Beyond x·y = k: Designing an AMM for Decentralized Prediction Markets | Eric Lee | ETHTaipei 2026

Good afternoon everyone and thanks to the AT Taipei team for organizing this great event. Really happy to join you today and share what we are building at Safe Market. Unfortunately, I could join you person this time, but I hope I'll get a chance to visit Taipei and meet you all face to face next time. And now, let's get into today's topic. Beyond XY equals K, designing an AMM for decentralized prediction markets.

Here's what I will cover today. First, a practical question is why do we need truly decentralized prediction markets as infrastructure? and then I will walk you through our AMM. It gives traders one quote while letting liquidity providers choose different levels of risk exposure. And finally, I'll introduce an idea we are exploring called future news and how prediction markets could become useful information infrastructure in the future.

And all right, let's begin. Actually, we are not trying to decentralize something just for the sake of decentralization. There's a practical need behind this. Think about poly market like someone from a completely different country with completely different interests might be browsing the same headline markets as you, but do you really care about the same questions? A community might care about a local funding decision.

A project team might care about meeting a deadline. These questions matter even if they don't attract a global audience. Institutions also have their own needs in who can participate, how the outcome is decided and how the market connects to their existing tools. We want to make it easier for them to create and customize markets around those needs and without waiting for a platform to prioritize their question. So what exactly do we need to decentralize to make that possible?

And let's break that down. There are three questions here. Who controls the assets? Who execute the trades and who decides the outcome? And Kelshi follows an exchange model and poly market separates his responsibilities and users can sign orders but an operator matches them and trades set on chain and outcomes go through proposal and dispute process.

So onchain segment doesn't mean every part of the market is decentralized. For someone building on top of this market the practical question is who can restrict access or stop trading? We want builders to be able to use the core market roles without depending on a single platform operator. And what would that let them build? Let's look at the next.

The idea is really simple. Like give people the tools to build and run their own prediction markets. These five steps cover the basics. Create a market, add liquidity, trade, resolve the outcome, and connect it to other applications. A community could launch a market around its own question.

An institution could make it part of a product or workflow. Neither should need to build an entire exchange from scratch. Our AMM focused on two of these steps, funding and trading. Traders want one clear quote and a simple trading experience. But LPS may want different levels of risk exposure.

How do we bring those needs together? That's where our design starts. Here's the idea. One quote for traders, different risk choices for LPS. Imagine some news come out and the people keep buying.

Yes, the pool selling more. Yes. And the price moves toward the edge. Some LPS may want to stop taking more exposure earlier. Others may be comfortable providing liquidity further toward that edge.

The three bars show the different choices. conservative, balanced and aggressive. But traders should have to think about these layers or split their orders across different pools. They should just get one quote and make their trade easily. That's what we are building.

Different explorer choices for app combined into one quote for trader. So how does it work? First, let's look at what the pool actually holds. Let's start with one unit of collateral. the back pair of tokens like one yes and one no.

If S wins, the S pays one and no pays zero. If no wins, it's the other way around. And uh in either way, the pair pays one in total. That gave us a common way to count the pool's inventory. Like for each layer, we add its curve collateral to its ES tokens and to its no tokens.

Those are the two equations on the right. This doesn't mean we have twice the money. It means the collateral backs both possible outcomes. And that's how much of that inventory do we actually make available for trading. We don't use all of a layers inventory for the current quote.

We first set a floor and only an inventory above that floor is available. Look at the bars. The pale part stays outside the quote. The blue part is what we call effective inventory. In this example, the conservative layer keeps more aside while the aggressive layer makes more available.

And now imagine people keep buying. Yes. As the layer is available, yes, inventory runs down. It gets closer to its floor. And once it reaches that floor, it stops providing liquidity in that direction.

That limits further exposure through quality. It does not guarantee that the LP would lose money. Next, how do we combine the active layers into one quote? Now, we take effective inventory from all the active layers and add it together like on one side and no on the other. Those totals give us one local XY= K curve.

So, the trader gets one quote even though several LP layers are providing the liquidity together. The key is that the curve applies while the active set stays the same. When a layer reaches its floor and a trading direction, we update the active set and build the next local curve. That's what we mean by like beyond XY equals K. We still use the formula partly, but our mechanism determines which inventory goes into it.

And next, how do we share the trades inventory change across these layers? Once we have the quotes, we need to share the inventory change across active layers. We do that using each layer's share of the active set life capital. That share is its weight. The distinction is simple.

The weight determines how much of the inventory change layer takes. The floor determines how much room it has left to keep quoting in that direction. And when a layer reaches its floor, we recalculate the widths for the remaining layers. And that's why one order may need several actual execution steps. And let's continue to follow that yes order.

The engine adds up the active inventory, finds an X boundary, and trades up to it. and updates the state finally. And suppose a conservative layer reaches its floor before the order is finished. We remove that layer from the active set for this direction, recalculate the curve and weights and continue with the rest of the order. The trader doesn't need to manage these steps manually.

And uh but if we reach the global price cap we we should stop and only part of the order is filled and the unfilled remainer is returned. So a layer's floor changes who provides liquidity. The global cap stop the order from going further beyond and that's what happens to fees. The fee combines the active layers fee rates using the same live capital weights. So when the active layers change the fee can also change.

One more point about the cap is that it limits the AMMS exposure not the event's actual probability. Trades in the offside direction can still move the market away from that edge. And that's the mechanism like one quote for traders to trade easily and different exposure choices for LPS. And now what can people use it for? We want prediction markets to be useful beyond betting and the question is what decision can this market help someone make.

Take a project deadline. If the market starts expecting a delay, the team could check what's going wrong and partners could adjust their plans. The probability is like a signal to investigate, not something to follow blindly. And that's how a market could fit into a community's decisions or fit into an institution's existing workflow. And if one market can help with one question, what could we learn by like connecting different signals across many prediction markets?

And that bring us to our new idea about the latest product called future news. For future news, we want to bring together event data from prediction markets around the world. First, we line up the events and their timelines. Then, we look for pairs uh where one market's probability tends to move before another. Uh we can put it simply like who moves first and who follows.

Next, we ask an LLM whether that relationship makes economic sense and what else could explain it. And finally, we test the signal through historical trading back tests, uh, including fees and liquidity using fresh periods we didn't use to find the pattern. And moving first doesn't prove position. And also in an LLM's explanation isn't proof either. This is research where we want to explore and the bigger goal is to turn the signals into a new way of reporting and possible futures.

And here's what what that could look like. A news report might say um the project has been delayed, but future news could report in a complete different way like the market now sees a growing chance of a delay and shows which related events might be affected. The difference is where we look on the timeline and what what has happened uh versus what may happen next based on probabilities changing. Right now news already includes forecasts. What we want to add is a view that connects live market probabilities with relationships between events and updates as those signals change.

These are possible futures, not facts yet. That's the information effort in infrastructure we want to build and we'd love to work with others on it and uh we'd love to hear what you'd like to build with S market and maybe you have a question for your community or use case inside your organization or your company and we'll be happy to explore it with you together and we are also looking for research collaborators on future news and people building with AI agents or stable coins who want to explore how prediction markets could fit into their work. And if that sounds interesting, please scan a QR code and get in touch with us. Okay, that's all the contents today and thanks again to it TA team and thank you all for listening. Thank you.