Schelling Points, Prediction Markets, and Consensus

Prior Readings:

• Nash Equilibriums
• Focal Points
• SchellingCoin

A Schelling point relates to two people trying to make the same decision, and using special or ‘meta’ knowledge about the problem to arrive at a decision. As an example, two strangers are told that they must meet each other in New York City tomorrow, but are not given contact information or any clues about where they should meet. In real world experiments, people commonly chose to look for the other person at the Grand Central Station information booth. The information booth is a Schelling point, as it is a common and very widely known meeting place.

In March 2014 Vitalik Buterin presented the idea of using schelling points to inform a blockchain about external events such as the price of a Bitcoin. This information is inherently unavailable, and must be acquired through some sort of portal. One way to create this portal is to set up a game where all participants must pick a price, and they are incentivized to pick the same price. There is supposed to be a schelling point around the actual price of the coin — the truth is an obvious default, especially in a system where the goal is to produce the truth.

This unfortunately starts to break down as you get into the specifics. Schelling points are inherently subjective, and there must be strong evidence for all participating parties of an obvious, coherent choice. For example, if you are trying to reveal the price of a bitcoin on March 1st, are you choosing the price as it was at midnight? At noon? What timezone? Some people might choose midnight UTC as their schelling point, while others may choose 10am EST because that’s when the NYC Stock Exchange opens. And, some people might choose the Coinbase price, while others may average all of the exchanges. It gets even trickier if exchanges start exhibiting erratic behavior, are showing different prices to different people, or are otherwise suspected of foul play.

The other examples of schelling points also end up having issues. In the case of finding someone in New York City, there are multiple ‘obvious’ potential places to meet. There’s the Empire State Building, there’s the Statue of Liberty, and in games where people actually had to find eachother, the participants often required multiple days of trying to successfully meet.

The example given in Buterin’s article also has problems.

Suppose you and another prisoner are kept in separate rooms, and the guards give you two identical pieces of paper with a few numbers on them. If both of you choose the same number, then you will be released; otherwise, because human rights are not particularly relevant in the land of game theory, you will be thrown in solitary confinement for the rest of your lives. The numbers are as follows:

14237 59049 76241 81259 90215 100000 132156 157604

For me, the answer appeared obvious without even reading the numbers. I chose the first number, because it seemed obvious to me that with no context, you just pick the first number and hope the other person takes the same strategy. My cofounder thought it would be a good idea to pick the lowest number (in this case, it is the same as the first number, but only because the numbers are already sorted). To Buterin, 100000 appeared as the correct choice because it stands out as special where all of the other numbers appear random. Out of 3 people, there were three different methodologies used to decide the schelling points. As you start to reach across cultural boundaries, the methodologies may start to vary even further.

A classic prediction market has people bidding on the outcome of a future event, such as the price of a bitcoin or the outcome of an election. People buy and sell shares of each potential outcome, and in the end the shares of the actual outcome are rewarded financially, while all other shares become worthless. As an example, you might have people trading ‘Donald Trump Wins 2016 US Presidential Election’ shares vs. ‘Donald Trump Does Not Win 2016 US Presidential Election’ shares. At the end of the election, all shares of the winning outcome can be traded for $1, and all shares of the losing outcome are worth nothing. Critical to the voting is the ability to tell what happened, so that the shares can be fairly cashed out.

There are 3 major pitfalls with reaching consensus around the outcome of an event.

• The event itself is not fully specified. For example, ‘The Price of a Bitcoin at Midnight, Jan 1st, 2016’ fails to specify a timezone, and it fails to specify which exchanges should be used.
• The event is vulnerable to manipulation. For example, the price might be driven way up by someone who stands to gain financially from the price being misrepresented. Or, an exchange could misrepresent the price for the purpose of manipulating the results.
• The event is difficult to verify publicly. For example, ‘Policemen in New York will pull over 400,000 cars in 2015’ is difficult to verify, though technically all instances of policemen pulling over cars are documented .

A good candidate event must be fully specified, difficult to manipulate, and easy to verify publicly. A good example of such an event is the ‘The winner of the 2016 United States presidential election’. The outcome of the election will eventually be known, and the winner will be obvious to everyone. It is extremely unlikely that two people acting in good faith will provide different answers for the winner of the election. The election is also difficult to manipulate, there are massive amounts of protections in place to make sure that the US election is fair. Finally, the results of the election will be broadcast all over the world from many different sources, and can be verified in many ways. A massive percent of the population will be able to confidently state the result after little effort.

These fully specified, manipulation resistant, and publicly verifiable events form a necessary foundation for sound prediction markets. Without this foundation, prediction markets are subject to confusion, manipulation, and abuse. Though sometimes tricky, many prediction events exist that satisfy all of the criteria listed above. As the world moves forward into the realm of decentralized prediction markets, it will be important to keep in mind the pitfalls associated with many naive prediction events.