Predictive Market Fundamentals: Leverage AI to create predictive markets at a micro level

原文标题：The prediction market primitive

原文作者:Hiroki Kotebe

Original source: medium

编译：Inception Capital

tl'dr

Sooner or later, the prediction market has been expected to take off, and a continuously improving user experience lays the foundation for this.
But to scale to billions of users, we need "something new" that goes beyond continuous user experience improvement, and that is to make AI the key and core of the machine.
An AI quartet of content creators, event referrers, liquidity distributors, and information aggregators can spark large-scale new activity in this space.
Integrating these AIs into the current prediction market framework enables prediction markets at a micro scale, making them personally engaging and relevant.
The original form of the prediction market paved the way for Tinder-like prediction market applications, embedding the prediction trading experience into our everyday digital presence.

Every decision starts with a prediction. Think about the potential of Bitcoin: "Buying Bitcoin now, will it double the investment by the end of the year?" If the prospect of "yes" is considered more likely than "no", then it is economically rational to decide to buy Bitcoin in the absence of a better option.

But why limit ourselves to Bitcoin? Imagine that we could build a market based on predictions of various events, such as who will be the next president of the United States, or which country will win the World Cup. Here, it's not the asset that's being traded, it's the prediction itself.

Forecasts shape the market, and the market validates our predictions.

The prediction market is called the "holy grail of cognitive technology" by Vitalik.

Predictive Market Fundamentals: Leverage AI to create predictive markets at a micro level

Vitalik is adept at seeing big opportunities before anyone else, so he's a great source of information for cutting-edge storytelling. He came up with the idea of an AMM (Automated Market Maker) on Ethereum in a blog post seven years ago. Then a guy named Hayden Adams answered the call and began building it, receiving a $60,000 grant. Two years later, Uniswap was born.

If Vitalik's blog post can spark an industry that creates more than $100 billion, then we should probably pay attention to his point. For example, Vitalik was excited about the use of predictive markets in governance in 2014 – a radical form of governance known as "politics of the future" – and now we see Meta DAO implementing this philosophy, with large venture capital firms such as Pantera also getting involved.

But what we want to focus on is his recent discussion about prediction markets and AI, as we're starting to see some big things taking shape.

The forecast market is ready to take off

The current market-leading prediction market is Polymarket, thanks to its continuously improving user experience, as well as the expansion of event categories and event offerings.

Source: Dune

Trading volumes have hit an all-time high recently, and with the US presidential election in November of this year, where Polymarket's users are mainly concentrated, it is likely to increase further.

There are other signs that the prediction market may take off this year. In addition to the all-time high in the crypto market in 2024, this year is also the year with the most elections in history. The world's ten most populous countries, including the United States, India, Russia, Mexico, Brazil, Bangladesh, Indonesia and Pakistan, will hold elections. In addition, the 2024 Summer Olympics are about to be held in Paris.

But considering that the monthly trading volume is still only in the tens of millions of dollars, and it could have reached hundreds of millions of dollars, let's consider some of the limitations of the current forecast market:

Centralized control over event creation
Lack of incentives for community content creators
Lack of personalization
Mainly concentrated in the United States, ignoring the huge international opportunities

But we need "something new"

We believe this thing is AI.

Humans need AI as a participant in the game. We expect to soon see AI (bots) involved in the prediction market alongside human agents. We're already seeing demos of this on platforms like Omen and PredX, and there are probably a number of others coming into the space. More on that later.

AI needs AI as the arbiter of the game. Although relatively rare, dispute resolution is important and necessary in the prediction market. For example, in a presidential election, the results can be very close, and allegations of electoral fraud will arise. Thus, while the prediction market may end in favor of Candidate A, the official election commission may declare Candidate B the winner. Punters supporting Candidate A would oppose the outcome by claiming fraud, while punters supporting Candidate B would argue that the Election Commission's ruling reflected the "true" outcome. There's a lot of money at stake, and who's right?

There are several challenges to answering this question:

Players may not trust human arbiters because they are biased
Human arbitration can be slow and expensive
DAO-based prediction solutions are vulnerable to Sybil attacks

To address this issue, the prediction market could use a multi-round dispute resolution system similar to Kleros, but using AI instead of humans to resolve early-stage disputes, and humans will only participate in arbitration of disputes that are ultimately unsolvable in rare cases. Players can trust the AI to be impartial, as it is not feasible to create enough training data to influence the AI. In addition, AI referees work faster and at a lower cost. xMarkets is working in this direction.

AI can spark desire

For prediction markets to really take off, they need to be able to generate enough interest for people to overcome the psychological barriers and really start trading prediction assets. For topics that many care about, like who will win the presidential election or the Super Bowl, it probably doesn't take much effort. However, the inclusion of only such topics severely limits potential liquidity. Ideally, a prediction market should be able to attract the liquidity of a particular event that is of high interest to a specific audience, just like targeted advertising, which we all know works.

To achieve this, the forecasting market needs to address four major challenges:

Event Supply: A highly relevant event supply is critical. To reach a niche but loyal audience, event creators must gain insight into their community's interests to drive increased engagement and transactions.
Event Demand: The demand needs to be high within a particular target community, taking into account their number of users and psychology.
Event Liquidity: There is enough diversity and dynamics of views within the target community to drive sufficient liquidity, engage both parties and minimize slippage.
Information aggregation: Players should have easy access to enough information to give them the confidence to place their bets. This may include background analysis, relevant historical data, and expert opinion.

Now, let's look at how AI can address these challenges:

Content creator AI: Content creator AI ("copilots") can assist in the creation of content that exceeds human capabilities or motivations. By analyzing trends in news, social media, and financial data, AI provides timely and relevant event topics. Whether it's a human or an AI content creator, it pays off when it comes to creating engaging content that keeps the community alive. Community feedback enhances AI's understanding of the community, making it an ever-improving content creation engine that connects content creators and audiences.
Event Recommendation AI: Event recommendation AI tailors event recommendations to users based on their interests, transaction history, and specific needs, with a focus on events that are controversial and trading opportunities. It adjusts to the user's behavior across different regions, cultures, and times. The ultimate goal is to provide highly targeted event recommendations, free from the distractions of today's prediction market platforms with content that is not relevant to the individual.
Liquidity Distributor AI: Liquidity Distributor AI handles counterparty liquidity risk by optimizing liquidity injection to narrow bid-ask spreads. To minimize risk, AI can employ a logarithmic market scoring rule (LMSR) AMM model specifically designed to minimize risk in low-liquidity forecasted markets. They can also be combined with reinforcement learning agents to dynamically adjust liquidity depth, protocol fees, and bond curves to further reduce risk. These AIs manage event liquidity from a pool of universal liquidity providers, rewarding contributions through accumulated fee income or platform tokens as further incentives. Overall, this means pre-adaptation to market changes, reduced slippage, and better price stability.
Information aggregation AI: These AIs utilize a variety of metrics (e.g., on-chain data, historical data, news, sentiment metrics) to make a comprehensive sense of events. Based on this, information aggregation AI can provide comprehensive forecasts, turning the prediction market into a preferred source of informed decisions and alpha. Projects can choose to limit access to the insights gained by information aggregation AI through tokenization, because in the prediction market, knowledge equals money.

Now, let's take a look at what it looks like when you put these together. Below, you can see the main components and how the prediction market without AI works (black), and the prediction market with AI (blue).

In a non-AI model, the content creator (usually the platform itself) creates events at will, provides liquidity (initially subsidized by its treasury), saves events to an event database, and promotes them to human players in bulk. That's how Polymarket works at the moment, and it works pretty well.

However, I think it could be better.

In the AI model, collaborative AI for content creators (copilot AIs) enables content creators to create and promote events within a targeted, general or niche community. Liquidity offering is powered by liquidity dispenser AI, which optimizes liquidity injection over time by learning from the player's order book and using external data from oracles and other data providers. Event Recommendation AI uses stored events and wallet transaction history in the event database to optimize event recommendations based on individual interests. Finally, information aggregation AI collects information from data vendors to provide education and context to human players and provides AI players with information about their predictive decisions. The ultimate goal is to create a finely tuned prediction market system that enables prediction markets to operate at the micro level.

A prediction market at this scale will bring a different user experience, more like Tinder or TikTok. Since events are highly targeted, they can be presented to you via a dynamic stream (feed) like TikTok, and even with today's wallets and blockchain technology, players can place bets by swiping left or right like Tinder does. Imagine people making micro-bets on events they personally care about while commuting to work or school.

Enhance the ability to aggregate information

One of the hardest outcomes to predict is asset prices, so let's focus here and see how AI performs on the possible edge of the prediction market.

In academia, research using AI to predict asset prices is being actively explored. Machine learning techniques such as linear models, random forests, and support vector machines have been shown to be more accurate than human judges when predicting cryptocurrency prices. These models found that behavioral indicators like Google search strength can explain changes in price.

IBM Research explores the application of human forecasting markets to commodity price forecasting and provides a persuasive case study that demonstrates the potential of combining AI with prediction markets. Their research highlights the artificial prediction market integrating diverse and evolving real-time sources of information to make better predictions in complex real-world problems, such as predicting volatile commodities (e.g., ethylene, hydrocarbons) that are traded on non-online exchanges. Here, AI agents are able to outperform standard machine learning models because they can gradually improve their prediction capabilities by learning from themselves.

Another study compared the performance of random forest regression and LSTM in predicting Bitcoin's next-day price, and the results showed that the former performed better at predicting less error. It also demonstrates the power of AI in aggregating information far beyond the ability of the average person to model 47 variables, including (a) Bitcoin price variables, (b) Bitcoin's technical indicators, (c) other token prices, (d) commodities, (e) market indices, (f) foreign exchange, (g) public attention, and (h) dummy variables over the course of a week. The most important predictors change over time, from the U.S. stock market index, oil price, and Ethereum price from 2015 to 2018, to the Ethereum price and the Japanese stock market index from 2018 to 2022. The study also found that the random forest regression model performed best with a one-day delay when predicting the next-day price of Bitcoin.

We can infer that in some of the popular prediction markets, the time required for a busy person to aggregate, analyze, and interpret large amounts of data to make good predictions is simply too little. Or, the problem itself is too complex. But AI can do that.

AI token recommendation

Pond is building a decentralized base model based on cryptocurrencies that has been applied to AI token recommendations based on on-chain behavior. Currently, they use large-scale graph neural networks (GNNs) to estimate the alpha probability of various tokens using on-chain behavioral data. GNNs are a class of AI models specifically designed to work with data represented graphically, so they are useful when there is a relational structure between data, such as the peer-to-peer transaction network of blockchains. Dither is another token recommendation AI that uses a time series modeling approach to token recommendations through a token-gated Telegram alert bot.

Solve the problem that the market is too weak

One of the main challenges facing the forecasting market is that the market is too weak to attract enough participants and volume. However, there is one major difference between the prediction markets of the 2010s and 2020s, and that is the possibility of universal AI participation. As Vitalik points out:

In addition, automated market maker (AMM) models that predict the underlying market can be improved. For example, when analyzing more than 2 million transactions on Polymarket, it was found that there are liquidity supply problems in convergent prediction markets using the traditional constant product AMM(x*y=k), including:

Convergence and liquidity removal. As the prediction market converges (i.e., the outcome becomes more certain), liquidity providers will have an incentive to withdraw their liquidity. This is reasonable behavior because of the increased risk of holding "loss-making" tokens. For example, in a "yes" trending market, "no" tokens become less valuable (i.e., impermanent loss), which is a risk for liquidity providers who may end up with some worthless tokens if they don't sell them early.
Deviations and inaccuracies. Reduced liquidity can lead to less accuracy and more bias in predicting market convergence. Specifically, in the volume-weighted price range of 0.2 to 0.8, "no" tokens are often undervalued, while "yes" tokens are often overvalued.

Source: Kapp-Schwoerer (2023)

To address these issues, the authors propose a "Smoothed Liquidity Market Maker" (SLMM) model and demonstrate that it can increase the volume and accuracy of convergent prediction markets. It does this by introducing a centralized function (similar to Uniswap v3) into the model, in which the liquidity provided by liquidity providers is only valid within a specific price range. The result is a reduced risk exposure, ensuring that the number of valuable tokens held by liquidity providers (e.g., "yes" tokens in a market where the market converges to a "yes" result) does not converge to zero at the time of price adjustment, unlike a constant product AMM.

LP - trade-off for traders

There is a balance that must be struck when choosing a centralized liquidity AMM variant like SLMM for convergent prediction markets. While you're trying to reduce the risk to liquidity providers, you're also reducing the incentive for some trading activity.

Specifically, concentrating liquidity can reduce the likelihood of liquidity providers losing money when the market converges to a definite outcome (from

and reduce early withdrawals), but it may also reduce the opportunity to profit from trading on small price movements (e.g. from $0.70 to $0.75) due to increased slippage, especially for large orders. The immediate result is that the trader's potential profit margins are squeezed. For example, if they expect the price to rise from $0.70 to $0.75, slippage may limit the amount of money they can use effectively to capture the expected increase. Going forward, it will be important to experiment with adjustments of various trade-offs in these market maker formulas to find the sweet spot.

conclusion

The prediction market is a powerful infrastructure. Of course, like any other crypto infrastructure, it comes with challenges, but we are confident that these challenges will be overcome. As these challenges are gradually addressed, we can expect to reuse this infrastructure to answer a variety of questions across a variety of digital environments. With advancements in targeting and liquidity solutions, we can expect the development of the prediction market in specific sectors. For example, take X (formerly Twitter) user as an example:

Will X launch Premium++ or equivalent before the end of the year?
Will X make the ability to edit tweets available to all users in Q3?
Will X report an increase in daily active users in its next quarterly report?
Will X's ad revenue increase or decrease in the next quarter?
Will X announce a new major partnership with content creators before the end of the year?
Will X release blockchain- or cryptocurrency-related features in Q3?

Interestingly, these questions don't need to be confined to independent prediction market websites. They can be integrated directly into X or other platforms via browser extensions. We may often see the emergence of micro-prediction markets in our day-to-day online experiences, adding opportunities for speculative trading to the ordinary browsing experience.

Some of the questions above were written by me and some by content generation AI. If it's hard to tell, it's because ChatGPT's content generation AI is already excellent. The same goes for information aggregation AI and recommendation engines built by other big tech companies (just look at the ads Google and Instagram show you). While it takes work and time to achieve the performance of these models, they demonstrate the viability of these AI categories. The main open-ended question that lacks precedent relates more to liquidity allocator AI, AI actors, and the self-improvement and goal-oriented development of AI – the evolution from basic machine learning to verifiable AI agents.

If you are developing in these areas, or if this article resonates with you, please feel free to contact us!