From Prediction Markets to Information Finance Why Polymarket Excites Me

Original title: From prediction markets to info finance

Author: Vitalik, founder of Ethereum; Compiler: 0xjs@黄金财经

Abstract: Election prediction is just the first application. The broader concept is that you can use finance as a way to coordinate incentives in order to provide valuable information to the audience.

One of the Ethereum applications that excites me the most is the prediction market. In 2014, I wrote an article about futarchy, a prediction-based governance model conceived by Robin Hanson. As early as 2015, I was an active user and supporter of Augur (see, my name is in the Wikipedia article). I made $58,000 betting on the 2020 election. This year, I have been a close supporter and follower of Polymarket.

For many people, prediction markets are about betting on elections, and betting on elections is about gambling — if it’s fun for people, that’s great, but it’s fundamentally no more fun than buying random tokens on pump.fun. Viewed from this perspective, my interest in prediction markets can seem confusing. So in this post, I aim to explain why the concept excites me. In short, I believe that (i) even existing prediction markets are a very useful tool for the world, but also that (ii) prediction markets are just one example of a much larger, very powerful category that has the potential to create better implementations of social media, science, journalism, governance, and other fields. I’m going to call this category “info finance.”

The Two Sides of Polymarket: A Betting Site for Participants, a News Site for Everyone Else

Over the past week, Polymarket has been a very effective source of information about the US election. Not only did Polymarket predict a 60/40 chance of Trump winning (while other sources predicted 50/50, which in itself wasn’t too impressive), it also demonstrated something else: When the results came in, while many pundits and news sources had been baiting viewers into hearing news in Harris’ favor, Polymarket just laid it out: Trump had a better than 95% chance of winning, and a better than 90% chance of seizing control of all branches of government.

Both screenshots were taken at 3:40 AM EST on November 6

But to me, this isn’t even the best example of how interesting Polymarkets is. So let’s look at another example: the elections in Venezuela in July. The day after the election, I remember seeing out of the corner of my eye some people protesting the highly rigged election results in Venezuela. At first, I didn’t think much of it. I knew Maduro was already one of those “basically a dictator” figures, so I thought, of course he’s going to falsify every election result to keep himself in power, of course there will be protests, of course the protests will fail — as many others have, unfortunately. But then I was scrolling on Polymarket and saw this:

People were willing to put in over a hundred thousand dollars betting on a 23% chance that Maduro would be overthrown in this election. Now I was paying attention.

Of course, we know the unfortunate outcome of this scenario. Eventually, Maduro did stay in power. However, the markets made me realize that this time, the attempt to overthrow Maduro was serious. The protests were massive, and the opposition had a surprisingly well-executed strategy that proved to the world how fraudulent the election was. If I hadn’t received the initial signal from Polymarket that “this time, there’s something to watch,” I wouldn’t have even started paying attention.

You should never trust the Polymarket betting charts completely: if everyone believed the betting charts, then anyone with money could manipulate the betting charts, and no one would dare bet against them. On the other hand, trusting the news completely is also a bad idea. The news has a sensationalist motive, exaggerating the consequences of anything for the sake of clicks. Sometimes, this is justified, and sometimes not. If you see a sensational article, but then you go to the markets and find that the probability of the related event has not changed at all, then it’s also justified to be skeptical. Alternatively, if you see unexpectedly high or low probabilities in the markets, or unexpected sudden changes, that’s a signal for you to read through the news and see what caused it. Conclusion: You can get more information by reading the news and the betting charts than by reading either one alone.

Let’s review what’s going on here. If you’re a bettor, then you’re betting on Polymarket, and to you, it’s a betting site. If you’re not a bettor, then you’re reading betting charts, and to you, it’s a news site. You should never trust betting charts completely, but I’ve personally made reading betting charts a step in my information gathering workflow (alongside traditional and social media) and it’s helped me get more information more efficiently.

Information Finance The Wider Implications

Now, we’re getting to the important part: predicting election results is just the first application. The broader concept is that you can use finance as a way to align incentives in order to provide valuable information to your audience. Now, a natural reaction would be: isn’t all of finance fundamentally about information? Different participants will make different buy and sell decisions because they have different views of what will happen in the future (besides personal needs, like risk appetite and desire to hedge), and you can infer a lot about the world by reading market prices.

To me, information finance is just that, but structurally correct. Similar to the concept of structural correctness in software engineering, information finance is a discipline that requires you to (i) start with a fact you want to know, and then (ii) deliberately design a market to extract that information from market participants in the best way possible.

Information finance is a three-sided market: bettors make predictions, and readers read the predictions. The market outputs predictions about the future as a public good (because that is what it was designed to do).

Prediction markets are an example: you want to know a specific fact that will happen in the future, so you set up a market for people to bet on that fact. Another example is a decision market: you want to know whether decision A or decision B will produce a better outcome based on some metric M. To accomplish this, you set up a conditional market: you ask people to bet on (i) which decision will be chosen, (ii) the value of M if decision A is chosen, and zero otherwise, and (iii) the value of M if decision B is chosen, and zero otherwise. With these three variables, you can determine whether the market thinks decision A or decision B is more favorable for the value of M.

One technology I expect will drive information finance over the next decade is AI (both large models and future technologies). This is because many of the most interesting applications of information finance are related to "micro" problems: millions of small markets where decisions, taken individually, have relatively small effects. In reality, markets with low volumes often don’t work efficiently: it doesn’t make sense for experienced participants to spend time doing detailed analysis just to make a few hundred dollars in profit, and many even argue that such markets can’t work at all without subsidies, because there aren’t enough naive traders on all but the largest and most sensational problems for experienced traders to profit from. AI completely changes this equation, meaning that even on markets with $10 in volume, it’s possible to get fairly high-quality information. Even if subsidies are needed, the amount of subsidy per problem becomes very affordable.

Information Finance Requires Distilled Human

Judgment

Suppose you have a trustworthy human judgment mechanism, and that mechanism has the legitimacy of an entire community trusting it, but it takes a long time and is costly to make judgments. However, you want to have low-cost, real-time access to at least an approximate copy of that “expensive mechanism”. Here’s an idea of ​​what you could do, proposed by Robin Hanson: Every time you need to make a decision, you set up a prediction market on what the expensive mechanism would do if the decision were called. You let the prediction market run, and invest a small amount of money to subsidize the market maker.

99.99% of the time, you don’t actually call the expensive mechanism: maybe you “reverse the trade” and give everyone back their input, or you just give everyone zero, or you see if the average price is closer to 0 or 1 and treat that as ground truth. 0.01% of the time — maybe randomly, maybe for the most heavily traded markets, maybe a combination of both — you actually run the expensive mechanism, and compensate participants accordingly.

This gives you a trusted, neutral, fast, and cheap “distilled version” of your original highly trusted but extremely expensive mechanism (to use the word “distilled” in LLM). Over time, this distilled mechanism roughly mirrors the behavior of the original mechanism — because only participants who helped achieve that outcome make money, and everyone else loses money.

A possible prediction market + community note combination model.

This applies not only to social media, but also to DAOs. A major problem with DAOs is that there are too many decisions and most people are unwilling to participate in them, which leads to either extensive use of delegation, with the risk of centralization and principal-agent failure common in representative democracy, or vulnerability to attacks. If actual voting rarely occurs in a DAO, and most things are determined by prediction markets, with humans and AI combining to predict voting results, then such a DAO may work well.

As we saw with the example of decision markets, information finance contains many potential paths to solving important problems in decentralized governance, and the key lies in the balance between market and non-market: the market is the "engine" and some other non-financialized trust mechanism is the "steering wheel".

Other Use Cases of Information Finance

Personal tokens - projects such as Bitclout (now deso), friend.tech, and many others that create tokens for everyone and make them easy to speculate on - are a category I call "raw information finance". They deliberately create market prices for specific variables (i.e., expectations about a person's future reputation), but the exact information revealed by the price is too vague and subject to reflexivity and bubble dynamics. It is possible to create improved versions of such protocols and solve important problems such as talent discovery by more carefully considering the economic design of the token (especially where its ultimate value comes from). Robin Hanson's idea of ​​reputation futures is a possible end state here.

Advertising - The ultimate "expensive but trustworthy signal" is whether you will buy a product. Information finance based on this signal can be used to help people decide what to buy.

Scientific peer review - There has been a "reproduction crisis" in science, whereby some famous results have become part of folk wisdom in some contexts, but ultimately fail to be reproduced in new research. We could try to identify results that need to be rechecked through prediction markets. Such markets would also give readers a quick estimate of how much they should trust any particular result before it is rechecked. Experiments with this idea have been done, and so far appear to be successful.

Public goods funding - One of the main problems with the public goods funding mechanism used by Ethereum is its "popularity contest" nature. Each contributor needs to run their own marketing campaign on social media to gain recognition, and contributors who are not able to do this or who naturally have more "background" roles have a hard time getting significant funding. An attractive solution is to try to keep track of the entire dependency graph: for each positive result, which projects contributed how much to it, then for each project, which projects contributed how much to it, and so on. The main challenge with this design is to figure out the weights of the edges so that they are resistant to manipulation. After all, such manipulation happens all the time. A distilled human judgment mechanism might help.

Conclusion

These ideas have been theorized for a long time: the earliest writings on prediction markets and even decision markets are decades old, and similar accounts in financial theory are even older. However, I believe that the current decade offers a unique opportunity for the following main reasons:

Infofinance addresses a trust problem that people actually have. A common concern of this era is a lack of knowledge (or worse, a lack of consensus) about who to trust in political, scientific, and business settings. Infofinance applications can help be part of the solution.

We now have scalable blockchains as a foundation. Until recently, fees were too high to really implement these ideas. Now, they are no longer too high.

AI as a participant. Infofinance was relatively difficult to work when it had to rely on humans to participate in every problem. AI improves this situation greatly, enabling efficient markets even on small-scale problems. Many markets will likely have a combination of AI and human participants, especially when the volume of a particular problem suddenly changes from small to large.

To make the most of this opportunity, we should go beyond just predicting elections and explore what else Infofinance can do for us.

Special thanks to Robin Hanson and Alex Tabarrok for their feedback and comments