My First Winning Agentic AI Trading Strategy On Polymarket

[00:00:00]Hello, hope you are doing well. So, I just got back from the US from the World Cup. Uh Norway won 4-1, so I'm super happy. But today, I wanted to talk a bit about a strategy I was thinking a lot about and I've been testing while I was traveling. So, I thought I can just kind of go through this strategy today and this is a strategy that is very good if you want to implement kind of Yeah, you can call it agentic AI or something because it's really nice to have any kind of smart model to help you figure out the details of this strategy. But the fundamentals is quite simple, I would say. But it's a bit of a different strategy when you kind of think about Polymarket that we are looking at today because on Polymarket, you usually thinking about being on the taker's side. You want to place a bet, right?

[00:00:47]But we are going to do something a bit different today that I kind of want to show you that I have some great success with the over the last few weeks maybe or months, I guess.

[00:00:58]So, I I just want to walk through kind of how I set this up and what you kind of need to get going if you want to start looking at this strategy. And of course, this is nothing to do with financial advice or anything like that.

[00:01:12]This is just a strategy I've been testing and having some luck with. So, basically how this strategy work is that we are looking for some expected value, right? Because on each trade here, we want to avoid a few things. Uh on the maker side, we we can call it maker side. We want to use that to avoid the fees that we pay on a Polymarket, right? Because we need if we just do it the taker side, we always pay some fees to Polymarket and that gets kind of expensive over time.

[00:01:44]Another thing we want to avoid is slippage. Uh if you don't know what that is, it's let's say you place a trade here on like I don't know. Let's say you place a trade here on like 40.

[00:01:56]Something like that. Uh and you want to do a trade on like the price 40.0, right?

[00:02:01]But there's a lot of people in queue and you don't have the perfect latency. And you get filled at something like.442 cents instead of the 40 you asked for because maybe you use some fill any price you get. Uh that means that you're kind of 2 cents down already and you get the fee on top of that. And that means that you kind of lost if you have like a small edge you were looking for. So this is what I want to avoid with this setup we're going to look at today. I want to avoid paying fees if possible and I want to avoid the slippage. So how can we actually do that on Polymarket? So like I said, we can't really be on the taker side of this. We need to kind of be on the maker side. Uh I can just pull that up if you want to. So you can see Polymarket has something called market making. So a market maker on Polymarket is a trader who provides liquidity to pre- precision markets by continuing posting bids and ask orders.

[00:03:01]So we're not exactly going to do like pure market making on this, but we want to take advantage of that mechanic.

[00:03:08]So I'm going to try to explain how I did this with like a simple setup here and how you can try to do that too. But there is a big caveat I'm going to talk about uh uh that you need to figure out and that is something that AI is really good at helping you with.

[00:03:24]So the idea is basically we need to find something that is called that is called a fair value price. And this is what I've been using AI to help me calculate by collecting tons of data.

[00:03:39]I'm going to talk a bit more how much data I did collect to kind of hone in on finding the perfect fair value price for my model. But let's say we have the fair value price.

[00:03:51]And what we can do then is try to leverage that by just Let's say the price on Polymarket now is 55 cents for upright. So that means that the market think it's a 55% chance that in this window we will kind of end up on uh up. So this is we're talking about kind of the BTC market in this example 5-minute up and down. So we just going to see if the Bitcoin price kind of ends on up in the 5-minute window before or down regarding to the previous price in the last window, right?

[00:04:31]So let's say the on the on the Polymarket the value is 55% chance of ending at up at an X kind of amount of time left of the window, right?

[00:04:44]But our model that we created with AI says that the fair value should rather be something like Let's say 0.51. So we Our fair value is 51% chance of ending upright. And that means that we don't We won't adjust this according to our price. So this is what we think is the true price, not the market price. And we want to try to take advantage of that by placing some resting orders. My AI model come after looking at all my data uh looked at it and it said that we need to have a 4-cent spread to be positive after looking at all my data. And let's say our AI model kind of finds out now that the fair value price for this setup on the BTC 5-minute up and down market is 0.51.

[00:05:41]But like I said, on the Polymarket the price is 55. So we are not interested in looking at kind of the market price. We want to check always compare it to our fair value price. And like I said, we have kind of hard-coded that we only want to look at things that has a 4-cent discount. And that is calculated from our fair value price. So, let's say we want to buy an up share in this market now. What does that mean? Of course, we want to look downwards at the price. So, we are only interested in buying upside.

[00:06:18]That is only going to be 47, right? So, we going to place our resting order bid at 47. So, the idea here does we we always want to to swap this or switch this. So, let's say our fair value price moves down to 45, then we want to set our resting order to 41, right? So, we always want to adjust this. And this is what our model is doing. So, we want to buy up at 47. This is of course like a big difference to the market value.

[00:06:51]But, what happens sometimes is that we get this impatient traders. So, they really want to sell.

[00:06:58]And if they can't really get the price, they just want to dump it. And we are kind of on the on the book with a 47 here ready to capture on that edge here, right? So, if someone really wants to get out, they could sell to us for 47.

[00:07:14]And we get the discount. And that is on the upside. If you kind of think if we want to buy on the downside, we just do like a simple calculation like the fair value is 51. We do like 1 minus I guess 51. That is 49 minus 4. So, that is going to be 45, right? So, that means that if we want to buy the down share, we are only looking to buy it for 45 or less, right?

[00:07:44]So, this is basically the strategy I've been testing out. So, I only want to buy at a discount because that means that we have the the plus expected value kind of baked into these prices, right?

[00:07:59]Because over time, if we always have the 4-cent discount, it doesn't really matter if the market resolves up or down if we always buy uh our um shares at a discount. But, this is uh where it kind of gets interesting because this means that this thing, the fair value price, right?

[00:08:23]has to be almost perfect if we going to if this is going to work. And that is what I've been using AI to help me figure out. So, that is the big hurdle here. You got to collect like a tons of data to kind of get this fair value price calculation correct. So, it doesn't really matter if I talk about the strategy because if you don't find a good way to calculate your fair value price, uh yeah, this is not going to work. So, that is what you need kind of to collect all the data from. So, I kind of want to show you how much data I collected before I kind of set my model to trust this price.

[00:09:01]So, here you can see what I collected. I collected for the 5-minute up and down fair value model we analyzed, 144,000 graded fair value snapshots, 2,000 resolved markets, 170 hours of live market data. So, basically 7 days or something.

[00:09:21]And you can see uh yeah, this is kind of grinding slowly over time. We have done 32 wins, but this was also early data. So, the late part of the model has been a bit stronger, but a bit um a bit more less frequent, but we are up like almost $70. And this is the strategy running like fully autonomous.

[00:09:44]So, what we did on when I kind of got back from um from the US now, we have adjusted the gap. So, now we are a bit less frequent, and hopefully that means that our uh yeah, kind of sharp ratio will be a bit uh more steady instead of having all these spikes. Uh I also looked at the 15-minute market data. Uh I'm going to come back to that in a future video uh how that is looking.

[00:10:10]>> [snorts] >> But basically, what you need to do is collect a lot of data if you want this to work. So, that is kind of the big hurdle. But what is good is that you you can use like Codex, Cloud Code, Open Code GLM 4.5.2 or something to help you collect this data via the API. So, it just takes time.

[00:10:30]So, everyone can do this. But of course, like I said, this is nothing to do with financial advice. This is just what I've been testing out.

[00:10:37]So, hopefully this was understandable.

[00:10:40]Uh I don't know really know how to I guess I could mention that we are not kind of trying to participate in the the market maker rewards or rebates program at this with this strategy. We are not looking to get any of that pool if you know what that is.

[00:10:57]We are only kind of looking to avoid the fees and the slippage part when we are doing the on the taker side.

[00:11:07]So, after playing around with this for months, I figured out that this is the way to go for me at least. But of course, there are there are there are options on the taker side, too, I think. But uh this is what I'm going to focus on going forward, I think. So, a couple of other things, too. Uh I asked uh Codex here 5.5, "Should we do some Monte Carlo simulations on this strategy?" And uh ID it kind of gave me here is that we can do that later, not now as the decision-maker. We just want more data first before we kind of start to look into that. And the Monte Carlo simulation could ask is $12 US dollar per quote too much? What drawdown should we expect? How likely are we to be down after 100 fills? So, this is that we kind of need more data before we're going to do that, but I'm definitely going to do it in the future. Another thing I asked the model is what about overfitting in this case? Uh this is of course uh a danger here. We saw that when we tried to adjust some uh parameters in the early models, uh early setup of this. But uh now we kind of make a 3-cent gap change after looking at the 7 days of data.

[00:12:15]Uh I'm just going to keep monitoring this and see what happens, but uh this is of course a real danger with this model. But uh the nice thing is that there's not really like a big It [snorts] is of course a drawdown here, but it's not as long as our fair value is good, and we can check that when we have enough data. Uh we're not really taking any like big risks, so you will not get kind of bankrupt using this strategy. Uh like I talked about in my previous video, the idea here as does this is going to be like a Yeah, what do you call it? This is kind of going to be like a long-running AI agentic uh strategy. Like I talked about my pod setup, so this is just one of the pods that is just going to run in the background and hopefully over months it has like a plus uh expected value or like a in the green. So, it doesn't really matter if it only makes like $25 a week because it doesn't really cost anything to run. Uh as long as we have the data and it looks pretty good, I'm just going to keep it running. So, this is this is not a part of like a You're not going to get like a be a millionaire from this strategy or anything, but it could be like a part in a bigger plan that you have. So, this is why I wanted to talk about this today.

[00:13:35]And like I said, I didn't really go into any like specific data or setups, or I didn't walk through how to set this up, but I thought it was just an interesting idea that we kind of came to set up by using both Fable. I used I used Codex 5.5. And uh probably you can use some lesser models for this because the math is not really you don't need like Fable to calculate this strategy, right?

[00:14:02]Uh but of course, like I said in the video, the fair value price.

[00:14:06]And here you need a lot of data to kind of figure out because the whole strategy kind of relies on this being as correct as possible. Of course, it will never be 100%, but uh you can kind of backtest and see how close you were, right?

[00:14:22]So, that is what is pretty interesting about this setup. And like I said, I'm going to keep running this. So, if you want to come back in the future looking at these types of videos, these types of strategies, and see how we're doing. I'm going to do update videos in the future on this.

[00:14:40]So, please like this video and subscribe. I also did sign up to Interactive Brokers because I'm going to start testing this Agnostic AI on more like option trading and stuff like that.

[00:14:53]So, if you want to follow along on that, too, please tune in.

[00:14:56]So, yeah. Thank you for tuning in, and hopefully I'll see you again soon.

[00:15:00]And yeah. Bye.