Kris Paruch - Designing Robust Prediction Markets on Ethereum

Transcript

Good morning everyone. Hi, my name is Kristoff  Par. I am from Vienna and I came here today just for the workshop uh about prediction  markets. So um by trade I'm a mathematician and um economist. I I work at the University of  Economics in Vienna. I have my own company, Token Engineering Labs, where we since almost 10  years now design protocols, design token systems, design incentives. The focus of my work  based on my education but also my interest um has evolved around market design has evolved  around the designing protocols that relate on to functionalities on the application layer and  mostly relevant um everything that has to do with information aggregation, everything  that has to do with exchange mechanisms,

everything and that has to do with price  discovery. And this brings us into the story or into the narrative of prediction markets because  there is um they are there's an advent in in crypto in web three right now where um prediction  markets have gained quite a bit of traction. So um what I want to talk you through over the  next 50 60 minutes or so is what prediction markets actually are, how they are designed, what  are the building blocks of prediction markets, what happens under the hood and how  you can build applications on top of it. So we want to cover the full spectrum from  mathematics, economics over token standards, um smart contracts representing the individual  functionalities and the assets and then we want

to talk about implementation effects and um some  of the some attention we will put towards market micro structure and market efficiency. So we want  to look how you could design prediction markets, how you can build them, uh which mechanisms  exist that you can use and how this choice that you make on the microructure side will uh  result in different outcomes or trade-offs on the trading or the performance level. So this is  an academically grounded talk. We want to base our um conversation in academic literature in  state-of-the-art economics, but we want to be very practical. So I want to show you how you  can actually talk through a full process where at the end you would eventually be able to say hey  this is how you can design and build a prediction

market. I assume that most of you here in this  room are familiar with prediction markets. Uh this is what you see here on the right hand side.  Um this is a chart that is from poly market one of the biggest um prediction markets out there.  And this is um representing two price pass of um the election in 20124 I believe it was um between  uh where Trump um was uh elected president of the US. So what you see here is a price development  of two assets. Um and you see that over time and by the end of of the of the poll when um when  the election actually happened. So November 8th, you see this price convergence towards one of the  outcomes. And this is what most people associate

prediction markets with. So they think of it as  prediction systems. They think of it as betting systems. Um maybe to some extent even gambling  because you can indeed place bets and participate in those in those systems and then eventually when  you are right or you you can even make profit out of it. Uh but the narrative today and kind of  like my selling point um is that um I want to say that the bridge markets are actually far  far far more than that. So they are not simple betting systems. There are actually information  aggregation mechanisms where each interaction by users or by traders is a signal that comes  into the mark market and then there are specific market functions that take in this signal and  transform it into some aggregate metric which is

um in this case the price that can be interpreted  as the probability of the event. So for you to remember one thing uh we are creating assets or  we are creating markets where the goal is to bet on future events on events that are un unsure yet  uncertain and we want to use those price paths of those assets that are being traded as estimates  of the probabilities of those events happening. And I want to show you how the different choice  and the different design of those markets will allow you a completely different interpretation  of what this probability actually mean and how you have to be very careful about how you design the  market because you can get completely different

results or completely different probabilities um  as an outcome. Prediction markets are designed um to facilitate information aggregation. So this is  the goal. We actually want to create a marketplace where the goal is to to aggregate this information  and spit out an aggregate metric. But what we also want to do and this is key because this is the  only reason why those markets have credibility is we want to incentivize truthful reporting.  Which means that we want to design a system where everyone who participates provides their pro true  estimate, their true probability estimate of the event rather than just participating with some  random input because only then the information

that we receive is valuable because only then we  can aggregate it and only then this aggregated metric is also very valuable because we then can  say hey if actually everyone um reported honestly then the aggregate signal is also honest and  therefore we can actually use this price that is um can be interpreted as a probability. But what  you see here is already the first hint at how an economist or an engineer would approach this  problem. Because what you would typically do is you would define some requirements and then  you would try to derive from first principle a mechanism or a set of mechanisms that actually  fulfills those properties and allows you to have a system that keeps those properties intact  and achieves what it what it is designed to do.

So the goal of the workshop today is along this  spectrum. So we want to understand what prediction markets are and what is being traded. What are the  assets that are actually priced there and how do you create them and how do you represent them on  ch on chain. We want to define and compare core market micro structures because I can immediately  tell you that there are multiple ways how you can design such a price aggregation mechanism. There  is not only one economic mechanism that achieves that. There is a multitude of me of mechanisms  that you can use. We therefore want to evaluate those different mechanisms under different  regimes. So we want to look at their performance.

We want to see how the different mechanisms behave  given particular signal inputs and what this means for us as market operators, market designers  and traders. We want to look at it also from the implementation perspective. So we want to look  at it through the lens of a developer and say how can we once we have this idealized whiteboard  solution that we know how it behaves and that we find attractive, how can we actually implement it  and how can we actually build such a system that achieves what we're looking for. And while we are  doing all of this, I'm going also to sneak in a little bit of the engineering workflow so that you  also have a structured process in in at hand that you can use for your future design projects where  you can say, "Hey, I remember this is what we do

did at the workshop. If I follow those steps, I  would actually achieve something. I would result um I would um end up with a mechanism that  has act nice properties that I'm looking for. So first let's take one step back and this might  be quite uh surprising because we are not going to talk about prediction markets at all in the first  two slides. What we're going to do is we're going to say hold back. Sometimes prediction markets  are not the right mechanism for you to use. So there is a whole family of different markets  uh of different mechanisms that you a menu of mechanisms that you can pull out and you can  say I want to build something or I want to use something and I want to achieve this but maybe  what is the best mechanism for my choice? And

this is interesting because right now prediction  markets are booming. And I want to emphasize in those first two slides that um maybe some of the  projects or uh some people building those things uh choose a system or choose a mechanism that  is not tailored to the purpose that it is used for. And we want to anchor this in in a paper by  Canada at all which is published in nature in 2022 uh where the authors basically review um consensus  making and combined decision making mechanisms in human and in animals. And what they say there is  that depending on the information distribution across the whole population. Meaning is it only  one person who knows everything and participates

in a mechanism or is information equally  distributed among group members and also dependent on the dimension whether we are defining  a consensus decision- making or a combined decision-making mechanism. We have a full menu of  different um algorithms and mechanisms that we can choose and that will drive our market design.  In consensus decision-m mechanisms, the goal is to make a decision together. So there has to  be consensus. The group can only move forward if everyone agrees that this is what they want. This  is the left kind of mechanisms that you have here.

And the simplest one if information is equally  distributed would be the unonymity rule which means that everyone has to agree that this is the  choice that they want to make and this choice has to be equal for everyone and then and only then  the group agrees that this is what they want to do. It is a very slow mechanism. So there are  tradeoffs between efficiency and speed of those mechanisms because you can imagine that this  takes quite long for a huge population to to to arrive at one one decision when everyone has  to agree. While at the other hand when you have the complete contrary mechanism where you say you  have desperatism or leadership so you basically select one person from the group and say the  group does what this one person says to do.

And this might be useful if information is um  completely skewed towards this one person. So like if he's the only expert, if only he knows  what is the right thing to do, then it makes sense to select this mechanism. Right? On the and  on the other hand, if everyone knows the same, take unmity in combined decision-m consensus  is not a requirement. So individuals also take individual decisions. the group arrives at a con  at some decision but they don't have to agree. So it can be something like random copying. So like  you do whatever your neighbor is doing and then everyone is doing this for some time and then  eventually the group arrives at one destination and then this is what they have just managed  to achieve and again there is a distinction

between how is information distributed. So now  if you zoom in only on the left hand side on the consensus decision-making mechanism you see  again a set of different methods that are examples of such mechanisms and now finally we see that  prediction markets are one of them. So this is why I'm bringing this slide in. What I'm saying  here is only if you want to make consensus me um consensus based decision- making as a group  and only if there is some prerequisites are met are met then prediction markets are  the right choice for you to implement.

Okay. So now let's look at it. So assume we have  decided that we want to build prediction markets. Um, who of you has seen already kind of  like how they operate or more technically what happens under the hood? The most common  distinction is between information markets and scoring rules. And then in 2003, Robin Hansen  came along who defined market scoring rules which sit in between those two of them. And  we want to emphasize why this um definition almost 23 years ago was so revolutionary and so  important and which important gap it closed. So imagine we want here as a group of five six seven  people make a prediction about the future. We can

create an information market. What does it mean?  It means that we create assets. For example, will the football club in which stadium we  are today win the next match? Yes or no? This is the question. This is what we want  to predict. And what we can do now is we can create an information market. So we can now say  everyone who is here can buy or sell opinions in this market. The problem that you have is that for  such a small group as we have here five six people you will probably not find a counterparty for your  trade. So maybe someone is trying is willing to

sell, maybe someone is willing to buy but at  completely different prices which means that the information density or the trading volume  is very very low which is what you see in this graph. So you see that simple information markets  they perform extremely poorly here on this side of the spectrum. This side of the spectrum is you  have very a high number of estimates per trader which means that you have not many traders in the  market. So simple information markets where you can exchange those opinions and we can trade  those opinions they work extremely poorly in a situation where there is not high volume where  there are not the many people participating. So

this doesn't work. This is unefficient.  It would be cool but it's unefficient. What are scoring rules on the other  hand? Scoring rules are an instrument or a mechanism from economics that has been  used since the 1950s where you can basically score participants individually. So you  don't need to find a counterparty. I would do something different in this case. I would ask  you all individually about your prediction of the future event. And I would I would ask you I  would ask you you you you would report the probability distribution. So you would say what  do you believe will happen or how likely do you

believe something will happen. Then we would wait  for the events to realize and then I would score you hence scoring rules based on the accuracy of  your prediction. So if you were correct if your prediction was good you would receive more points  or more rewards. If your prediction was poor, if it was inaccurate, you would receive less  reward. Those are scoring rules. And you see immediately that for this to work, you don't need  high volume. You don't need many participants. You can score everyone individually. So it works  even with one person, right? The problem is, and you see this here, you actually don't get um  pulled opinion. So it's difficult to aggregate all individual opinions into one market belief. So I  am able to score you individually. I know exactly

what you predict the event probability to be. But  I'm not certain and I cannot aggregate this into a market probability. So I cannot actually say  the market believes that the game will win will end in a win 60%. This is not possible. So we get  to the opinion pooling problem for scoring rules and we get to the thin market problem for simple  information markets and those are common and old historic mechanisms that have been used and can be  used for this information aggregation to work. But in order to close the gap 2003 Hansen introduced  market scoring rules which are basically the best of both worlds. So what happens there is I also  have a scoring rule but this scoring rule is true

for everyone. So there is only one scoring rule  and every one of you can interact with this one single scoring rule and if you don't believe that  the scoring rule is set correctly you can change it but you can change it in a particular cost.  So you have to pay a price to move the market belief to a different uh point of the curve.  What market scoring rules achieve are if you have only one participant so in the thin market  case they collapse the scoring rules because they are scoring rules. So like it's the same thing but  in for high volume if you have many participants they become information markets commonly known as  automated market makers. So this is how uh where amm comes from where you have those scoring  rules and you just kind of like have high

uh market volume. So now you now you already  see that by those different underlying logical algorithms that you can use to make a market  prediction or to pull opinion from people, right? You get to completely different market micro  structures because in this case you get to an information market which is an orderbook  style trading mechanism and in this case you get to an automated market maker a bonding  curve style mechanism. And the goal of our session is to compare those two and see when  does which make sense? What does it mean? What are the consequences? How do systems perform? Um  how do you actually select one of those if you want to build a prediction market based on which  which criteria do you select actually want like

how do you um pri prioritize in a selection and  please remember um we are using those mechanisms because at the very first slide we had those two  requirements where we said we want to facilitate information aggregation those achieve that  and we want people to report truthfully those achieve that as well because scoring rules  and in particular proper scoring rule which rules which are a subclass of scoring rules have this  incentive compatibility property. So the rewards are maximized for participants if they report  truthfully. So for every participant it makes sense to say okay if I really tell you what I  really believe um I will maximize my potential reward. So you already get a mechanism that  elicits truthful information. You want this

check and the other is information pooling.  You can pull those signals into one common market belief. Perfect. This is what we wanted.  Okay. So now you see again family of mechanisms um selection based on requirements and ticking  off all of those requirements to say okay we are on the right path. What are then claims?  So which assets are we actually trading? How do we represent our beliefs? Um, and  what are the things that we are actually interested in finding the price? Um, some  of you might have already heard of arrow de securities which are exactly the assets that  we are trading on prediction markets. They are specific synthetized assets that pay out one  if the event occurs and pay out zero otherwise.

And the interesting thing about prediction  market here is that it's kind of like a reverted logic of designing because typically you design a  marketplace to facilitate an exchange of an asset. So you might have a commodity, you might have an  asset that you want to trade and you want to build the marketplace to facilitate exchange, right?  This is the purpose of the market. Here we're turning the tables because we're saying we don't  care about those assets. Those are synthetized assets. they just kind of like we define them  like that. But um the purpose of the market is to aggregate information and to reveal information  to give us some metric that we didn't know before.

Because we create a market, we create those  assets and by allowing people to trade on them, we will find their true opinions because we will  elicit truth true information and we will pull those information into one common signal which is  the market price which can be interpreted as the market probability of the event realizing.  Perfect. So please remember this. This is quite important. We want to find a venue for the  exchange and pricing of those assets, right? And you you you see here that those assets can be  multi-dimensional. So typically or intuitively you would be thinking about two-dimensional binary  assets. So for example, a yes or no question. Um will Trump win the election? Yes, no. And then  you can say yes, if he wins you receive $1 and

um no, he loses you receive zero dollars. So you  lose. And this asset is being traded um between zero and one um and it always pays out either  zero or one at resolution. So you can immediately see that the price at which it's traded is  the market belief of the event happening. What you can do just um um a mental note because  this will become important. You can always mint a full set of those claims by providing one unit  of collateral. Which means if it pays out one, I can always say I am giving this contract $1 and  I create one yes and one no share because those

events are exclusive and exhaustive. So they cover  the full event space and then know that exactly one of them will happen. So I can actually make  this primary market mechanism of collateralizing um my contract and creating a full set of  shares. All of the events um outcomes can be can be represented because I know exactly that  in aggregate they are worth exactly one. Perfect. Um so again for a bi binary event I think we  talked this through already. We have a yes and no. Pace one if the event occurs, pays one if the  event doesn't occur. And then we it would be very convenient if we want to build it. So like  now let's talk about building it for the first time. It would be very convenient to think about  as funible year C20 tokens. Yes. Because this is

what we want, right? Like we have a yes and no.  Like they all the same. So we can just create a marketplace and then we can say perfect. Uh I know  that like if we trade them separately uh I know that P of yes uh price of yes and price of no will  be exactly one. So like I will even if I create um marketplaces for them I know that like the  people will come in and arbitrage the differences out and you will have this um this parity here.  It's very convenient. It's very nice to to say hey let's build it let's implement it maybe um let's  design a prediction market where we represent yes and no by year c20s and then we just like can  start trading and then we will have everything that we want and it works it's fine um and uh but  it it has some limitations so we will see what the

limitations are and once you have defined those  assets once you have represented them as ERC20s you now have to think about primary markets  and secondary market primary market we talked Um already we said that we can mint a full set of  shares at any time by collaterizing one one unit of of um um by bonding one unit of collateral.  So like this is this is perfectly feasible and the secondary market is now our choice. How  do we design this? Right? Um we are the market designer. We are the market operator. So we can  choose which secondary market mechanism we want.

We can make one marketplace for yes and then we  know that no will be exactly dictated by this. We can make two marketplaces separately and then we  will have arbitrasures ensuring price adjustment or we can even um create both assets and just  pull them together in a bonding curve and and let the market find the equilibrium or or the  market price. Um this is all our choice. We can do all of this. So this is perfectly fine. Now we  are entering the design space and how we want to treat this economically. But the key distinction  that I want to make at this stage is that we have something that we say is the claim representation  and settlement layer. So those ERC20s that we talked about like what is the arrow the broad  security representation? How do we represent it technically? What is it that represents the  claim and how do we settle it? How do we get

our rewards after uh after resolution? And we  have on top of the of the claim and settlement layer we have the market layer where actually  those assets can be traded and price discovery can take place and now we have to think about both  of those things separately. So what is the best way to represent those assets and what is the best  way to find uh to discover prices of those assets and so far we haven't really gone we haven't  gotten really far but um but we have agreed on ESC20 is fine for the representation and um for  trading we can do any of those right okay but um things can get tricky quickly uh they can  get tricky Okay, quickly because um we are

not bound to have a binary market. So like we  can have more sophisticated event bases. So we can have n mutually exclusive and exhaustive  outcomes. For example, you have 10 candidates, right? And you can say which candidate will  win the election or you have something like uh 100 percentage points and the question is how  many percent will the winner of election get, right? And then you can ask any number between  zero and 100. Zero probably not, but any number can be is an eligible number to be to be bet  on. which means that now you are segregating or partition partitioning partitioning the market  um the outcome space into into more outcomes and um all of those those other properties still  hold. So like you now have to have multiple n

shares to represent those outcomes. All of  them have those properties um have this um yes uh pay out one or and zero in in case the  event occurs or not. the prices of those add up to one. So like all of those properties that  we had with two assets still hold. Um and we can design those binary categorical rangebased or even  continuous markets. Like we can get even very very sophisticated by not betting on discrete outcomes  but on continuous outcomes or on parameters of of continuous distribution functions. But the problem  is this does it work with one ERC20 per outcome design? So was our intuition right? Can we use  this? The answer is probably not because it grows in size pretty quickly. So if we have this event  space, we have so many outcomes, then the the size

of the space gets huge pretty large and then we  would have to have so many assets, so many claims, so many bilateral exchanges between them, so so  many arbitrures, so much so much complexity coming in. It just doesn't pay off. It's not worth it.  And it's ex extremely operationally inefficient. So this is why we have a new question. We have a  new goal. Now we have to seek a system where we can have one collateral backing many claims.  Many claim types can coexist in one contract and not one contract per claim. Claims can  be split and recombined. So like we want to have this efficiency and this settlement in the  representation already and we even can represent

joint and conditional outcomes. So we can bet on  things I want to bet on B only if A happens. So more combinational um event spaces and this is  where the conditional token framework comes in and some of you might have heard of it. It's  something that Gnosis introduced in 2017 2018 um where they said um yeah we want to have a token  standard that exactly achieves all of this without getting too deep into the the ERC20 one token  per contract logic. So um what does it mean? ESC 1155 we started here at the ERC20 level.  We had one contract per funible asset but now um we extended it to an ERC 721 where we have  one contract for multiple NFDs. It's still not

the thing that we're looking for because we want  to have one contract for multiple fungeible tokens and we want to represent all of those claims  within one contract and we want to represent it as IDs within the RC155 and allow for all of  the claims and all of the combinations of claims and all of the um complexity that is inherent to  those markets to be represented by one contract and the pricing happening in this contract. So,  so this is a very convenient standard and most of the prediction markets that um are built today  leverage this standard. So, pull the market um is built on um on CFD. Um all all of the currently  this is where kind of like state-of-the-art uh claim representation. And remember, we made a  distinction between how do you represent claims

and how do you do price discovery. So it would be  interesting to see if we can um we can actually trade those assets and how can we find prices for  those assets and the answer is yes. Agnosis event came out not only with the framework but also  they provided two prototype implementations one uh for an AMM and one for a market scoring  rule uh such that there is kind of like an out ofthe-box uh contract and algorithm available  where you can using this standard already um build price discovery on top of it. But let's  go back one step. Let's uh think about what the CFD is. Um it encodes exactly what we want to  encode because it has those primitives primitives.

It has the condition. So the question that we  are asking and we define who is responsible for resolving this question or who is responsible  for providing eventually the answer to this to this this question. We have outcome slots.  So how many different outcomes exist and what they are. Um we can define subset of outcomes. So  we can make conditional or outcomes or we can say we want to bet on A and B. Uh so kind of like  exclude only one option but like bet on on so who's Trump is not going to win but someone else  is going to win uh all of the other candidates.

Um we have the position the collateral we have  the token ID which represents the claim and we define a payout vector that tells the contract um  it in which proportion assets have to be paid out once the resolution is known. So very convenient  standard you can use it and you can build on top of it. When you think now about the life cycle  of a position in this standard what you would be doing is again distinguishing between the  settlement layer and the market layer. So first you would define the position on the settlement  layer. You would say who is the uh oracle? What is the question? What are the outcome slots? Then  uh you define the collateral for the position. you can provide collateral and split as we mentioned  into a full set of outcomes because this you can

do always at any time. Um um so for an end token  market you provide one unit of collateral and you provide a full vector um of of outcome shares.  And then once you have those assets represented on the on the representation layer, you can now  enter the marketplace and now we can say let's find prices. let's allow users to agree on um  on the probabilities by um by mutual exchange. Um so this is what happens next. This is where  the marketplaces come in and then you can again once this is find found you can go back to the to  the CTF layer merge those claims distribute and so on and so on. you uh resolve um the question  when the outcome is known and then eventually users who hold claims can redeem collateral by  claiming and burning those claims. So this is

how the position now works and you can go you can  you can use it to to realize a full market cycle um on the settlement and market layer.  Economically you distinguish between uh collateral positions and redeemed payout. So  first you enter the market with 100 USDC. This is what you have. You provide 100 USDC. You create  positions by minting a full out full uh set of shares and then once the outcome is known one of  them becomes worth and the other becomes worthless exactly the collateral that is behind it. So  this is the economic life cycle of a position.

Okay. So now that we know um what claims are, how  they can be represented, what the life cycle is, now let's step step up again and let's  talk um on the market layer. What I have, what I'm showing here is um a three-dimensional  uh bonding curve where uh you would be able to trade those assets on the current prices  that are represented by the green and blue arrow in those directions by moving  the inventory of the AMM from this point to any other point and by moving by moving  remaining on on the on this cost function.

Um this provides always liquidity and allows  you to basically have this this liquid market uh um and and yeah facilitate the exchange.  So you don't need a counterparty. But this is just an example of saying okay um this  is the automated market maker right. But um as mentioned before we can we can also already  talk about um continuous double auctions which are the ones that you would do on an information  market where you would have an order book based. How do you define prices? Prices mathematically  are defined by the directional derivative in the direction of uh one CL one outcome. So as you  have seen on the previous slide the small arrows

are the prices and you get those directions by  taking this derivative. So the price of the asset E momentarily is by looking at the cost function  and deriving in one direction and asking how much would it cost you to move move the the curve  one unit in this direction. Um the full vector of prices is nabla c. So all of the directions  and then you get to those p1 p2 pn and we already know that they will add up to one. So this is what  we're looking for right. So you can you can derive them individually by taking derivatives on the  curve but um you can also check that uh if you do this eventually they will add up to one. This is  how you would do it on an AMM. How would you do it on an order book multiple price definitions exist.  So um a very intuitive is obviously saying the

last clearing price. So obviously the market price  is the last price that cleared. So like if you found someone to trade, this is what the current  probability is. But if you maybe don't have anyone to trade or like you want to be more formal,  then you can say we can define the mid price as the market price. You just collect those orders in  the order book and say what is the middle between the highest um sell order and the the lowest sell  order and the highest buy order. And then you can say this is the market price. But this can be  completely um imaginative, right? Because if I want to sell at 20 and you want to buy at 80, um  probably probably the other way around. I want to sell at 80, you want to buy at 20. Um it's unfair  to say that the market price is 50, right? Like

like no one is agreeing on this price, but this  would this definition would tell so right? So you can approximate it. You can say, okay, there is  this definition of a market price. Um but maybe this is not an actual market price. And especially  you you get the you see the problem when we talk about this low volume or low liquidity situation  where only us two exist. It's actually not possible to say what the current market price  would be. Right? There is one sell order, one buy order, one midpoint, but like is this  really the probability of the event right now? I don't know. Here you have always instantaneous  prices. You are always able to trade. So like here you actually can find the price. So this  looks very attractive in a low volume low equity scenario and this we have to see right what  what it means. We know that those prices add up

to one. We know that you can normalize them and  then you have this probability definition. This is just a formal way of saying what we already  know. So we don't have to look into this too much. Perfect. So now we know um kind of like the  primitives that you would use to design a market. uh we know so bonding curves or like market  scoring rules and um and uh information markets we've talked about the representation of assets I  wrote the bro securities we talked about a token standard that is helpful in implementing them  the conditional token framework and we talked also about definitions of market venues or how  you could actually define prices how you can defi derive prices how you can facilitate information  aggregation or price discovery given that you have

this situation. So now let's talk about what  is the minimal implementation of a prediction market. So like what do we actually need out of  kind of like those elements that we've introduced to build our marketplace and let's start maybe  from here because this is what we already know. So we have the substrate for event contingent  claims some implementation some representation on chain of those claims of those assets. Voila,  you can use the conditional token framework. We haven't talked about this at all because the  presentation doesn't have enough time to discuss this in more depth, but um somewhere in between  you also need an oracle. So you need someone who is responsible for the resolution of those  markets. So not only providing price discovery but actually injecting the true outcome or the  true realized um event outcome into the contract

such that the contract knows um how much to  pay out to whom because this is all encoded but it needs the flag. It needs to know what was  actually realized what is the outcome and then it can derive all of the payout rules and all of  the claims that achieve that. This is what you also need. So you need to have something it can be  centralized which is maybe a bottleneck but also it can be very sophisticated decentralized oracle  um for market resolution. Then you have to have trading mechanisms which are exactly the market  venues the marketplaces that we've introduced that we've talked about and then finally uh you need  something that the user can interact with. So a front end something that um connects to those  contracts and where positions can be created, positions can be stored, uh assets can be managed  and so on and so on. And if you implement this,

you have a prediction market. So you have  something where users can log on then they can interact with the contracts, they can interact  between um their um holdings. So they can they can actually trade assets um on the protocol. there  is something that tells the contracts what the outcome is and how to resolve this and you have  the asset representation. This is what you need. Those are the building blocks and we've covered  actually quite a bit of them. When you look at the smart contract, I'm going to just go through  those slides very quickly. When you go when you look at the smart contract layer um this is what  you need to build something to represent events create markets trade shares resolve the market  distribute winnings and manage the state. So those

are actually the elements or the contracts of the  protocol that you're going to build to handle all of those things. You can think about the more  sophisticated middleware infrastructure. This is not required in the minimal implementation but  you can think about more sophisticated solutions uh including wallet management, a transaction  engine, backend services, everything that makes the system more efficient, more convenient for the  user, more performant and then on the front end finally um cool features or cool ways for users  to interact with um manage their holdings and facilitate transactions. So those are the things  that you would need. Okay. So now let's look at um what I promised. So now let's look at the  implementation of um at the comparison of

the performance. So um I want to bring  your attention to to this chart here. And if you have followed the conversation today  very closely, you will immediately see what's happening here. So this is a price chart coming  from one simulation model where the same question was answered by agents that were participating  on two markets simultaneously and one of them was a bonding curve market and one of them was was  an order book market. And what you see happening here is that for the market scoring rule for  the bonding curve market you have very low uh volatility and you have continuous pricing and  you are within the band of let's say 45 to 43

to 48 something like that. So like very close to  the to the to the midpoint of 50. If you look at the at the box uh box uh plots, you you see also  very low volatility and even for sim multicolor runs of this experiment, you stay close within the  range. So what what does it tell you? It tells you that this market has continuous liquidity. There's  very good price discovery. Um there is there are no major shifts. There is no big price jumps.  Quite the contrary, you see on the um CDA side, we mentioned this that here um the price only  moves when either a trade occurs or when you kind of like have an agreement between different orders  placed on on the um on in the order book and you

can say okay there is actually something happening  that drives those prices and those events are much more selective and they are not that common.  So you see that for example here very long time nothing happens with respect to the price  but obviously signals are coming in. So people are placing different limit orders but they're  just not realized or they are not changing the mid price because kind of like those two orders  are still pending. And then when potentially an order comes in that is clearing the market, um  you see a huge price jump immediately because you kind of like um get this liquidity immediately  sweeped out of the market. And performance-wise, what you see happening here is that um let's say  at step 20, you have a very high discrepancy in

prices between those two markets and you now  have the problem of um the interpretation of probabilities that we started with initially,  right? Because imagine that you have started um you have created a market that is um an order  book market. What would it mean? it would it would mean that the probability of the event happening  is let's say I don't know 25% on the order book but if the same thing was h was traded on a  continuous market you would be around 50 which is kind of like twice the probability right so purely  based on the choice of the market micro structure you have a completely different aggregate  aggregated metric it nothing else changed So agent beliefs are the same, agent um signals  are the same. It's just kind of how the market

translates those signals into a pulled opinion or  a pulled aggregate. This is completely different. So you immediately see how you can drive um the  market metric the price or the probability by the selection of the micro structure. So you have  to be very careful what you do. Um and obviously there are guidelines how you would do it because  we would imagine that um those effects will would become less extreme or less prominent if we had  higher market volume. So if there is more activity on this market on this order book market it would  probably converge to something similar than than the continuous market but we're not quite sure  and um in in the literature or the academic way

to approach this uh to answer this question is  you can act there are actually there's a bunch of papers that is comparing um market mechanisms  and you can do it in three different ways. So you can look at it from the Xandex interim and exposed  perspective which is nothing else than saying um exposed how well did the market perform overall.  We uh we look at metrics like price deviation, welfare, trader profit. So we look at something  completely that is that is not the focus on on of our conversation today uh because we have  not uh put any attention on on those questions um right now but they are quite important as well.  So like you can perform this analysis you can say okay how well did the market perform we define  those metrics and then we see how it goes. But

what is more interesting to us today are those  x and x interim evaluations where you would here in this case um look at um the mechanism and you  would try to evaluate which mechanism produced the most accurate forecast. So uh you would u wait  for the for the event to to occur. You would see the market resolution. This is a requirement  for this uh for this analysis. And then you would score the event outcome against the submitted uh  forecasts and you would be able to see okay how did the market perform that we've run on the um  on the order book side how did the market perform that we've run on the um on the market scoring  rule side and then we can compare which one was better right like we can do an x ante evaluation  but uh kind of like obviously after the fact after

we know how how the market market went we can we  and see which one was better. The more interesting approach um if we have not resolved yet so like  during an ongoing market is if we look at an X interim evaluation and we would be asking  the question of which mechanism incorporates information faster. Um so we would be looking at  signals coming in. We would have both markets run in in parallel and then we would just look at  them and see okay where does the information come in quicker. So like which market translates  this information immediately into something that that we can um that we can uh interpret as um  as an effect on our market price. And the bottom line of all of those events um those um this  research and those analytics is and they are quite

um they quite agree on on this on this bottom  line is that um there is a threshold. So there is a particular number of trades or a particular  depth of liquidity that that you require until which it is better to trade on a continuous market  a market scoring rule and above this threshold or above this this volume you would be better off on  the order book and uh this is quite clear. So like there is there is a particular threshold and the  recommendation is um before you set up a market before you choose the market micro structure what  you do is you try to estimate how much trader per trader activity you will have how much interest in  this market you will have how is the information

distribution um across participants in the market.  How many experts exist worldwide? How many people do you think would even participate? And if you  feel like this is something that will draw a lot of attention and a lot of volume then you can  go with an order book and this and then you are completely fine because those markets are cheaper  more efficient and and quicker for high volume. If you feel like you will not get much attention  and very difficult to get even kind of like first trader activity going on then you should start  with a continuous market and in between um the proposal is to look into hybrid approaches. So you  start with an automated market maker until you get a little bit traction and then you transition to  an order book. So this is maybe the answer why

um on poly market you can trade on events where  basically everyone is an expert. So like everyone believes that they have a good opinion. for sports  event politics where you have very high liquidity and very easy questions to answer right like who  will win uh what will be the score when will be the goal whatever where people just like um are  interested in participating because it's for fun and it's um and they can they don't have to think  much about uh whether their opinion is valid or not they just participate and those markets  operate they work but their business model is from generating fees so they want to kind of like  have this um this interaction with users and they want to to to um to earn money based off of this  activity. What I was trying to lead to today is

that um prediction markets are more than that.  They are not uh purely um betting systems. They are also information aggregation systems. And if  you design them wisely or correctly, you can even use them in niche markets. So you can use them  in a situation where only a handful of experts exist. for example, five or 10 people worldwide  that would know something about an event. Um, and you would very much like to get this  information out from them. Uh, but um, on an order book market like you would not get this  price finding um, clearance, right? So you would get into the problem that we had on the simulation  side, right? Like you would probably get into into this here. And this is why for such use cases it  is um advised to create maybe a market that is

um continuous that allows uh all participants  to trade immediately and discover prices by uh by active engagement. The downside of those  continuous markets is that they require subsidy. So they need to be funded. So it's costly to  operate an automated market maker. And there is um or a market scoring rule um better and um there  is a bound a lower bound on um an upper bound on on losses that that can incur. So you have to look  at those u systems quite differently. So you have to look at them through the lens of an information  revelation system. You want to incentivize people to participate. You want the experts to share  their opinion and you have to pay them for

revealing information. So you have to subsidize  them. You have to give them something that they feel is worthwhile for them to to capture by  correcting the market state to a belief that is more accurate with their predictions and that  as the externality of the system you actually create information. So you have created a market  to discover prices to discover probabilities. So, so this is kind of like the the bottom line of the  presentation that prediction markets are more than simple betting systems and based on the micro  structure that you choose, the implementation that you choose and so on. Uh you're completely  fine with defining those systems. What I did and this will be the closing line is oops. Yeah. So  uh you see here this is an experiment that I ran

uh where you don't you can't read it pretty  well here but um what happens is that those are all different market micro structures  so different um market scoring rules with different parameterizations and different um um  functional forms and you see immediately that when a signal comes in they translate those  prices completely differently and and this is maybe the nice thing to see that uh depending  on what you want to achieve, you can select from a menu of different market designs and then  get the realization that supports or to achieves um um your use case that supports your use case  that you want to look at. Perfect. So this is um everything that I wanted to show you. This  is the literature that I'm using. Um this is

the the publication in nature where you would  be able to step to take one step back and look at the information aggregation in general  and see whether prediction markets are even um the right mechanism to choose. Um if you want  to look into comparisons those are the three ones that I recommend you the most because they compare  between two market structures. um one is ex post, one is ex x exanten, one is x interim and then  you have those two seminal publications Hansson um 2003 2007 where he introduces um information  markets and and um market scoring rules uh very important piece of work. Perfect. Thank  you so much for um for your attention. This is the end of the formal u part of the talk or the  workshop, but I'm happy to stay here for the next

10 minutes or so to get a more like conversation  or interactive um flow going. Thank you so much.