Transcription
So, the other day I was scanning the SEC website where basically big Wall Street insiders, CEOs, and big whales have to disclose their trades. And this got me thinking. And by the time we find out through the markets or through the news, it's usually way too late. The guys at the top already know the best opportunities get kept between the elites. And to even get a hedge fund to look at you, you need tens of millions.
So, I decided to build my own AI hedge fund. Five agents modeled after the greatest investors of all time. Real data, real analysis, and running 24/7. But, I didn't just want to build your generic chatbot that tells you to invest in Apple. I wanted to prove it actually works. So, I decided to run an experiment. I wanted to see if my AI hedge fund could beat the top hedge funds on Wall Street. Without cheating, without staging any data, no future prices, no hindsight. Just 3 months worth of data, the AI makes it pick, and we compare it to what the actual hedge funds did.
Now, I understand this is quite the challenge, but in the era of AI, I feel like it's my purpose to empower us, the brokeys. I gave her about 25% of that money. She got 25 grand, I got 75 grand. I'm the genius behind it. I'm the tech. With the insider information these Wall Street guys seem to have. And by the way, this is definitely not financial advice.
And if you follow this channel, you've probably seen me build similar stuff like this before. I built an AI hedge fund on NN with five investors analyzing the same stock, which got over like 2.4 million views. But, it had a massive problem. Every agent saw the same data. So, you basically just got a different flavor of the same answer. And that's not how a real hedge fund works. At a real hedge fund, what the macro guy sees is completely different to what the technical guy sees. They're looking at completely different angles, which is why it makes the disagreement so important.
However, this thing is completely different. With this model, each hedge fund manager only sees the actual data that they are good at reading and would actually have access to inside the hedge fund. And instead of just giving me a suggestion over Telegram, I actually have a visual diagram. Show me the thought process behind every agent's decision. And I can basically watch these five billionaire hedge fund managers all do in real time.
But, first I actually need to build this thing, which turned out way harder than I thought. Now, the default watch list that most of these generic AI trading systems will give you is Apple, Microsoft, Nvidia. But, this is just boring. Let's face it. Any random Joe with ChatGPT can, you know, give you the same investment advice. But, real AI hedge funds give you opportunities that no one's even looking at. So, we're going to do things a little bit different. The composite score combines all of them. The quality gate is at 0.35. Below that, it doesn't even get passed through to the agents. So, only the most interesting picks actually make it through. Think of it like a bouncer at a club entrance. If your name's not Joe, you ain't getting in.
To ChatGPT, I've given you my old AI hedge fund with five agents, same data, text output. I want you to rebuild this information with asymmetry between agents, meaning they won't all see the same data. I want it to have real-time signal detection, a visual dashboard, and critically I need it to run a blind backtest against a specific date. Can you help me with the architecture?
Oh, I'm all in for this upgrade. So, imagine your five agents like a team with different specialties. Firstly, each agent get a different feed. One's on economic news, another's on market depth, another on social sentiment, and so on. They all shout out their signals into a central hub. That hub, your strategy engine, will have a visual dashboard, so you can see who's calling what. Now, the blind backtest, you pick a date, the system pretends it doesn't know the future, and it runs the signals as if it's that day. Architecturally, you want a modular pipeline. Separate data feeds, a signal aggregation layer, a dashboard UI, maybe Next.js for fun, and a backtest runner that can replay historical data. Let's make it slick, yeah?
And what would the tech stack be? Would it be Docker, Timescale, Celery, React? What would you suggest?
All right, let's piece the puzzle together. Back end side, I'd definitely suggest Python. Flask or FastAPI is slick for your API layer. For tasking and real-time signal handling, Celery is a classic for distributed tasks. Timescale or another time series DB will be your bread and butter for storing and querying all those signals. Then slap on a React or Next.js if you like dashboard for that clean front end. And yeah, Dockerize it all so you can run this beauty anywhere. It'll be smooth.
Okay, that seems like a big build. Do you think I can get this done in 5 days?
5 days is pretty ambitious, but hey, you thrive on big swings. If you're reusing some old components and keep the scope tight, like an MVP version, you can push it. Prioritize. Core agent logic first, signals, basic dashboard. Perfection? No. But, you'll have something light to iterate on. You know the drill. Sprint hard, vibe code the rest.
So, can you give me a prompt to give to Claude code to actually build this thing?
Perfect. Thanks for the comprehensive instructions. Just one thing before we begin. Don't ever tell me what to do again, you little.
Day two, seven database tables, four ingestion jobs, price data every 5 minutes, fundamentals daily, insider trades daily, and news every 15 minutes. 420 price bars, 687 insider trades, 200 news articles, all completely automatic.
Day three, the agents. This is where it stopped being a project and actually turned into a real hedge fund. Five investors, Buffett, Munger, Cohen, Dalio, and Ackman. But, here's what makes this thing different from every other AI hedge fund that you probably seen out there. Information asymmetry. In a real fund, the macro strategist doesn't see the same data as a technical trader. Different information, and that's the whole point. So, I built a data practitioner. Buffett only sees fundamentals, PE ratio, revenue, cash flow, debt. No charts, no news, just the business. Munger gets fundamentals plus news. Ackman gets fundamentals plus insider trades. So, basically when a CEO is offloading loads of stock or or buying huge volume, Ackman sees it all. Cohen only sees live price action, candles, volume, support, resistance. But, basically he has no idea what the company even does. And Dalio sees price action plus the news, so he can really get that macro lens of what's going on. But, the key here is that they cannot see each other's data. When Buffett says buy on cash flow, Cohen says sell on the chart. The disagreement is real. Neither is wrong, they're just looking at a different side of the picture.
Now, that is genuinely looking really good.
Day four, the dashboard. 10 pipeline nodes when the system's running, animated dots flowing along the connections showing the data. Three panels, pipeline, activity feed, opportunity ranking, and every agent's verdict, confidence, reasoning, and a committee consensus. But, the question I really needed to answer is, would this thing have been right 3 months ago? Because anyone can go and look at a stock and say, "Oh, I would have bought that 3 months ago." I need to prove that this thing can actually predict, which brings us to the part of the video that actually matters.
So, a few days later, it was time for integration. Basically, we had to wire this whole thing together. And well, it didn't go too smoothly. The regime detection completely broken, keywords didn't match the signal type, a whole waiting system was doing completely nothing. But, we managed to fix it. However, the CIO was way too conservative. Everything was basically getting rejected, and the threshold was set for blue chips, which we really didn't want cuz this is just boring. We wanted this thing to show emerging sectors, so that was the next fix. But, then come another issue. We were using the OpenAI API for this just to keep costs down, and we hit the rate limit on it. 100 API calls, a 30,000 token limit, the system just kept hitting a brick wall. And the retry or fallback logic fixed this, but it was just painfully slow. We had to do something about it.
So, then I tried the rerun. All fixes are done. Scanner still says zero queued, and the system was saying they'd already analyzed everything, which was just complete. It was still running the old keys from the old run. So, I had to manually flush the keys. Six bugs, one day, but we had to get it done for the video. And to be honest, this is what building actual production software looks like. Not clean montage, but 5 hours of why isn't this working? Followed by that one epiphany or line that basically fixes everything. But, by the grace of God, we managed to get everything fixed, which finally means that we can carry out the experiment that I've been trying to do the whole video.
Right, so we finally got the system running. The agents were analyzing properly, the pipeline was running smoothly. But, none of this actually mattered if we couldn't get it to pick stocks. Now, I could let this thing run for 3 months, come back to it, and figure out how it's done. But, this YouTube game's hard. I need you to please like and subscribe so I can keep making videos like this. But, I don't have 3 months, and you know, really I could just cherry-pick the results, and you wouldn't know any different. So, I had to make this a scientific proven experiment, but without wasting 3 months of my life. So, I designed something called a blind experiment.
Why would we do a double-blind study? Well, the main purpose of a double-blind study is to minimize the effects of the experimenter bias. In other words, the results of the research are less likely to be affected by external factors, such as the experimenter's verbally or non-verbally communicating their assumptions about the treatment's efficiency, or maybe the expectations of the participants. So, in psychology, a double-blind experiment is where neither the subject or the researcher know who's actually receiving the real treatment. This removes all biases, so you literally cannot cheat, even subconsciously.
So, step one, I pick pick from exactly 3 months ago, and I load the database with data only up to that point. No future prices, no future news, no insider trades. The AI cannot see anything past that 3 month cutoff. Step two, I let the system scan its full 20 stock watch list only using the data that it has, the historical data that I loaded into the database. Full pipeline, signal detection, quality gate, five agents, committee, and CIO. So, it can run exactly like it would in real time. So, I analyze in January's data as if it was today. Step three, the system makes its pick and then I record everything before I actually look at what happened. And then step four, only now do I actually look at what happened with the stocks over that 3 month period. And then step five, I compare the AI agents' returns against the S&P 500 and against the reported returns of the actual hedge funds during that 3 month period.
So, the time has come to actually run this experiment. So, I'm about to build a database. I'm about to run the blind test. So, let's see what happens. So, I've now give it the instructions to run this test. Um we're going to see what it's going to do. So, we've got it all here. We've got the database, the blind experiment results, you know, the back end blind run it so it's going to run it separately so we can't cheat on any of the data. Um and we'll see this thing's running. So, let's just see how this thing goes.
So, we now have the results and I promise you I've not looked at these yet. I literally just loaded this and we have to scroll down to view them. So, we can see the cutoff date 1st of January 2026 and then the evaluation date too. So, this is when we're actually getting the data too. So, this is actually scary because, you know, I've been building this thing for, you know, 5 days now. I'm going to finally going to see whether it was good or not. So, let's have a little look at the picks. Okay. That's not actually bad at all. So, the AI portfolio returned 2.89% and the S&P returned 3.9% so, Oh, I've just read this wrong. So, we're minus 3% during this period. But the S&P 500 was actually minus 3.9. So, although we lost money, the AI beat the S&P by 1.1%. That's absolutely insane because I built this thing in my bedroom. And then we can see the list it's actually allocated to so you can actually verify this. And then phase four, the AI portfolio returned minus 2.89% versus minus 3.9% beating the benchmark by 1.01%. Critically, the AI's conservatism, 77.5% crash protected it during what turned out to be a rough Q1 for 2026.
So, now I want to want you to show me the stocks side by side versus what the AI picked versus, you know, the S&P 500. And it's even created a blind experiment dashboard for us so we can see our actual returns versus the S&P 500. So, let's have a look at this one. We have plus 46% on VRTA AI Infra. This looks like a space stock and the actual return was 27.62% and we allocated 1.5%. Um and then we had a nuclear stock which we made 11% on. Um and then we lost quite a few of these. So, we lost 0.83% on that. BST, we lost nearly 8%. On biotech, we lost nearly 8%. And then 27.55% on on this one. So, although there was a few losses, we do have some really good winners here as well. And then we can see the full watch list um here. So, I'll publish these results as well so you can have a look at them yourself. But that's cool. I think that's really cool. And we can even look at the actual return here and the contribution. And we can even see the consensus on it. So, we can see if there was a split consensus whether everyone agreed or didn't agree. So, I think that's really cool, to be honest. Um this has been a really good experiment and we can even see what the agents agreed on which is absolutely amazing. So, I'm happy with this result. Um well, in terms of the actual AI hedge fund index, we actually lost to them by 2.46 points which is not great, but I definitely think in future experiments, we can beat these guys. And we'll never actually fully know cuz individual hedge funds like Citadel or Bridgewater don't actually declare these numbers. But if we take the index, then yeah, they won this time anyway.
So, can an AI hedge fund beat Wall Street? Well, it can definitely beat the S&P 500, but this time to actually beat the hedge funds, not quite. But here's what I think actually matters. This full experiment was verifiable and provable. The picture autonomous, the results are real, no matter what they were. And the fact that I can sit here and build something in my bedroom which actually beat the market in many aspects is absolutely insane. The field of finance is absolutely getting leveled by AI and now is the perfect time to jump onto it. And if you're watching this video thinking, you know, I want to build stuff like this for your own personal finance or for foreign or for trading, whatever you want to use it for, then I can help you build these projects inside my school community below where I can jump on a one-to-one call with you and guide you through the whole process and give you all the templates if you don't want to start from scratch. And you can also get access to this project inside the community. So guys, I hope you enjoyed the video and I'll see you guys in the next one. Peace.