market. The market is the interface through which every individual implicitly contributes their private information to produce efficient prices.

The same principles apply to markets for financial assets, where the different views and strategies are implicitly aggregated into one price. And, especially in financial markets, participants who contribute bad information - who often see the market move against them - are cleared out and their wealth transferred to those who contribute better information.

The promise of prediction markets is that the same market forces that set commodity and stock prices can assign a probability to an event occuring, by facilitating trading of a contract that pays out if the event occurs.

Just as commodity markets allow for cooperation between people who speak different languages and have different religions, prediction markets can synthesize the opinions of people with different analytics frameworks, worldviews, priors, and models into a single, definite, point estimate of an event’s probability.

But just as some markets are more efficient than others, prediction markets are better for some questions than others. And prediction markets have particular challenges with election outcomes.

Empirical inefficiency

Election betting markets frequently produce non-credible results. These markets consistently feature:

• Long-shot candidates projected to significantly out-perform polling
• Strange rank ordering of long-shot candidates, such as Michelle Obama sometimes given better odds than announced candidates
• Losing candidates taking an unreasonably long time to fall to 1% probability of winning

In the 2020 presidential election, as has been covered in Asterisk Magazine, PredictIt gave Mr. Biden only a 90% chance of winning Georgia, Michigan, Arizona, and Pennsylvania as late as December - even after those states had formally ratified Mr. Biden as the winner. The probability of those states flipping to President Trump was not 0%, but it was certainly much lower than 10%.

Also after the 2020 election, in PredictIt’s “electoral college margin of victory” market, most bettors who bet against Mr. Biden winning bet on President Trump winning 280+ electoral college votes. Even if President Trump flipped the result in every state where he had legal challenges, he would still not have reached 280 electors.

When a market’s price differs from your intuition, you should usually assume the market is incorporating information you don’t have. That’s what makes markets so efficient at pricing most things, and it’s why prediction markets have so much promise. But in this case, in hindsight, we can say that clearly PredictIt’s market gave nonsensical results.

And that was the election when prediction markets were supposed to do well. Every political observer in the US knew about PredictIt, and their moves were covered in the press. Billions of dollars were wagered over millions of daily trades. And yet they produced such noncredible results.

Again in the 2022 midterms, the betting markets underperformed. An analysis found that FiveThirtyEight’s polling aggregation model was more accurate than any betting market website across many House races.

Fundamental limitations

This pattern of inefficiency can partly be attributed to a solvable cause. Elections draw the interest of unsophisticated bettors who are biased by their tribal allegiances, who overestimate the likelihood of unlikely events, and who simply believe strange things. In theory, that creates an opportunity for smart money. But because PredictIt limits the amount that an individual can bet on a market, the smart money is overwhelmed. I expect that if that limitation is lifted, the markets will become more efficient.

But still, there is a fundamental limitation in how accurate prediction markets can be. This limitation is due to two aspects of election markets. First, transactions in purely speculative markets contain less information than transactions driven by real needs. Second, election outcomes is a question that can be addressed well by polling and statistical modeling.

Transactions in product markets contain real information about people’s willingness to consume. Transaction in financial asset markets contain real information about people’s risk tolerance and sectoral allocation. Even derivates markets are driven by real risk management demands. But in election markets, the information contained in a transaction is no different from the information in a sports bet: “this person thinks that they’re right, and they want to wager”.

A person’s willingness to wager on their belief may still contain useful information, but only in markets where there is significant unexpressed private information. That’s the case in commodity markets, where we can’t in practice continuously survey the preferences and plans of all producers and consumers. It works a lot better if you let people just transact and let the market find the price. But for many simple outcomes, like the point spread on a sports game or the outcome of an election, it’s possible to find the market clearing price without running a market - through statistical modeling.

In election outcomes, there is very little useful private information for the market to elicit. No one knows that the outcome will be. The polls are all public, and the models are too. (There are private campaign polls, but they tend to be less accurate than the public ones). You can do no better than polling people, adjusting the results with demographic models, and then aggregating the polls to simulate election outcomes. So in election markets, the average bettor is a fool and the marginal dollar contains negative information.

Addressing responses

The primary response to my claim that election markets have a fundamental limitation is that people who disagree with the market can bet against it, and over time the market price will incorporate the public statistical models, plus private information.

My claim is that polling and statistical modeling are so good at predicting election outcomes that, in comparison, there is very little value remaining in incorporating private information. The vast majority of bettors who believe they have private information will be wrong, and election markets will tend to be more wrong than statistical models. In order to counteract this effect, smart bettors who use statistical models would have to dominate the market. In that scenario, the market is providing no value and the smart money would be better off setting up their own betting site as bookmakers.

Another argument is that election markets allow people with financial stakes to election outcomes to hedge their risk. But election risks can already be hedged with OTC swaps, which are far less regulated. And the prices of those swaps are set by statistical models, not markets, because markets are not needed for price discovery for this type of product. The existence of swaps for hedging election risk is particularly devastating for election markets because it shows that you don’t need market-based price discovery for election hedging markets to exist. The social utility of election markets begins to look dubious.

Another argument for the social utility of prediction markets is that they will make horse race election coverage more objective and quantitative. This hasn’t been the case in sports, where sports betting has drawn the attention of commentators away from the substance of the game and towards coverage of exotic prop bets.

I can think of some scenarios where individuals may have useful private information. Maybe a candidate has a scandal that will soon become public. Anyone who knows about the scandal before it breaks can make money trading on it. Maybe forecasting the results in small, local races is aided by local knowledge. That hasn’t been the case so far, as FiveThirtyEight has outperformed PredictIt in house races, and I think there is a trend towards local races becoming more nationally correlated, but it may be the case one day.

It’s also the case that there are plenty of degrees of freedom in creating statistical models. Credible models can differ from each other significantly. Prediction markets are a way to synthesize the results of multiple competing models. But in practice price discovery will always be dominated by unsophisticated bettors.

What prediction markets are good at

I understand why people like election prediction markets, because I like them myself. The FiveThirtyEight model is just one model. The RealClearPolitics polling average is just one average. It’s nice to have one place where all the viewpoints can meet and produce a single price. It’s the most buzzy application for prediction markets. It’s frustrating that this isn’t working. But prediction markets, in other applications, are still a power idea.

Prediction markets are useful for quantifying non-quantitative problems. Election outcomes are so amenable to statistical approaches because they are statistical questions. But many important events have a significant non-quantitative element to them: who will be nominated for VP, who will be House speaker, will Ukraine join NATO, will a tax on unrealized gains become law, etc. Predictors can take rigorous approaches to answering them, but it is much harder to have confidence in any one prediction on these questions than in any one election outcome model. A prediction market is probably the best way for synthesizing heterogenous views into probability ballparks.

Prediction markets are also useful to incentivizing people to build models or to publicly express their model’s predictions. I like markets like “how many launches will SpaceX have this year”, “how many people will board a plane next month”, and “how many deliveries will Tesla make next quarter”. Some of these questions are implied in stock valuations, and it’s nice to disaggregate them.

I would like to see prediction market advocates focus on these applications.