Clément Lesaege - Futarchy and How Prediction Markets Help Us Make Better Decisions

So beforeh going to futarch we have to first uh study prediction markets. How how do they work? How do we structure them? So very simple prediction market that we had for the US election. So we ask who will win the US election and we can tokenize both outcomes.

So you have some Trump token, you have some ARS token and if Trump wins the Trump token will pay out. If ARS wins the ARS token will pay out. And if you look at the price of the Trump token you can deduce an implied probability of Trump winning. So that gives you probabilistic predictions. uh but we can do even better than just uh look looking at what will happen.

We can look at what will happen conditionally on something else. So here for example we had a market on the bitcoin price if Trump wins or the bitcoin price if aris wins and we could see that the market was giving a higher price with Trump winning than Iris and when Trump won we also saw this materializing in the markets. So let's go on the structure of a prediction market. Basically you have a a market which is a questions. So you put some underlying token for example die and now you're going to get a complete set of all the potential outcome.

So here which party will win you get one token for the democratic party one republican token and one other tokens. People can trade those like order books AMM OTC like any way you want. uh and you can see this link between the token price and the probability of the event happening. If we don't just want to trade outcomes of like singular outcomes, we can trade values. So let's say we want to predict what will be the inflation.

So we create a scalar market. So we're going to ask what would be the inflation in 2026. We put a range 0 to 10. And if the inflation is lower than the range, you will have the down token which will redeem. If it's higher than the range, you will have the up token which will redeem.

And if it's within the range, which it should be, if your range was good, they will both redeem proportionally. So here, if you have like 8% inflation, the up will redeem for 0.8. So if you get the price of the up token and you assume that the value will be within the range, you can get an estimate of the value. So here up token is at 0.

8. So the expected value of inflation would be 8% uh with this market. So no contingent markets. So you still have your first market here. You put some die, you get a complete set of the potential candidate of the Democratic party, uh Biden, Aris, uh someone else.

But now with the Biden token, you can put it into the second market which will ask which party will win the election. And now you have a token which is a Biden Democrat token which redeems if the Democratic Party wins and Biden is a candidate. So if you look at the price of this token in comparison of the Biden token, you get a implied probability of Biden winning if he is a candidate. Okay. Okay, so now you have all the basic tools and let's see how we can put all of that together uh to to make better decision and to make some futarchy.

So first like short history of governance system. Um basically you have two issues in governance system. One is a bias which is the system targeting something other than the common good and then you have the variance basically is a system good at eating what it's targeting. So autocracy like you know like uh regal monarchies, dictatorships uh they tend to be the worst for both bias and variance because they have eye bias. They target what the autocrats want and they have I variance because autocrats are alone.

They don't really they cannot know all the topics. So they're not going to be very good at eating what they target. So you have the worst possible. Then you can get into an oligarchy. So in this case you will have some elite class which is going to make the decisions.

And here you still don't have a good uh bias because you are targeting what's good for the edit class. But you more or less solve the variance problem because you have multiple decision makers. They can talk with each others. They can find who is the best at this particular topic. And in this case you will have a lower variance.

Now if you get into uh a democracy where like which will be a direct voting democracy well you will have in this case no bias because people will want what is good for them and most of the time the combination of what people want for themselves uh will lead to something pretty close to the common good but then you still have a eyeiance because obviously not everyone can be knowledgeable on each topic. Uh not everyone can spend a lot of time uh studying each each question each proposal. So you will have a high variance because you're not going to be very good at at targeting at sorry at eating what you target. So the question is can we uh create some newance system uh which can solve this and get both a low bias and a low variance. So one thing you can notice is that often in the political debate people tend to agree on what they want.

Like you ask what do you want? They say wealth, health and happiness. Uh so people tend to agree but then when you ask them how can we get that that's where people disagree like less taxes higher taxes sacrifice a good goat um you have each debate in general on how to get what we want while what we want tend to be more or less consensual. So yeah so can solve that and can we both get a low variance and a low bias. So the idea is to split the decision process into two phases.

The first one is voten value. So we vote on what we want. What do we want to optimize? Do we want to optimize for the wealth of the country? Do we want to optimize for the wealth per capita?

Do you want to optimize for some other metric like healthcare metrics, HDIs? And then we're going to make market where we're going to ask if we pass this proposal what would be the value of this metric and we pass a proposal which will lead to the best metrics. So let's assume that you are a single uh topic voter. You are only interested by the bitcoin price. You don't care about anything else.

You are in the last US election. You look at the markets and you say, "Yeah, I just want to increase the price of my bitcoin." In this case, you can vote for Trump because Trump will be the one which will lead to the highest expected value of your bitcoin portfolio. Obviously, in practice, uh people tend to have a wider range of of stuff they care about. So how can we structure those markets?

Um so let's say we go back to the election. Let's say we are the Democratic party and here what we want we want to win the elections. Uh we have a list of candidate so Biden ARS some others. Uh we have the first market we which were candidate will be the democratic candidate. We get the ARS token the Biden token some other token.

And now we put those into the second market uh which ask which party will be in the election. And so now we have a token which is a Biden Democrat. So Biden is a candidate and the Democratic party wins. We have an Aris Democrat. Aris is a candidate um and Democratic party wins.

And when we look at the price of the Biden Democrat in respect to Biden, we get uh that it's 30. So you will have 30% chance of winning if you put Biden as your candidate. And you can do the same thing with ARIS. In this case, you have 40%. So it tend to be better.

So you can switch your candidate to Aris. You still lose but you had more chances of winning putting Aris than keeping Biden. But in practice most of the time we want to optimize on metrics not just outcome. Uh and well remember the scalar markets. So we can combine that with scalar markets.

Uh for example we could have a metric footarchy. Uh we could ask which project will get the highest will bring the highest amount of users if we fund them. So those can be uh communication project and uh we want to optimize the amount of user in our platform. So you have all those communication project which come they ask hey like can you give us some money and we're going to do some communication for you. So you put your die you get to kind of each project and now on each project you ask if we fund this project to communicate on our project what would be the amount of user that we will have like in six months.

So people can trade they can trade up token they can trade on token. Uh when you look at the range and the price of the up token you can see the implied number of user that you will get. So you see that the project B is the best one. So you give the funding to project B. Uh and at the end in the next six months you can look at the amount of user you have and you can resolve those markets.

So note that as a trader you can use the same fund to trade on all of those market at once because only one market will be relevant like project A C and D because they are not chosen they are not funded all the trade which happen there they have no practical effect because they all redeem for uh like token which we do not redeem for D only the B token redeem for die. So only the trade between uh B up and B down will matter in the term of profit and loss of the trader. So we can go back on what to optimize. So already talk a bit in the case of state like GDP SGIS. Um but now if you look at crypto we have three main metrics that crypto project like.

So you have obviously the amount of users uh which sometime may be a bit hard to get but you could always like run some studies to find how many people are actually using your product uh because you want to try to avoid to do double users which are using multiple addresses TVL that's still a pretty nice metric even if one of the previous speaker was proposing a metric order to to bias and the fees uh which your project is So it really depend of the project. If you are a lending platform, it can be lending fees. If you are L1 or L2, can be the fee given to the sequencer or the fee burn in the case of E is um but you can also sometime want to optimize and which may be subjective. So there is all this work which is being done currently on impact evaluation. So here you don't really have a objective metric to resolve to but you will use a committee you will use different sort of mechanism to evaluate the impact of the project after the fact.

So it's still subjective but it tend to be uh way easier to have a consensus on that because it tend to be easier for people to agree on the value provided by a project compared to the potential value that this project will provide in the future. So we know some project are using retropf like optimism. So you can just uh you can just use that you can just use the retropgf uh as your impact metric or you can develop your own impact metric and in this case you ask if this project in is funded what will be the value of the impact metric and you you find the project which will have the highest impact. Um another way of doing futarchy is what I will call price based futarchy. So here uh we have two tokens.

So you have one token which is relative to the project that you want to make decision for. So in this case you have the go token. If you ask knows this we have a token that we call the reference asset here it's it's die which is more or less $1. And now what we can ask people we can ask them to trade in the child universe of the proposal passing or not passing. So how do we do that?

So you can come with GNO. You put it into the market and now you get yes GNO and no GNO. So the yes GNO redeems for GNO is the proposal passes. The no GNO redeem for GNO. If the proposal does not pass, you can do the same thing with D.

You get yes D and no D. And now what you do you can trade yes GNO for yes D, no GNO for no D. So basically you can do some buy and sell of go and die conditionally on the proposal passing or conditionally on the proposal not passing. Keep in mind that you are not exposed in any way on the likelihood of the proposal passing or not. You can be completely neutral on whether the proposal will pass.

You are not betting on a proposal passing. You are just doing conditional trade on the proposal passing or not passing. And now you can look at the price of the yes go compared to the yes die. And that's going to give you a conditional price. So you can see here that is a proposal pass.

You have a conditional general price of 122 die. If it doesn't, it's 131. So here you could conclude that if this proposal passing would be bad for the price. Uh and in this case, well, not pass it. So here I'm I'm putting that in the context of crypto with crypto tokens.

But you could do something similar with with stock uh for for publicly traded companies. So we can see how it looks like on an interface. So that's done by target five. Uh you can see this graph. So people are doing some trade.

So you can see like the if yes, you can trade. If no, you can also trade and that gives you those two conditional price. In general, you tend to have one higher than the other and you tend to have the spot price uh in the middle. Uh and towards the end you also tend to have the spot price which tend to uh reach uh the price of the decision which will lead to the highest price. uh if people assume that the decision maker are going to follow the market.

Um so that's how it's look like in practice. Um so now how do we how do we use that? Uh so we can really go from anything like it can be simple advisory to some fully autonomous system. Uh you can really go full you know like we have this project which is on Solana which is metad so it's a DA which works only on futarch. So you have some tokens.

It's trying to optimize the price of its own token. But there is no vote. Like you have no delegate. You have no one voting on anything. Anytime you want to make a proposal, you make a proposal and people can trade this metad token conditionally on the proposal passing or not passing.

Uh here you can see that the price if approved uh is better than the price if rejected. Uh so the proposal ended up being approved and executed. So that's very interesting for me. It's also a bit scary because it's the first time I see something on Solana which is more advanced than ERAM but it's very interesting. So advisories that has been tried actually even in 2020 by Nosis where they had done this impact module.

So this is basically just a module you put on snapshot uh and you can see the impact on the price of the token of your proposal. So it can just be used as advisory. So when you are voting you can take that into account. Maybe you may still want to vote the proposal even if it decreases the price because maybe you want to do it for ideological reason or you think the market is wrong. So you can still vote whatever you want but you get some market feedback before voting.

What you can also do uh you can also delegate to futarch. So you can have like a food address which is already going to vote the way of futarch and you could say hey I'm gonna put like I don't know like 10 or 25% of my tokens that I'm going to delegate to this nonhuman delegate which is just following foot advice. Um then we have specific application of futarch. Uh so you have this idea of the futarch optimizer. Maybe you don't want to put your world d into futarch but you can take some of the asset of your d it can be the d token it can be like e or tables that you put into futarch optimizer.

So this is a subdow uh you give it a target metric. It can be increasing the proposal of your D or it can be increasing the number of user the TVL whatever you want give it asset and now like anyone can come up with some proposal you will have some futarchy uh period where people can trade and then it's going to execute the proposal if it will lead to a higher value of the metric or the asset you want to optimize for obviously when you have a proposal passing you may say oh yeah well it's already going to be good you know like if I want to do if I want to increase the TVL. Obviously, spending the money is going to increase the TVL, but you have to keep in mind that if you are spending the money in the optimizer, well, there will be less money for the future proposal. So, market participant should take into account the value created by the proposal, but also also the value destroyed by the proposal because that's less assets for the next proposal. uh and we could even argue that's even closer to real DAO because what we've been calling DAO will tend to be organization which are still humanled uh but here you don't have much human decision you may have some human decision in the term of traders but it's not directly a decision it's more like a a trade uh and know like if you get the AI agent like know you could even have organization which work with almost no human involvement at all um then you can have specific application So for example here with CLOS we're going to we're going to run this where we want to have some project to be creating TVS to value secure for CLOS and so the project can come they apply uh and know we have markets which ask if we give an investment into this project uh what will be the amount of TVS that this project will bring uh people can trade and in this case we will give the the investment to to the project to the purple project because the market estimate that it will give five million of TVL which is significantly greater than the other projects.

So yeah, so thanks for for listening and we're gonna go to to the questions. Thank you. Thank you everyone. Do we have any questions? No questions on the board.

Question over there. Hi uh thank you for the presentation. Uh question so I think I understood how it works more or less but is it not very easy to manipulate the outcome if uh right we have these small tokens for single very specific use cases that might have low liquidity I guess so it's very easy to manipulate the price of a yes or no outcome on future key on the future key system. So if you're trying to manipulate a price uh you're basically acting de facto as a bad trader because most traders they try to optimize their profit but here you are acting as a bad trader because you are trading uh let's say you think that it's going to get uh 5 million of TVS but then you try to put it down and you're going to trade it and sell it at uh so that it would estimate 1 million of TVs. So you are not trading the value that you are currently estimating but you are trading a value to try to buy the market.

So the thing is that if you do that you are creating an opportunity for good traders to go counter trade you like they can take the other side of the trade. Uh and by doing it you are acting as a de facto subsidy to the market. So like if you have a market and a lot of trader which are bad which can be bad either because they have other intention or they can be stupid like it doesn't really matter. Uh well that just create more opportunities for good trader to come there trade against them and make some profit. So manipulator uh will actually make market more efficient because they just create a more positive s game for the people who are trading and are not manipulators.

Okay. But if we use that futarch key system to actually make the decisions, right? If it's if I'm, you know, if I want to push for an approval to be passed, then I might buy the futile key vote saying yes, right? And that will kind of bias the vote towards. So it will bias the actual outcome of the of the proposal.

Yeah. But then anyone can go counter trade you. And and if you actually try to do that at scale, that's even more attractive for people to go and count trade you because they have a lot of money to be made by counterter trading you. It's not a close system like well you know there is a amount of dye sure but in practice even if you're a manipulator you will only have a very small portion of the dice supply uh and good traders if they see this kind of opportunity of things which are highly mispriced uh they would just come there and trade you. And you you could see that in in not even food just in a regular market.

If you can look at the election market as soon as you have something which is clearly mispriced where trader are going to come because they have some opportunities and they are putting back the price uh to more reasonable values.

Do we have any more questions?

Gentlemen at the back. Here we are sir.

Thank you. Uh actually I want to plus one the question that was just be before me because something in your presentation I didn't quite understand is at the beginning you sound something very general generic like we could do that for policies asking what wellness is or how to best vote on a policy but then your answer just now is like it cannot be general because if I start doing photography decision then it's easy to skew the system and you cannot really counter bid if it's like a general we're always going to follow the market and I had the same so I question my first question was One one question by one question by one. You can ask another one after so I think you may have misunderstood one point uh which you are not uh trading on which decision will be made. You are trading on uh what will be the value of a metric if this decision is being made. So as a trader uh you should not really be concerned about what the decision is going to be.

You are only to be concerned about what will be the impact of the decision. And with the same uh money you can trade in the universe of all decision being made at the same time. You can say okay like if this is this project that will be the the TVS. If this is this project that will be this TVs and you can use the same capital to trade on all those universes at once. But at no point you need to have an exposure on what will the actual decision be?

I understand that. Wouldn't that mean that the actual decision needs to be decorated by the actual photography? Because if it's automatic decision based on what the market would think, then there's no the correlation between what the decision would be and how I can influence the decision to be which I think was the question before, right? Yes. And I also do not I don't know how to match these two points for me.

They're contradictory. So wait, first imagine the decision is independent. Are you okay in this situation? H.

So your concern is that if the decision depends on the market, right?

Yes.

Yeah. So the decision can depend on the market but the impact of the decision will not depend on the market. Well, at least not significantly. So there is obviously uh some sort of um um selection bias in that if the project C is the one chosen, it's chosen because it's the one which will lead to the ISTVs. uh so it's going to be more likely to lead to higher TVs because that mean that the other traders would have thought that it would lead to the highest TVs but this is something that trader can take into account.

So they can they can take this bias into account when they're trading and even this kind of bias uh tend to apply to all potential outcomes equally. So it's normally most of the time it shouldn't change the order. I know that you can construct very specific edge cases where it could change the order but there very specific edge cases and it's very unlikely to happen in practice but or maybe your issue is that you could badly trade on decision which are not going to be chosen. So you could also say yeah you can badly trade on decision which are not going to be chosen but then it doesn't do anything because the decision are not chosen anyway. So trying to bias some market which are conditional on the decision which end up not being chosen has no impact on the system so it doesn't cost an issue either.

Oh I think he had a second question.

Okay let's let's give someone else the opportunity to ask a question or we have one question on the board as well unless uh

okay

how accurate are production market in your opinion? uh I would say in the vast majority they are pretty accurate uh there are indeed some edge cases which tend not to be well covered in the current implementation uh for example you can look at the JUS market where it was trading that J is going to come back at 2% uh but if you look more deep into this market you can understand why uh so there are a few reason to that the first one is actually the market is not estimating the likelihood of Jesus coming back but it's estimating the likelihood of the oracle hum that they just come back which is significantly more likely. It's way more likely that the oracle is wrong and broken than to actually judge coming back. Uh so that's one part of the the bias you will have and that tend to be showing in events which are very unlikely. Um and another part is also that current market so I'm talking about poly market cash all of those uh they are using uh dollar as the base asset.

So even if you go and you see oh Jesus J is coming back is at 2% I could just buy the no token of that and get 2%. Yes you can but 2% in six months that's basically 4% per year. Uh you could basically just put your money into ESDA instead and get 5.5%. Uh so you may have some issue with cost of capital and this one can be solved via having good underlying uh which is the beering.

uh you have issue with the oracle which can be solved by having good oracles and yeah hi I'm keros making good oracles um so even if you have a low cost of capital you may still have issues because trader may want to do uh more stuff with their money instead of trading on this uh on this market which uh is very unbalanced and uh we have ways to do that also like we have multicategorical markets uh that we can use uh to be able to trade on the long tail of probabilities. You could also have combin combinatorial markets which can do that. So I would say in general prediction market are good. You may find some edge cases where not the correct structure was used and which could lead some bias but in general you I think most of the time you have a way to make a good and accurate prediction market.

Thank you. Do we have any more questions? One question. Thank you. Um so stepping back a little bit about the trading topic of prediction markets.

What is your opinion about actually using the accuracy of prediction markets in detecting basically biased votes or outcomes? the ones that were tampered with or the ones that got for example in a in a voting system for example um let's say that there there was a vote that based on prediction market should have come out at a specific range but then the outcome was like drastically different. Well, maybe you could use it uh as something to, you know, like specifically watch this particular vote. Uh but often you will have votes uh which um may go against what the market may have looked at the most likely outcome uh just because the people who are trading may not really get much information. Uh so for example you had markets about the papal election and you probably had a lot of people who were very knowledgeable about which one was likely to be elected as the pope.

Uh but those they will be the cardinals which were first ethically probably prohibited to trade on those market and also even physically prohibited to trade on those market because they were in the punk lab. uh so in this case you only had external traders which had very little information and in this case uh well you could not have a good outcome. So prediction market they do not create information prediction market they aggregate information and so if you have no information to aggregate in the first place uh the market is probably not going to work or if the market is large enough attractive enough in this case you can create incentive to to create the information. So in the case of papal election h it's pretty hard. Maybe you could have I don't know like if it was a very very big market you may have some crazy guy trying to spy inside of the conclave.

Uh but more realistically uh for example in the US election you had a French trader who ended up um commissioning its own neighborh poll. So they were basically asking to people who do you think your neighbor would vote for? uh which is something that traditional pollster did not do but because this guy had a lot of money that he wanted to trade on those market for him it was worth it to pay people to make those polls because it could help him trade better. So yeah, most of the time you aggregate information and if you make a very very big market, you may even create incentive to to create information.