¶1So, if you go to chat GPT now and you say, "I want to buy a stock on Nasdaq this week. What stock is good to buy?" You will get an answer, right? And same if you say, "I want to place a position on the NFL game between the 49ers and Seahawks this Sunday. Who do you think will win?" You get an answer. And I'm not saying this is worse than doing like just a 50/50 or like a blind pick on the Nasdaq, but I have been kind of running my sick intelligence startup now for like 6 months.
¶2I've been kind of diving into this topic, and I thought I have some interesting things to share because there's not much you need to do now these days if you have access to AI to really help you like make much better decisions and not just give away your money to Calshi without doing some kind of research and trying to get the most out of the tools you can actually access now using AI. So, I went ahead and I asked Claude to actually create like an animated video what we're going to go through today, and I just wanted to play that straight away because it kind of explains everything we're going to go through today. So, now you kind of know up front what I'm just going to go through today, and hopefully this can kind of give you some understanding of, yeah, how you can kind of get maybe that small edge or at least know what to avoid. So, yeah. Let's just play the video and then we kind of get into the more practical side after.
¶3>> Ask a chatbot who wins Sunday and you get a confident guess. Not a price, not a plan, just vibes. So, we built a real model instead. Every NFL play since 2013. And the most valuable thing it told us, sometimes the best move is to pass.
¶4Here's the game plan. We'll take Sunday's 49ers versus Seahawks game and build our own forecast from real data. Why? Because a position only makes sense when the price is wrong. >> [music] >> How?
¶5Five steps. Read the price, strip the vig, grade every play, test it against the market, and then decide. Take a position or pass. And here's where AI does the heavy lifting. It calculates the fair price.
¶6It cleans the data. It writes the machine learning model, runs it, and then explains the results in plain English, so we actually understand what it's telling us. First, every price is a forecast. Seattle at 59 cents on Kalshi means the market thinks Seattle wins 59% of the time. Market providers build in a cut.
¶7Add both sides of the market, and you get 103%. That extra three is the vig, the stadium markup on your hot dog. Strip it out, and you get the fair price. Seattle 59.2. Then the data.
¶8453,000 plays, collected and cleaned. One file even listed the wrong quarterback. The play-by-play and the injury report told the truth. Next, we grade every play with EPA, expected points added. Think of it like a GPS for the drive.
¶9First and 10 at your own 25 is worth about one point. Complete a pass to midfield, now it's 2.6. That play earned plus 1.5. Add it up, adjust for who each team played, start from 80% of last season, and every team gets a power rating. Pass and rush.
¶10Offense and defense. Seattle also gets about two points for home field. Ratings become a point spread, Seattle by two or three. And the spread becomes a win probability. Our model, Seattle 56 to 60%.
¶11Then the real test. We replayed every game from 2014 to 2025 using only what was known before kickoff. Coin flip, >> [music] >> .247. Our model, .224. The [music] market, .212.
¶12Lower better, and the market wins. That's not a failure. It's the point. The market is millions of dollars of sharp opinion. Our model is your referee.
¶13It tells you what fair looks like, so you never pay a markup on a coin flip. So, we only take a position when a price is clearly wrong. Our rule, at least 2% edge after fees. Like a kicker, you only try it when you're in field goal range. Sunday's game, fair price Seattle 59.
¶14Kalshi asks 59 cents, 60.7 with fees. Edge, minus 1.5. San Francisco, minus 2.9. Neither side is in range. And when you do find an edge, Kelly sizing tells you how big to go.
¶15We use a quarter of it. With a thousand dollars of capital, that's often about $14. Never go for it on fourth and 20 with your whole season. And this goes way beyond football. Whether it's a Sunday game or a Nasdaq stock that's priced for perfection, the same math helps you avoid bad positions, and that levels the playing field.
¶16But remember, this isn't the truth. It's statistics and probability. Even a fair price can go the wrong way. That's the playbook. Find the fair price.
¶17Only take a position when it's wrong. And when it isn't, the best move is to pass. Play smart and trade responsibly. >> Okay, so in the video it's said that we will use our model to find the fair price, right? But it's not that simple, right?
¶18Because beating like a prediction market like Kalshi won't actually know what actually the fair price really is is really hard, right? But this video isn't about that. This video is more about learning to actually use AI to make better decisions and learn something along the way, right? Because sometimes, maybe you start trusting your own model more than you trust the market, and this is where you can find some edges. But, this needs a lot of examples.
¶19You can't just do that on one example. But, I think this is a good start today, right? So, before we go through my AI forecasting system, I want to tell you a bit about an event that comes up in a couple of weeks, and that's the NVIDIA GTC Berlin 2026. Okay, so in about 2 weeks now, NVIDIA GTC Berlin 2026 kicks off. I'm going there in person, so I just wanted to share a bit about it.
¶20Maybe you kind of want to do that, too, and maybe you want to just uh attend virtually. So, I wanted to share a bit what is [music] actually happening at GTC Berlin. So, we have all the sessions. Of course, we have Jensen's keynote. That is pretty highlighted.
¶21But, I wanted to share a bit more what else is possible at the GTC Berlin 2026. >> [music] >> So, we have a lot of sessions under open models and data sets, and I think there are some really interesting things here. So, the one I'm definitely attending here is the Where is it? I think it's down here. Agentic AI in production from open models to agents you own.
¶22So, this is kind of blown up lately. You can see how kind of the uh the use cases have switched from [music] using only kind of proprietary models over to more open-source models, at least in businesses. And I think this is just going to continue, and NVIDIA is covering a lot of open models and data sets at the conference this year. And of course, they are building their own Neumotron models. Also, agentic AI and closed conference sessions is pretty popular.
¶23Agentic AI in productions from open models to agents you own, we just mentioned that. And this one looked pretty interesting. So, this was self-improving AI systems. That's a pretty hot topic these days with the RSI and stuff like that. So, I'm definitely going to catch that.
¶24So, yeah, that's basically what I wanted to let you know. NVIDIA GTC Berlin starts on October 20th to 2022 in Berlin, of course. So, if you're going, let me know. Would be nice to meet up. I'm going in person.
¶25So, but if you can't go in person, yeah, you can attend virtually. So, follow the link in the description and register now and attend the sessions that you think is interesting. So, yeah. Thanks to NVIDIA for kind of highlighting this part of the video. Now, let's go back to the project.
¶26Okay, so let's do this now. So, this is kind of my AI forecasting system. This is like a simpler version that I have been using kind of in my my Sick Intelligence startup. But basically, the idea is don't ask, just like we did in ChatGPT, who wins? We kind of want to ask, is the price wrong?
¶27Uh if it is, then we're just going to pass up on this position, right? But we want to set some rules first. I always do that. So, we want to think about how many how much stake we want to kind of place in like on the Nasdaq, on a stock, on a position on Cal sheet, because we kind of want to know that up front. And if we think our model finds some kind of edge, we want to kind of think about in before we actually calculate that, when we are going to take that position or not.
¶28So, let's say it's just 2% if that happens or more, we're going to do it. And kind of how small your position stake kind of based on your Yeah, you can call it bankroll, right? So, we're going to do like a Kelly sizing, something like that. So, I just start with that and I say I'm ready to place a position on an event. I have a $1,000 edge I need to place two and I want to use quarter Kelly sizing.
¶29This is like a math based way to kind of place a correct size, as you probably saw. Okay, so now this is kind of our rules for today, right? And now we're just going to move on to try to find a fair price, okay? So, we want to try to find a fair price for this event here on Calchas. So, let's say this is the stats now.
¶30So, let's just screenshot this. I think that's super simple. All right, we just screenshot this. Okay? And we placed it in here.
¶31Second. And I just do something like, let's find the fair price with no vig on this event. So, basically we got this screenshot and I got like a second opinion from this site here because it's nice to have something to compare to, right? So, now we have two different events and I'm just going to let AI handle this. Like we know, AI is really good at code and it can calculate this.
¶32Providers build in a hidden fee called a vig. Remove it and you will get a fair price. So, this is kind of step one here. So, you can see here like with the with the vig removed, we kind of the margin was like almost 3% so like 103% total. But we remove that and we compare this to Calchas.
¶33So, now we kind of have our fair price. And yeah, that is good. And now we can kind of see what we need to be the 2% expected value. So, we need like on Seahawks in this case, we need 55 cents or lower to clear out 2% after fee. So, how do we how can we kind of get to that to that number, right?
¶34So, now we need to collect all the data, right? So, let's try to do this as simple as possible now. So, I kind of noted down some free tools here that is kind of sources of data. So, let's just copy these tools first. Okay?
¶35Um all right, let's just copy this. And we can paste this in here, right? And then I'm just going to grab a very simple instruction here. Uh with all this as as pro quant forecaster, create a model to give us a prediction on the game, work hard and precise. Let's just try that and see where that takes us.
¶36So, we are on Opus 5.5, right? And hopefully now these models are not so good at like machine learning, writing code and stuff, and also fetching data, cleaning data. So, this should really be enough to make these simple instructions. But here, of course, you can be more specific. I just want to see what we get if we kind of run it like this.
¶37So, let's just let Opus work this a bit and I'll come back when we have something that is uh yeah, kind of a bit more um ready to go here. >> [music and singing] [music] >> Okay, so then after about 17 minutes, we got our model. We did collect every data. We got our model run and we came back with a result here, right? So, you can kind of see how good is the model while I clean Seattle.
¶38And we can see all the results we got here. So, the raw model got 63.6 for a Seattle probably probable win. Uh that kind of differs a bit from the market, right? And we can do like a model blended with the market to get a more close to the market. But this is things you can dive into yourself.
¶39So, basically, this is just to give you like an introduction into start using more AI active in decisions like this because I think like putting in math and stuff into these things when you have these models now is very quick and it can make you like do better decisions. So, I asked it to create like a simple HTML overview here of the the outcome. So, let me refresh this. And you can see position call, no position at the market price. And it says, "Don't touch touch SF." Is that 49ers, I think?
¶40And here we got the result. So, 63.6 is kind of our model our model, but the market is at 59.7. And we get does it clear 2% taker at 57? No. Resting limit at 57.
¶41So, we could do that. We could place an order at 57 and hope that it fills. Uh or we can do like a taker at 56. Uh and we get the stake. So, 14 and 10.
¶42So, if we get in at 57, we will have like a four per- 4.2% edge uh uh with our model, right? And you can see we got a bunch of other recommendations here, too. So, if you go back to kind of my my steps here. So, we did the own forecast, right? We ran the model.
¶43And then we can ask a question, why did the market uh why might the market be be right and me wrong? So, we can just ask that here where we did the model, right? And you can see here now why the market is probably right and the model is wrong. The back test already showed it, right? The market knows things the model doesn't.
¶44That's true. Rumors and injuries and stuff like that. Seattle's first rating was mostly last season. And you get a bunch of And then you can kind of read all this. You can do more research.
¶45Uh this could be anything. This could be on stocks or everything. But the basically today I just wanted to show you like a simple workflow that doesn't take too long. And when you build some kind of models, you can reuse them. And like I said, help you make better decisions.
¶46And if you're going to read when you might be right instead, you need something the market hasn't priced yet. Injury news before the line moves or a price that is stale after news breaks. This could be stocks, right? And everything. But of course, stocks is really competitive.
¶47So, that's not very easy for like a single person to do. But you might have some knowledge the market doesn't have. You never know, right? And systematizing it, automating this is something you can do, right? So, one thing I didn't add here that was step nine, and that was building this into an automation.
¶48So, we can always keep scanning for the edge that we talked about, right? And when we find that edge, we also have a system that places that position autonomous. So, this is the way I have kind of my system set up, so it always looks for that edge compared to our model, and it places that position autonomous based on sizing and stuff like that. So, yeah, that's just what I wanted to talk a bit about today, and I think it's really interesting now to start trying to look at these different models as we did today. So, this is kind of what I learned from my um my startup, Sick Intelligence.
¶49Uh I've been trying to develop these types of setups for like 6 months now. And yeah, I I have I have improved a lot. Like, I learned a tons, and now I'm starting actually to do this on a more bigger scale using more compute and using more advanced data pipelines and stuff like that. So, go check out my website, Sick Intelligence. You can sign up for early access now if you want to.
¶50Uh I am running like uh midterms uh bench here. So, you can see now, if we look at uh the Senate, or let's look at the House here. You can see we are pretty close to the market. We are like Our model is at 88% Democratic, and the market is like at 91. For the Senate, it's uh a bit lower, actually.
¶51Senate is more 56 on the Democrats, but it's climbing on round two here. So, I do like an update every third day. So, definitely go check out my website. I have some fun predictions here you can check out. Uh just some fun virtual one.
¶52And I have my introduction video. So, definitely go check that out if you found this video interesting. And yeah, hope you enjoyed it. Hope this makes you use AI in more decision-making like this, because it's really worth it. And this could be adapted to most uh most most things that has something to do with numbers, right?
¶53So, thank you for tuning in. Have a great day, and don't forget to check out NVIDIA's GTC 2026, and I'll see you again soon.