Given the importance of the topic to this blog, I thought it best to create a discussion of prediction markets here that I can refer back to. In the process of researching this topic, I found some other resources of varying quality which I will be referring out to as necessary. The best analysis is Paul Sztorc’s Prediction Market Sequence, which is unfortunately kind of buried under a subsection of the Bitcoin Hivemind website, and is also a combined ~75 pages (!). I would definitely recommend it to anyone who wants more details, but if you only want to read 3000 words, this blog post will cover a lot of the main ideas.
What is a prediction market?
Prediction markets are exchanges where you can buy and sell “derivatives” (also referred to as shares, contracts, options, depending on context) on whether a given event will occur. They can be simple yes or no questions, or they can be more complicated, which we will cover later. The market price reflects what people believe to be the probability of the event occurring. A simple example would be a prediction market on whether Donald Trump will win the 2020 election. Suppose each contract resolves to $1. If you buy a “yes” option at 40% or 40¢ and then Trump wins, you’d be able to redeem it after the event at 100% or $1. If he loses, the share value would be 0% or $0. Other prediction markets can use any pricing scheme reflecting 0-100%.
PredictIt.org, a well known prediction market site, uses shares that resolve to $1 or $0 with trading occurring in between those prices until the event occurs. A 70¢ price would indicate a market belief of 70% in the event occurring. This is similar to betting markets on external events (like betting on the outcome of sports events) and also futures markets, where additional contracts can be created as long as there is interest in trading them, unlike stocks or bonds which are issued by an organization in limited quantities.
For more info, check out the first three sections of Paul Sztorc’s first PM paper.
What’s the point of a prediction market, is it just a way to gamble?
They can be used for gambling, but that is not their most interesting use case. Prediction markets aggregate information, and so they are sometimes called information markets. They offer a profit incentive for someone to share information publicly or for you to investigate something to try and uncover information. For example, in 2016, there wasn’t a great deal of presidential election polling in Wisconsin and other Midwest states. You might have been able to leverage information you found out about Trump’s support among Obama voters in those areas to place bets on Trump doing better in the electoral college. While talking heads on TV get paid for giving opinions, they don’t get paid for correct ones; prediction markets change that, allowing for a direct reward mechanism for correct predictions. Prediction markets thus improve the quality of predictions while simultaneously cutting through the self aggrandizing opinionators who have nothing on the line.
Aggregation of information is a concept so broadly applicable, it’s hard to convey the potential impact of its widespread use. At the very least, we would have a significant improvement in forecasting due to the profit incentive of those with useful information as well as those who can find out information and profit off its discovery. In the best case, we can aggregate all human knowledge into singular probabilistic forecasts about the future, providing real world application and feedback of empirical methods for everyone’s benefit.
If prediction markets are so great, why aren’t they being used already?
They are being used in limited ways, although they face many challenges. One is legal; the US has made it illegal to gamble on the internet. That means no large scale prediction markets can function in the US. PredictIt.org, a well known politics prediction market, has limitations on the amounts individuals can stake in a market as well as the number of individuals allowed in a single market. This limits the liquidity and financial interest in prediction markets. There is also the fact that knowledge about them isn’t widely dispersed; it’s a niche topic for obscure blog posts and academic papers.
Even if banned publicly, private organizations have an incentive to use prediction markets to improve their forecasting. Why haven’t more done so?
Prediction markets don’t necessarily mesh well with human organizations; managers may not be optimizing for the best information, but rather the decision making policies that make them look good. Dan Ariely writes how resistant companies are to conduct experiments that would actually give them better information; experimenting on customers is seen as immoral even if that results in a company never improving the products it provides for fear of trying new things. In addition, as Bryan Caplan points out, markets themselves aren’t very popular generally, and the Senate shut down a Defense Department experiment on prediction markets because Ron Wyden found it “grotesque”, despite the obvious national security incentives to learn as much information about security threats as possible. Paul Sztorc discusses similar challenges prediction markets face in his first paper, linked earlier. Prediction markets challenge the status quo and conventional wisdom and thus may never be adopted in organizations with established hierarchies.
I believe that in addition to the points Sztorc makes, prediction markets are simply expensive to set up. You need a market infrastructure (likely a digital platform, secure database architectures, etc) but also a mechanism for resolution, a decision structure for what markets will be created, and of course money to seed the market (If you’d like to learn more about market seeding, this blog post dives into Hanson’s logarithmic market scoring rule). You also need people to know to go to your market to make predictions; if no one is going to make predictions or buy shares on your market, other people won’t be incentivized to either. This is a network effect that may be hard to overcome.
So why are we talking about them if they can’t exist in real life?
One reason is to increase the awareness of the potential of prediction market to the readers of this blog. But the big reason I’m talking about prediction markets right now (as well as Paul Sztorc) is because of the rise of cryptocurrencies. Many of the social challenges to prediction markets can be overcome by instituting prediction markets on decentralized platforms and trading the contracts in cryptocurrencies. If you need more background on cryptocurrencies, I have some previous blog posts.
Prediction market projects on distributed blockchains, like Bitcoin Hivemind and the recently launched Augur, cannot be blocked by traditional legal prohibitions. They also don’t need permission from the established experts in a field in order to make predictions. They are, however, more expensive than prediction markets like PredictIt or Betfair; they may require additional money to transact and specific cryptocurrency expertise to enter, and they risk uncertainty in how dispute resolution will occur in this new space. Nonetheless, many of these risks and costs may decrease over time as this technology becomes more familiar. Their location on a blockchain of course also means they are not just censorship-resistant, but effectively immortal.
Furthermore, prediction markets can do much more than the simple binary option market we discussed before.
What else can prediction markets do?
Ok, it starts to get complicated here. I’m mostly pulling from Paul Sztorc’s second paper on prediction markets (Unlocking the Power of Prediction Markets). Prediction markets can also create scaled markets instead of simple Yes/No binary markets. This is like trading a stock. You can buy Apple at 100 and sell at 200. The only difference is that the prediction market ceases after the event occurs, so instead of buying Trump at 40¢, you could buy Trump at 250 electoral votes in August 2016 and your share would mature at 304, since that is the total number he received when the electoral votes were counted in December. On such a trade, you’d receive a profit for guessing that Trump would receive more than 250, and the total profit would relate to how much more the ultimate “price” or quantity beat the price you paid.
The more interesting addition is the use of multiple dimensions. For example, you could create a market with 4 outcomes trading across two dimensions: Unemployment up year over year and whether Trump adds tariffs to Chinese imports. The resulting percentages should yield a relationship in how the world views these events; it’d be likely additional tariffs would increase the expected unemployment, but the prediction market would tell us if the expected effect were certain or uncertain. If we used a scaled dimension, we could even see the magnitude.
This is pretty powerful. I highly recommend Sztorc’s paper on this, as even more complex markets could tell us more detailed info (and he has nice illustrations). However, we should note that more complex markets require more complex contracts embedded into the blockchain. Blockchains don’t scale well (so far), so complex contracts may result in expensive transaction fees, and the multiple bets you can place may mean overall liquidity is low. If liquidity is low, people may not feel like betting is worth their time, since large bets might not be possible, reducing the incentive to research hidden information.
What are some examples of prediction markets?
The simplest prediction markets are already in existence at places like Betfair and PredictIt. For those betting markets to be useful to us, those websites have to share our interest in betting on things. And while this is generally true for broad markets (like prediction markets for president or senate seats), there are very few places you can easily place a bet on something other than company or commodity stock prices, political races, or sports. Moreover, current prediction markets have various problems, like legality and limitations on bets.
So there are simple prediction markets, like binary options on whether an event will occur, that could be useful to those who are curious about the world. Examples might include current events, like North Korean nuclear tests before the end of next year. North Korea might even buy shares anonymously (especially if this is a cryptocurrency market) which could actually directly convey information from inside North Korea in a way we never have before, albeit by directly financing a dictatorship. We will talk more about the criminal uses of prediction markets later.
Paul Sztorc points out in “Extra-Predictive Applications of Prediction Markets” that prediction markets could offer an easier way to bet on company stocks. You wouldn’t need a brokerage account with a bank, just access to a cryptocurrency. Perhaps this is more difficult for less tech savvy users to get than a bank account, but perhaps not. Many foreigners or people without good access to banking might not be able to trade in stocks. This would allow them to have a portfolio. It’s Finance 101 that you should diversify your assets to save for the future and hedge against risk, yet many people without access to banking and financial markets have no way to diversify their assets. With a prediction market tied to the S&P 500 price, you could simply buy a contract and you’d immediately have an asset that tracks the general stock market. In fact, you could use something similar to track any publicly known stock or ETF.
This would also allow for insider trading similar to the North Korean nuclear weapons market. Interestingly, this is a strength of the market, causing any insider knowledge to be quickly dispersed to the outside world. That’s actually another fascinating application Sztorc discusses: whistleblowers.
Whistleblowers risk lawsuits, job loss, prison time, and their lives, and yet they are guaranteed nothing in return, even if successful. Can we do better?
Sztorc goes on to point out that whistleblowers could collect money for their knowledge, providing more of a safe haven for coming forward sooner.
There are other fascinating applications for public construction, which would allow public betting on whether a project would complete on time. This could be used to keep public officials who awarded the contracts accountable and called out directly when making corrupt deals. You could also make markets estimating what the expected level of depreciation are for various new models of cars to assist people in buying cars that will retain their value. Nominal GDP futures markets are an idea that has been discussed on this blog as a useful tool in monetary policy, providing feedback for setting interest rates.
Prediction markets don’t even have to be used just for information gathering, as Sztorc points out. They can be used as insurance as well. There could be a market for natural disasters which allows people to place bets on events that they don’t have any additional knowledge on, but simply want to be insured against poor outcomes. They can then collect the winnings of the bet as their insurance claim.
The possibilities for simple single event prediction markets are countless. But there are even more interesting examples when we get more complicated markets.
How can we get more complicated than North Korea using insider trading to make money off of their own weapons program through online cryptocurrency betting markets?
It’s time to explore multidimensional prediction markets and in particular conditional prediction markets. Robin Hanson has suggested conditional prediction markets for publicly traded companies. One dimension would be the company’s stock price and the other would be whether they fire their CEO. It would provide immediate input from the market on whether people believe the CEO is actually useful and adding value to a company. If the projected company value was similar to or higher if the CEO was fired, that would provide good justification for getting a new CEO.
I also like the idea of a conditional prediction market for candidate policies, conditional on them getting elected. This would provide feedback on whether anyone actually believed the candidates’ promises to deliver on various policies. One can imagine cultural norms arising where candidates had to provide proof that they had purchased significant amounts of shares that they would carry out the policy if elected. Political campaigns might then be more substantial, focusing on policies candidates had demonstrated they were actually committed to.
Robin Hanson has also suggested the idea of Futarchy, where prediction markets are created to predict the outcome of various policies, conditional on their implementation. He proposes the concept of “voting values, but betting beliefs”, where legislative bodies would vote on what we would want as an outcome of some policy, and then commit to implementing whatever policy is favored by prediction markets. It’s an interesting idea although it mostly moves the debate in legislatures from “what policy should be undertaken?” to “what policy indicator are we trying to maximize?”. That may be a better debate, but if poorly chosen, the indicator could maximize something to the detriment of the polity. It’s probably still worth investigating on a narrower scale, and certainly the existence of the markets themselves could only improve the information available to policymakers.
Paul Sztorc also discusses some complex applications of multidimensional prediction markets in the Bitcoin space. Unless you are really interested in Bitcoin, they probably won’t catch your eye, so I’ll just mention them briefly, as they are technically impressive. The first is a way to allow Bitcoin investors to gain more information about possible hard forks (changes in the network and blockchain) and insure against them occurring or not occurring. Hard forks are turbulent times in a blockchain, but current discussion is just theoretical arguments of what might happen. Network effects could mean people like the idea but don’t switch to a new fork, but a prediction market could give immediate feedback and allow people to purchase shares of post-fork-Bitcoin while insured against the possibility the fork does not go through.
Also discussed is the idea of “stable” coins which would be futures markets on the exchange rate between Bitcoin and the US dollar, as well as issuing stock through a prediction market. The most interesting to me was the idea of efficiently funding public goods. The basic concept is that you create a prediction market asking whether a public good will be provided (like a lighthouse) and then people who want the public good buy shares of “no”. They won’t lose any money unless someone builds the lighthouse. People intending to build the lighthouse can buy lots of shares of yes, and, when it’s built, the lighthouse builders would reap the benefits of the shares, and those who wanted the public good would pay for it, only if it existed. You can also create lots of different possible “yes” shares, based on publicly available information, like “what letter will be painted on the side of the lighthouse?”, which would allow different teams to compete. For the full write up, check out the paper.
Sztorc concludes with generalizing this idea to “smart contracts” based on prediction markets. This is built on top of the idea that prediction markets are a robust way to determine not only the future, but how the future eventually resolves (i.e. there is a mechanism that ultimately is used to determine what happened in the 2016 election to determine whose shares are paid).
What about insider trading you were talking about earlier? Can you use prediction markets to do other illegal things too?
Mostly likely yes. Just like Bitcoin and other cryptocurrencies, prediction markets can be used for illegal transactions since they aren’t tracked well by the state, and they can be used to bet on unsavory things or incentivize criminal behavior. Insider trading is one of those activities. However, I tend to buy the libertarian argument against the illegal prohibition of insider trading. Crimes require a victim, and insider trading provides information, while its prohibition does little to stem the flow of information to those who have little concern for the law anyway.
However, stranger markets exist as well. Sztorc devotes most of his final paper to discussing prediction markets in deaths, which could be manipulated by assassins. Sztorc is skeptical this could work, as any markets that believe someone is highly unlikely to die are automatically creating an incentive for an assassin and so the market would self regulate back to equilibrium where the predicted death is both more likely and less profitable. I’m still somewhat concerned. The bottom line is that death markets provide an avenue for assassins to reward themselves that does not exist today. Perhaps such incentives would be small compared to the difficulty and bodily harm risked by assassins, but it’s unquestionable that not everything about prediction markets is purely good. That’s expected though; information is powerful and does not always lead to to purely good outcomes.
Conclusion
The benefits and potential of prediction markets to improve the world is vast. As Sztorc states: “Their greatest benefit lies in their unlimited ability to scale”. Conversations are one to one, classes are at best a few lecturers and a few hundred in the audience, while prediction markets can take input from everyone in the entire globe and condense that into comparable and understandable probabilities. Being able to retrieve the world’s knowledge by doing something as simple as checking prices online is nothing short of a revolution in information. With the advent of decentralized blockchains, prediction markets are here to stay; we should embrace their power and potential.