How My AI Agent Discovered a 993% Return Strategy on Polymarket

[00:01] wanted to share this video about my AI agentic research uh in trading I've been running lately. So, I set up this earlier this week like a 24/7 running AI

[00:13] agent. I was looking for interesting strategies we can dive deeper into on I targeted poly market but it can be any market you want to look into. So, you can see here we had a good uh results. We found a strategy that uh in some

[00:28] We found a strategy that uh in some types of events can give us like a 100% guaranteed profit if we can write the right execution and we have the right conditions to enter this type of trade here. So I wanted to talk a bit about h

[00:42] why this works and kind of show you a bit how I set this up because it's not that hard to set this up if you use like codeex cloud code kim K3 or any other

[00:54] like AI agentic setup. So hopefully this can give you some inspiration how you can start looking for this interesting strategies yourself. So let's just head over to poly market and I will kind of explain thish strategy here uh first. So

[01:09] here you can see on my account uh the four trades, right? And they look pretty strange, right? So you can I guess I can zoom in a bit. So you can kind of notice right away that uh we have will reformation market caps between 1.2

[01:24] reformation market caps between 1.2 billion and 1.4 billion close on IPO day at market close I guess. So you can see we also bought no for 4 and a half cent we also bought no for 4 and a half cent 42 shares. And if you look here, we have

[01:38] the same on the opposite side. So we also have yes at 42.5 shares. So this means that this has to resolve, right? So let's say this resolves to yes now,

[01:50] right? Uh and if you go back to our drawing here, you can see if we look at will uh resolve. So if it resolve to yes, you can see we put in 190. So this

[02:02] will be like minus $1.9 right but it this has to resolve to 100. right but it this has to resolve to 100. So this will be like plus is it $4250

[02:14] uh of course minus what we put in. So this will be profit around like $40 38 I guess or something guaranteed because it has to end either up no or yes. So this

[02:26] is like an arbitrage. So this has to be like uh at the end uh plus 38 guaranteed other side. I guess it's a bit more if

[02:38] it resolves to yes, but basically we kind of end up with two guaranteed $38. So this gives us like a 70 something around $75 guaranteed uh profit. So these are the types of

[02:53] trades we are looking for. So basically I built like an automated system. You can call it like a bot to look for exactly these types of markets on the exactly these types of markets on the poly markets. So this is running 24/7 to

[03:07] look for exactly conditions like this. So we can enter on both sides and we can kind of secure the arbitrage. Uh there's also something called merge. I'm not going to go into that. This means that we can kind of capture the value right

[03:21] away. So we don't have to wait until this has resolved. Uh it's a bit technical if you haven't been on Poly Market uh before, but I just basically wanted to explain what we are looking for and what the agent found for us. Uh

[03:35] for and what the agent found for us. Uh it's not like a unknown strategy, but uh I'm going to talk a bit about execution and this is really important in cases like this. Uh I'm going to talk a bit about that afterwards. But let's just

[03:48] head over to Codeex and I'm just going to talk a bit about how I set up this AIC research to find these types of opportunities. Poly Market homepage. Let's say you selected Poly Market for your platform

[04:04] to do research on. It could be anything, but in our case, let's just go with Poly but in our case, let's just go with Poly Market. And you see they have some new interesting things they are promoting here. They have something called combos.

[04:16] combo. So they have combo cup or look at what that is. And let's say today we wanted to do research on kind today we wanted to do research on kind of these combo markets. Okay. So what I

[04:30] of these combo markets. Okay. So what I uh want to do then if I want to kind of set up an agent to just start looking for this um basically I want to do uh a few things before we actually run the prompt. So, the first thing I do is uh

[04:45] I'm going to go into codeex. Uh yeah, for this test I'm not going to pick the max model, but I would at least pick like high or something for reasoning. Maybe for like a real run, I would probably pick like max or something for

[04:58] reasoning just because it's a bit of like math and stuff. So, but for this uh example, we can just set the model to high GPT 5.6 so you can do fable, you can do claude opus 4.8, eight, Kim K3 doesn't really matter too much because

[05:14] all models now are really good at coding. They're good at data analysis. really matter too much what kind of model you use. Just use the one you feel comfortable with and you have like a subscription on or something. Try to

[05:27] avoid using the API because just gets so expensive and you kind of need to have some return of investment, right? So, uh I'm going to select the this one, right? And the next thing I'm going to do is I'm just going to go to for my platform

[05:42] selection today. It's Poly Market. They have free documentation. They have a free websocket. That means that we have a access to a lot of data. And that is kind of the what we need to do. We need to have the data. Without data, we can't

[05:56] really do anything, right? So, I'm just going to copy this quick start to the API here, right? Uh just need that just to point it at the documentation. I know like GPD 5.6 will probably find the page for documentation

[06:11] uh itself doesn't really matter too much. So what is nice here we can run experiments uh without having an account without connecting to our account or nothing like that. The API is free and we can just yeah do some research

[06:25] without doing that. So, I just want to show you kind of like a simple prompt we can use now to just start doing some research on the combo markets uh for um

[06:37] uh on poly market. Haven't done it before. So, let's just see what happens before. So, let's just see what happens here. So, I'm going to start now with something called slashgo because I don't want this to stop. Even though I am in

[06:50] like yolo mode, I still want to do a slash goal. You don't have to do that, but for me, I kind of do like a lot of parallel stuff and slash goal kind of works great. Then I know it's not going to stop, but then I need to be specific

[07:04] when I want this to stop. So, let me just do like a simple prompt here. And yeah, let's try to run it. Okay, so I kind of did this on the fly now, but this is like the gist of what I run. So, I'm going to read through this. Uh, if

[07:18] you can just skip ahead. uh but basically SLG goal we are doing our research on on poly market today the documentation you need for this is what we linked we are targeting the combo

[07:31] market as this is a new feature on poly market we need to research this from like a quant pro algo trader perspective where like math and data are the only parameters to decide if the strategy is a plus expected value we're not going to

[07:48] driven Your task now is to understand the Your task now is to understand the market events etc. and aim to find three interesting hypothesis of strategies we can dive deeper into. So work on this

[08:02] until you have three strategies for us. So this last part here, this is kind of where it knows when to stop. So that is when we have three strategies. If you don't give it like a goal that has like a defined ending, it can just go on

[08:18] forever. So basically I want to try to have like a goal it can kind of reach within like a reasonable amount of time. Uh so it kind of stops after a while. So

[08:30] we just going to run this now and we're going to see what comes back. So this is basically what I did earlier. I had a lot of markets I wanted to check out and I just kept running this until we found some interesting hypothesis. Then we

[08:43] kind of went over to the data collection and that is of course really important and that is of course really important because the data needs to be spoton. But even though I had all the data for this kind of strategy, it still took a lot of

[08:56] adjustments on the parameters until we actually could have the good enough execution to actually find and land these strategies. So I did actually lose some money in the beginning because we did

[09:11] something that is kind of we we filled only one side here. So let's say we only one side here. So let's say we enter on yes here right 42 shares at 4 enter on yes here right 42 shares at 4 cent 4.5 cents but we did not

[09:25] cent 4.5 cents but we did not enter this side and that means that this is at the risk right because we don't have the arbitrage and if this resolves to no then we kind of lost our initial $1.9 here. So there are risk about this.

[09:40] $1.9 here. So there are risk about this. You don't fill every time both sides. So you will need to take on some risk but uh with the execution parameters and if uh with the execution parameters and if you kind of leverage uh codeex cla kim

[09:54] you kind of leverage uh codeex cla kim to help you nail down the execution, you should be pretty risk averse. But of course this is something you have to work on. I spent like a couple of days kind of honing in on this exact

[10:07] strategy. So, our execution was kind of good and we didn't enter so many toxic trades that kind of only filled on one side. So, you can read more about this. Uh you can just ask Shhat GPT or something if

[10:21] you kind of want to know a bit more about this. But basically, this is how I found this strategy. And so far, this has been going pretty well for me. And this just keeps running 247. So 24/7 and it's very passive because

[10:36] the system is automated. So it just keeps looking for these types of setups and it wants to strike when it sees them. So that is how I set this up. But this is just one of many because you want many of these strategies and yeah

[10:52] there are probably a lot of out there that is kind of unexplored. So let let's let's just let this run and see if we can find anything in kind of the combo market. uh we can maybe get some indications of what data we need to

[11:05] collect. So I'm just going to take you back when this is uh the goal is completed. Okay. So you can see after running about 16 minutes that was a pretty quick uh goal actually. But basically now you can

[11:18] see completed the combo quant research run. Full evidence formulas rules risk are in the brief and we have three hypothesis. So one is the one cent longshot floor capture and we have like a description

[11:34] of the strategy here. The second one is like a same event correlation surface trading and the third one is executable payoff bound arbitrage. So now we kind of have three different hypothesis we can start looking more into and start

[11:51] running some data pipelines on if we wanted to. So here kind of my next follow-up would just be something simple uh like uh let me just do something do we need to collect uh for a three

[12:06] strategies to be able to run the test to see if any strategy has some expected value. This needs to be very precise and is what makes or breaks the research. So now we can just ask in this case uh 5.6 six high. What data do we actually need

[12:22] and is it possible to get that data? So this is again the kind of the the the this is again the kind of the the the route I took to kind of get my um to get my yeah this strategy up and running. I asked to what data do we need? Is it

[12:37] possible to get that data or do we need to run like a a few days of collecting our own data? And this is what we did in this case. We ran like a collector over

[12:49] a few days and we got the data we needed to be more precise when we did some calculations or when I say we I mean of course the LLM did some calculations uh to figure out that we can actually run this. So the reason while while this

[13:06] run the reason why I am have been lately kind of pivoting onto this is because this is what the models has been showing to be really good at right they have

[13:18] been really good at everything that has to do with like data code and math so far so that is why I kind of pivoted into this because this is kind of the perfect trifecta we need data it's really good at kind

[13:33] filtering data, collecting data, cleaning up data. We need code to actually run the tests, really good at code and math. Of course, we need math. We need like a strong mathematic understanding of what is at least

[13:48] statistical math. What is like expected value in this case? What code do you need to write to find this? What data do you need to use? So we can kind of combine these three strong sides of a

[14:02] LLM into something and that is what I'm going to do. So I'm kind of learning bit a bit more about uh systematic trading, algorithmic trading, quant trading to capabilities from an large language model. And so far this has been going

[14:18] really well for me. I haven't like put any like big money behind it yet. I'm kind of in my research phase and I've been having a blast. I got to say I had so much fun with this, more fun than I ever had with the AI uh when it comes to

[14:33] ever had with the AI uh when it comes to this because these three uh yeah parameters here are really strong. So we don't get any weak sides of the model actually uh when we are working on this. Of course there are some biases and

[14:47] there are some average stuff. So you got to be kind of precise and that's why I'm trying to learn a bit more the basics so I can ask better questions. Right? So yeah, let's just wait for the response here and see if we are able to collect

[15:03] the data we need. Okay, so we got our response back here. You can see here we kind of look at the data that makes or breaks all the tests. So we need some RFQ request. We need to we need a complete quote life cycle full

[15:18] depth leg log books and yeah you can kind of go down here so basically I just asked a simple question simple question is it possible to get the data we need to evaluate all three strategies yes uh it's not possible from just the public

[15:33] historical data alone so here we will need to actually set up some pipelines to collect our own data right uh yeah you can also read here what data we need to set up and actually gather. [snorts] Uh so this is possible and basically I

[15:50] think this just showed what I wanted to share today. Basically this is the way I've been going and kind of utilizing the agent capabilities of these models. I try to always have some kind of research running and even though some

[16:04] almost all of the time it doesn't lead to anything. Uh I still learn along the way and I get more information and I feel like I'm better suited to kind of go on to new research project but I always try to keep something running and

[16:19] kind of leverage the subscription I have this case on OpenAI. So I always have something running that does some research, collect some data and basically yeah try to find something that is

[16:32] interesting that we can kind of test on the live. So here I would not go live. I will probably set up like a paper trade or like a shadow trade and just collect the data and run this uh for example against our strategy for some days just

[16:46] against our strategy for some days just to see if we get anywhere and if I find something and if we are lucky I guess we can end up with some kind of strategy that we have here that over time hopefully we will find more of these

[17:00] hopefully we will find more of these arbitrage landings landings arbitrage uh trades and we can make some Yeah, more of like a passive or like an automated income on this. It's not going to be huge amounts of money, but if you can

[17:13] get some volume behind this, it can be like a nice side hustle over time, and about data analysis, you learn about how to use these agents, how you can set

[17:25] them up. You learn a bit about finance, and you Yeah, it's just overall positive if this is something you want to get into. And on like Hyperlquid or Poly Market, you don't really have to do any like know your customer or something

[17:39] like that and you can get the data for free. So that is also something that is really nice. So yeah, that is basically what I wanted to share today. I think just a simple research loop I've been running lately with some good results

[17:52] and hopefully this can give you some inspiration of trying to do the same to find your own strategies that could work for you and the way you want to do things. Uh I will be doing more updates. I have a couple of interesting research

[18:05] running now uh in the background. So if they lead to something, I will probably do an update video. Maybe we can go a bit more into the execution because that bit more into the execution because that is very important on this uh very

[18:18] uh there's a lot of competition to get in the queue, but we can talk about that another time. So yeah, hopefully you enjoyed this, gave you some inspiration. want to, subscribe if you want to see more stuff like around AI agentic

[18:33] trading and how we can actually use these models uh with all of their good capabilities to do this trifecta here that I talked about to find some cool things that you might be interested in.

[18:46] So yeah, hopefully I'll see you again soon. Have a great weekend. Maybe on the soon. Have a great weekend. Maybe on the Sunday I have a new video and we speak