Vitalik Buterin: Prediction markets will become increasingly important Ethereum applications

This article was published on the "Babyt" WeChat public account, original author: Vitalik Buterin, translated by Si Fei.

Note: In this article, Vitalik Buterin reflects on his experiences with prediction markets and summarizes the four major issues currently facing prediction markets. Vitalik also predicts that in the coming years, prediction markets will become increasingly important Ethereum applications, and the 2020 U.S. presidential election is just the beginning, with conditional predictions, decision-making, and other applications to follow.

Special thanks to Jeff Coleman, Karl Floersch, and Robin Hanson for their critical feedback and comments.

Warning: I express some political views.

Prediction markets have been a topic of interest for me over the years, allowing anyone to bet on future events and using the odds of these bets as a credible neutral source of predictive probabilities for those events, which is a fascinating application. Related ideas (like futarchy) have always intrigued me as they can serve as innovative tools for improving governance and decision-making. As shown by Augur, Omen, and the recent PolyMarket (note: these three applications are built on Ethereum), prediction markets are also a captivating application of blockchain.

During the 2020 U.S. presidential election, it seemed that prediction markets finally entered the public eye, with these blockchain-based markets growing from nearly zero in 2016 to millions of dollars in volume by 2020. As someone very interested in seeing Ethereum applications cross the chasm and achieve widespread adoption, this certainly piqued my interest. Initially, I tended to simply observe rather than participate: I am not an expert in U.S. electoral politics, so why should I expect my views to be more accurate than those of others who were already trading? However, in my Twitter circle, I saw many very smart people I respected arguing that the market was indeed irrational, and if possible, I should get involved and bet against them. Eventually, I was persuaded.

I decided to conduct an experiment on the blockchain application I helped create: on August 1, I purchased $2,000 worth of NTRUMP on Augur (if Trump lost, holding this token would yield $1). At the time, I had no idea that my position would eventually grow to $308,249 and earn me over $56,803 in profit. What happened in the following two months proved to be a captivating case study on social psychology, expertise, arbitrage, and the limits of market efficiency, which has significant implications for anyone interested in the possibilities of economic system design.

1. Before the Election

My first bet on this election was actually not on the blockchain at all.

Last July, when Kanye announced his presidential run, a political theorist I usually respect immediately claimed on Twitter that he believed this would split the anti-Trump vote and lead to a Trump victory. I remember thinking at the time that his view was overly confident and might even be the result of over-internalizing heuristics (i.e., if a view seems clever and unconventional, it is likely correct). So, I certainly offered to bet $200, with my own boring bet being that Biden would win, which he gladly accepted.

The election in September caught my attention again, this time due to prediction markets. The market gave Trump nearly a 50% chance of winning, but I saw many very smart people in my Twitter circle pointing out that this number seemed too high. This led to the familiar "efficient market debate": if you can buy a token for $0.52 that will pay you $1 if Trump loses, and the actual probability of Trump losing is much higher, why don't people just come in and buy the token until the price rises higher? If no one does this, what makes you think you are smarter than everyone else?

Before election day, a Twitter post by Ne0liberal did an excellent job summarizing the accuracy of prediction markets at that time. In short, most prediction markets (non-blockchain) used before 2020 had various limitations that made it difficult for people to participate with small amounts of cash. The result was that if a very smart individual or professional organization saw a probability they thought was wrong, their ability to push the price in the direction they believed was correct would be very limited.

The most important limitations pointed out in the article were:

1. Low betting limits for everyone (below $1,000);
2. High fees (for example, PredictIt charges a 5% withdrawal fee);

This is where I disagreed with Ne0liberal in September: while the boring old-world centralized prediction markets may have issues with low limits and high fees, cryptocurrency markets do not! On Augur or Omen, if someone believes the price of a certain outcome token is too low or too high, there is no limit to how much they can buy or sell. However, the price performance of blockchain-based prediction markets was synchronized with PredictIt. If the market was indeed overestimating Trump due to high fees and low trading limits, preventing rational traders from outbidding overly optimistic traders, then why would the blockchain market, which has none of these issues, show the same price?

The main response from my Twitter friends was that blockchain-based markets are very niche, with few participants, especially very few who have much knowledge of politics. This seemed reasonable, but I was not very confident in this argument. So at that time, I bet $2,000 that Trump would lose, but did not place any further bets.

2. During the Election

Then the election happened, and in the initial panic, Trump initially won more seats than we expected, but Biden ultimately became the winner. As far as I know, whether the election itself validated or refuted the efficiency of prediction markets is an easy topic to explain. On one hand, applying standard Bayesian formulas, I should lower my confidence in prediction markets, at least relative to Nate Silver. The prediction market gave Biden a 60% chance of winning, while Nate Silver gave Biden a 90% chance. Given that Biden actually won, this proves I live in a world where Nate gave the more accurate answer.

But on the other hand, you could make a case that prediction markets better estimated the margins of victory. Nate's median probability distribution suggested that about 370 out of 538 electoral votes would go to Biden:

The Trump market did not provide a probability distribution, but if you had to guess the probability distribution from the statistic "Trump has a 40% chance of winning," you might estimate Biden would receive about 300 electoral votes. The actual result was 306, so in my view, the net score of prediction markets vs. Nate is ambiguous.

3. After the Election

What I could not imagine at the time was that the election itself was just the beginning. In the days following the election, Biden was declared the winner by major organizations and even a few foreign governments. As expected, Trump raised various legal challenges to the election results, but these challenges quickly failed. However, during that month, the price of the NTRUMP token remained at 85 cents!

Initially, it seemed reasonable to speculate that Trump had a 15% chance of overturning the election results, after all, he appointed three Supreme Court justices, and at that time, partisan tensions were escalating. However, over the next three weeks, it became increasingly clear that the challenges were failing, and Trump's hopes of winning seemed to diminish day by day, yet the price of the NTRUMP token did not change (in fact, it even briefly dropped to around $0.82). On December 11, more than five weeks after the election, the Supreme Court decisively rejected Trump's attempts to overturn the vote, and the price of NTRUMP finally rose… to $0.88.

In November, I finally became convinced that the market skeptics were right, so I got involved and bet that Trump would lose. This decision was not about money; after all, just two months later, I could make a lot of money just by holding Dogecoin, but I chose to donate it to GiveDirectly. More precisely, I was not just participating as an observer but as an active participant in the experiment, which helped enhance my understanding of why others had not purchased NTRUMP tokens before me.

4. Operational Process

I purchased my NTRUMP tokens on Catnip (a front-end user interface that combines Augur prediction markets with Balancer constant function market makers). So far, Catnip is the simplest interface for conducting these transactions, which I believe greatly enhances the usability of Augur.

There are two ways to bet on Trump losing through Catnip:

1. Directly purchase NTRUMP tokens on Catnip using DAI;
2. Use Foundry to access Augur features, allowing you to convert 1 DAI into 1 NTRUMP + 1 YTRUMP + 1 ITRUMP (where "I" stands for "invalid," which will be explained later), and then sell YTRUMP on Catnip;

Initially, I only knew the first option. However, I later discovered that Balancer provided more liquidity for YTRUMP, so I switched to the second option.

There was another issue: I had no DAI; I only had ETH. I could choose to sell my ETH to obtain DAI, but I did not want to disturb my ETH holdings. If I won a $50,000 bet against Trump but simultaneously lost $500,000 due to ETH price fluctuations, that would be a shame. Therefore, I decided to maintain my ETH exposure by opening a collateralized debt position (CDP) on MakerDAO (now also known as a "vault").

A CDP is a way to generate DAI: users deposit their ETH into a smart contract and are allowed to withdraw up to 2/3 of the value of the deposited ETH in newly generated DAI, which they can reclaim by returning the same amount of DAI plus an additional interest (currently 3.5%). If the value of the ETH collateral you deposited falls below 150% of the value of the borrowed DAI, anyone can enter and "liquidate" the vault, forcibly selling your collateralized ETH to buy back DAI and charging you a hefty penalty. Therefore, having a high collateralization ratio is a good idea in the event of sudden price fluctuations. For every $1 I withdraw, there is over $3 of ETH in my CDP.

The following diagram illustrates my entire operational process.

I did this many times, and the slippage on Catnip meant that I could usually only conduct transactions of about $5,000 to $10,000 at a time, so the price would not turn against me (when I skipped Foundry and directly bought NTRUMP with DAI, the limit was close to $1,000). Two months later, I accumulated over 367,000 NTRUMP tokens.

5. Why Don't Others Do This?

Before I got involved, I had four main hypotheses about why so few people were buying at 85 cents:

1. Concerns that Augur's smart contracts might have issues or that Trump's supporters might manipulate the oracle (Augur REP token holders vote on one or more outcomes) to return incorrect results;
2. Capital costs: to purchase these tokens, you must lock up funds for more than two months, which prevents you from using those funds or making other profitable trades during that time;
3. Technically too complex, making it unsuitable for everyone to participate;
4. In reality, there are far fewer people with sufficient motivation to seize an odd opportunity than I imagined, even if it is quite obvious.

All four points make sense; the risk of smart contract attacks is the biggest risk factor, and the Augur oracle has never been tested in such a controversial environment. The capital cost is real, even though betting on prediction markets is easier than betting in the stock market (because you know the price will never exceed $1, but locking up capital competes with other profitable opportunities in the crypto market). Additionally, trading in dapps is technically complex, leading to a natural degree of fear among people.

My actual experience entering this financial domain and observing the evolution of market prices taught me a lot about these assumptions.

6. Fear of Smart Contract Vulnerabilities

Initially, I thought "fear of smart contract vulnerabilities" must be a major concern for people. However, over time, I became increasingly convinced that this might not be a dominant factor. One way to examine this is to compare the prices of YTRUMP and ITRUMP.

ITRUMP stands for "Invalid Trump," where "invalid" refers to outcomes triggered under certain special circumstances: when the description of the event is ambiguous, when the market resolves, when the outcome of the event is unclear, when the market is unethical (such as assassination markets), and other similar situations. In this market, the price of ITRUMP has consistently remained below $0.02. If someone wanted to profit by attacking the market, it would be more profitable for them to buy ITRUMP at $0.02 rather than YTRUMP at $0.15. If they purchased a large amount of ITRUMP, they could force the "invalid" outcome to actually trigger, yielding a 50-fold return. Therefore, if you are concerned about being attacked, buying ITRUMP is by far the most rational choice. However, very few people chose to do so.

Of course, another argument against the fear of smart contract vulnerabilities is that in every crypto application outside of prediction markets (such as Compound and various yield farming schemes), people show a surprising indifference to smart contract risks. If people are willing to put their money into various risky and untested schemes (even those promising only 5-8% annual returns), then why do they suddenly become overly cautious here?

7. Capital Costs

Capital costs (i.e., the inconvenience and opportunity cost of locking up large amounts of funds) are a challenge, and I have become more aware of this than ever. Just from the Augur perspective, I needed to lock up $308,249 of DAI for a profit of $56,803, which is roughly a 175% annualized return. From the current perspective, this return is quite substantial, even compared to various high-yield liquidity mining activities in the summer of 2020. However, when you consider that I needed to operate on MakerDAO, the situation becomes a bit worse. Because I wanted to maintain my ETH position, I needed to obtain DAI through a CDP, and safely using a CDP requires a collateralization ratio of over 3. Therefore, the total amount of funds I actually needed to lock up was about one million dollars.

Looking at it now, this return does not seem very favorable. Moreover, if you consider the potential for hacking or the possibility of truly unprecedented political events, the appeal of participating diminishes significantly.

However, even so, assuming a 3x capital lock-up and a 3% chance of an Augur contract attack (I purchased ITRUMP to hedge the risk), this could reduce the risk-neutral rate to around 35%, and if you consider how real people perceive risk, this percentage might be even lower. This trade remains very attractive, but on the other hand, it is now very understandable that such numbers are not appealing enough for those who have experienced 100x fluctuations in the cryptocurrency market.

On the other hand, Trump supporters did not face these challenges: they only invested $60,000 to cancel my $308,249 bet (due to fees, I won less than this). When probabilities are close to 0 or 1, as in this case, the game is very unbalanced, favoring those trying to push probabilities away from extremes. This not only explains Trump's situation but also why various popular niche candidates with no real chance of winning often receive as high as 5% winning probabilities.

8. Technical Complexity

Initially, I tried to purchase NTRUMP on Augur, but technical issues with the user interface prevented me from placing orders directly on Augur (others I spoke with did not have this issue… I still am not sure what happened).

The UI of Catnip is much simpler and runs very smoothly. However, automated market makers like Balancer (and Uniswap) are best suited for smaller trades, and for larger trades, slippage is very high. This is a good microcosm of the broader "AMM vs order book" debate: AMMs are more convenient, but order books are indeed more efficient for large trades. Uniswap v3 is introducing an AMM design with better capital efficiency, and we will see if it can improve the situation.

There are other technical complexities, but fortunately, they all seem relatively easy to solve. Interfaces like Catnip have no reason not to integrate the "DAI->Foundry->sell YTRUMP" path into a single contract, allowing you to purchase NTRUMP tokens in this way in a single transaction. In fact, this interface could even check the price and liquidity attributes of both the "DAI->NTRUMP" path and the "DAI->Foundry->sell YTRUMP" path, automatically providing you with the better trade. Even extracting DAI from a MakerDAO CDP could be included in this path. My conclusion here is optimistic: the technical complexity issues are the real barriers to current adoption of prediction markets, but as technology advances, using them will become much easier.

9. Lack of Confidence

Now, we have the final possibility: many people (especially smart ones) suffer from a problem of excessive humility, making it easy to conclude that if no one else is taking any action, there must be good reasons why that action is not worth taking.

Eliezer Yudkowsky elaborates on this in the latter half of his excellent work "Inadequate Equilibria," arguing that too many people overuse "humble epistemology," and we should act based on our reasoning results, even if the results indicate that the vast majority of people are irrational, lazy, or wrong about certain things. When I first read these chapters, I was not convinced; it seemed that Eliezer was just being overly arrogant, but after going through this experience, I saw some of his wisdom.

This is not the first time I have witnessed the merits of believing in one's reasoning. When I initially started working on Ethereum, I was plagued by fear, worried that the project had some reasons destined to fail. I inferred that a fully programmable smart contract blockchain was clearly a significant improvement over what had come before, and surely many people had thought of it before me. So I fully anticipated that once I published this idea, many very smart cryptographers would tell me the reasons why something like Ethereum was simply impossible. However, no one ever did.

Of course, not everyone has the flaw of excessive humility. Many who predicted Trump would win the election could be said to have been fooled by their own excessive contrarianism. Ethereum benefited from my repression of humility and fear in my youth, but many other projects could benefit from more knowledge-based humility to avoid failure.

However, in my view, as the famous Yeats quote goes, "The best lack all conviction, while the worst are full of passionate intensity," this is more true than ever. It seems that spreading the message to society that the solution is simply to trust the existing outputs of society, whether those outputs come in the form of academic institutions, media, government, or markets, is not the answer. All these institutions operate precisely because some people believe they do not work, or at least some believe they may be wrong at times.

10. Lessons from Futarchy

Witnessing the importance of capital costs and their interaction with risk also serves as important evidence for judging systems like Futarchy. Futarchy and "decision markets" are often an important and potentially very useful social application of prediction markets. Predictions about who will be the next president have little social value if they are slightly accurate. However, conditional predictions have a lot of social value: how likely is it that if we do A, it will lead to some good outcome X, and how likely is it if we do B? Conditional predictions are important because they not only satisfy our curiosity but also help us make decisions.

Although election prediction markets are far less useful than conditional predictions, they help reveal an important question: how strong is their resistance to manipulation, or even just bias and erroneous views? We can answer this question by observing how difficult it is to arbitrage: suppose the probabilities currently given by the conditional prediction market (in your view) are wrong (perhaps due to under-informed traders or obvious attempts at manipulation). How much impact can you generate by setting things up correctly, and how much profit can you earn?

Let’s start with a concrete example. Suppose we are trying to choose between decision A and decision B using prediction markets, where each decision has a chance of achieving certain ideal outcomes. Assume you believe decision A has a 50% chance of achieving the goal, while decision B has a 45% chance. However, the market (which you believe is wrong) thinks decision B has a 55% chance and decision A has a 40% chance.

Assuming you are a small participant, your individual bet will not affect the outcome; only when many people bet together will it have an impact. So how much should you bet?

The standard theory here relies on the Kelly criterion. Essentially, you should act to maximize the expected logarithm of your wealth. In this case, we can solve the outcome equation. Suppose you invest part of your funds to buy A tokens at a price of $0.4. From your perspective, your expected new logarithmic wealth is:

The first term is the 50% probability (from your perspective) of betting on a return, where your invested portion grows by 2.5 times (because you bought dollars at 40 cents). The second term is the chance of betting with no return at 50%, where you lose the portion of your bet. We can use calculus to find the way to maximize this; for the lazy, you can use WolframAlpha, and the answer is: r = 1/6. If others buy in, the price of A in the market rises to 47% (B drops to 48%), we can redo the calculation for the last trader, who will flip the market to correctly favor A:

Here, the expected logarithmic wealth maximization r value is only 0.0566. The conclusion is clear: when decisions are close and there is a lot of noise, it proves to make sense to invest only a small portion of funds in the market. This assumes rationality, and most people invest less in uncertain gambles than the Kelly criterion suggests. Capital costs are even higher. However, if an attacker really wants to force outcome B for personal reasons, they can use all their funds to buy that token. Overall, the game can easily favor the attacker by more than 20:1.

Of course, in reality, attackers are rarely willing to bet all their funds on one decision. And futarchy is not the only mechanism vulnerable to attack; stock markets are equally fragile, and non-market decision-making mechanisms can also be manipulated in various ways by determined wealthy attackers. But in any case, we should be cautious not to assume that futarchy will push us to new heights of decision accuracy.

Interestingly, the math seems to suggest that when expected manipulators want to push the outcome toward an extreme, futarchy will perform best. One such example might be liability insurance, as someone wishing to improperly obtain insurance would effectively try to drive the market estimate probability of adverse events to zero.

11. Can Prediction Markets Improve?

The final question to ask is: are prediction markets doomed to repeat the same mistakes? Like when it assessed the chances of Trump overturning the election at 15% in early December, or even after the Supreme Court (including the three justices he appointed) told Trump to go away, it still gave a 12% chance of overturning the election? Surprisingly, I believe prediction markets will not repeat their mistakes, and I see some optimistic reasons.

1. Markets are Subject to Natural Selection

First, these events have given me a new perspective on how market efficiency and rationality actually arise. Proponents of market efficiency theory often claim that market efficiency arises because most participants are rational (or at least rationality is more important than any superstitious group), which is an axiom. However, conversely, we can view what is happening from an evolutionary perspective.

The crypto market is a young ecosystem. Despite Elon’s recent tweets, this ecosystem remains disconnected from the mainstream and lacks sufficient expertise in electoral politics. Those political science experts find it difficult to enter the cryptocurrency space, and there are many forms of crypto that are not always correct, especially in the political realm. However, what happened this year is that in the cryptocurrency space, the capital of prediction market users who correctly anticipated Biden's victory increased by 18%, while the capital of those who incorrectly predicted Trump's victory decreased by 100% (or at least the portion they bet).

Thus, there will be selection pressure. After ten rounds of such predictions, good predictors will have more capital to bet, while poor predictors will have less capital to bet. This does not rely on anyone "becoming wiser" or "learning their lessons" or any other assumptions about human reasoning and learning abilities. It is simply the result of selection dynamics; over time, participants who excel at making correct guesses will dominate the ecosystem.

Notably, prediction markets perform better in this regard than stock markets: stock market "whales" often get lucky with a thousand-fold return, which adds a lot of noise to the signals, but in prediction markets, prices are constrained between 0 and 1, limiting the impact of any single event.

2. Better Participants and Better Technology

Second, prediction markets themselves will improve. User interfaces have already seen significant improvements and will continue to improve. The complexity of MakerDAO->Foundry->Catnip operations will be abstracted into a single transaction. Blockchain scaling technologies will improve to reduce participant fees (ZK rollup Loopring with built-in AMM is already running on the Ethereum mainnet, theoretically allowing prediction markets to run on it).

Third, the demonstrations we have seen of prediction markets functioning correctly will alleviate participant concerns. Users will see that the Augur oracle can provide correct outputs even in very controversial situations. People from outside the crypto industry will see that the process is effective and will be more inclined to participate. Perhaps even Nate Silver himself will use some DAI and utilize Augur, Omen, Polymarket, and other prediction markets to supplement his income in 2022 and beyond.

Fourth, the technology of prediction markets themselves can improve. Here is a market design suggestion I proposed that could enhance capital efficiency while betting on many unlikely events, helping to prevent unlikely outcomes from receiving unreasonably high probabilities. Other ideas will surely emerge, and I look forward to seeing more experiments in this direction.

Conclusion

Through an incredible direct experiment with prediction markets and how they clash with the complexities of individual and social psychology, it reveals a lot about how market efficiency operates in practice, what its limitations are, and what can be done to improve it.

It also showcases the capabilities of blockchain, which I believe is one of the most valuable applications of Ethereum. Blockchain is often criticized as a speculative toy, doing nothing meaningful except for self-referential games (liquidity mining, whose returns are often paid in other issued tokens). Of course, critics fail to recognize exceptions; I personally benefit from ENS and even benefit from using ETH for payments when all credit card options fail. However, in recent months, we seem to have witnessed the rapid development of Ethereum applications that provide concrete help to people and interact with the real world, with prediction markets being a key example.

I expect that in the coming years, prediction markets will become increasingly important Ethereum applications, and the 2020 election is just the beginning. Future prediction markets will receive more attention, not just for elections but also for conditional predictions, decision-making, and other applications. If prediction markets operate mathematically in the best way, what amazing promises will they bring? Of course, this will continue to clash with the limits of human reality, and I hope that over time, we will gain a clearer understanding of where this new social technology can provide the greatest value.