YT YouTube2Text

YouTube2Text → Transcripts → Ho Trasformato Claude in un Trader Automatico 24/7

Ho Trasformato Claude in un Trader Automatico 24/7 — Transcript

by Giuseppe Castagna · 2,957 words · 434 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00In half an hour, I turned Claude into a
0:02professional trader that works 24/7,
0:04makes forecasts, reads the markets, and
0:06invests for me. And it became the
0:08perfect employee; it never gets tired,
0:11doesn't ask for time off, works 24/7,
0:13and above all, in the world of
0:15investing, it never panics. It is
0:17always logical and perfect. In this
0:20video, I will tell you about its first
0:22week on the job, what it does at 6:00
0:24AM with the markets closed, but above
0:26all, I will show you how I built it
0:27without knowing how to write a single
0:29line of code and I will show you how
0:31you can do it too, very simply, to
0:32change the way you see investing and
0:34artificial intelligence. Let's begin.
0:37Before I show you Castagna Capital,
0:39which is my automated trader, and
0:41especially before showing you how I
0:43made it, we need to make a deal:
0:45because when it comes to investing,
0:47money, and AI, YouTube is full of
0:49nonsense, and 90%of the people showing
0:51you videos like these want to sell you
0:53a dream, namely that with AI, trading
0:56will make you rich overnight without
0:58knowing anything and without doing
1:00anything. On this channel, I don't
1:02think like that; I show you how things
1:04really went, and my first automated
1:06trading experiment with AI was done in
1:08a video that I'll leave in the
1:09description, which you should also see
1:11up here, and I decided to ask Claude to
1:13turn € 100 into 1,000. It suggested
1:15that I start trading crypto because it
1:17was the only market open 24/7, a bit
1:19fast, but anyway, in that video, you'll
1:21find exactly how it went. After that
1:23experiment, which went very well
1:25overall and was watched by hundreds of
1:27thousands of people, I decided to learn
1:29from the mistakes I made, and the main
1:31one was that it gave me the signals and
1:33then I had to execute them, and more
1:35than a mistake, it was a hassle. This
1:37time, I've completely flipped the
1:39script, so it does everything, and I
1:42watch and see what happens. And since
1:44I'm not totally insane, nobody gives me
1:46money for free, just as I imagine they
1:47don't give it to you. This experiment,
1:49obviously, is divided into two parts.
1:51In the first, we will create the system
1:53and run it to see how much money it
1:55makes with a demo account, so we'll do
1:57the setup together and it gets refined
1:58based on the investment needs or
2:00desires of each of us, and in the
2:02second, we'll put in real money. Second
2:05point, this is not financial advice, it
2:07is not investment advice; it is an
2:09experiment to understand how this
2:10system works. Third point, I will show
2:13you everything, even the things that
2:15didn't go well. So, if we're in
2:17agreement, let's start with our setup,
2:20and first I'll show you our Castagna
2:22Capital. With an employee of yours, a
2:25human trader, the first thing you'd do
2:26is ask what they buy, how much they buy
2:28, what their knowledge is, and what
2:30they check to decide whether to buy or
2:32not. We’re doing more or less the
2:35same thing here, so to hire our
2:36employee, we’ll draw up an actual
2:37contract that will then obviously be
2:38inserted into the Claude code. My
2:42decision for this test is to have it
2:43work on only eight assets, so as not to
2:45drive it crazy with the billions of
2:47investment options that exist in the
2:48world. And the assets will be the US S&
2:52P 500, the tech index, small-cap
2:54companies, then Apple, Microsoft,
2:55Nvidia, gold, and government bonds.
2:58That's already enough material to do
3:00well. The goal isn't to day trade, but
3:03to beat the index, so it must buy and
3:05hold. And if beating the benchmark
3:09index seems like nothing to you,
3:11consider that there are managers paid
3:14millions of dollars who can't do it. So
3:17if it manages to do that, we’re
3:19already—the cool thing is that Claude
3:22is extremely strong at reading the big
3:24picture and reasoning. In fact, for
3:26several months now, it’s had a
3:28function called agentic financial
3:30analysis that does exactly that: it
3:31reads the financial situation and makes
3:33a predictive argument about what might
3:35happen. ChatGPT does more or less the
3:38same thing now, although they’ve
3:40blocked the ability to execute
3:42automatic trades, but whatever system
3:44you use here, it’s a matter of code
3:46and APIs, so the intelligence can be
3:48swapped out based on the best model at
3:51the moment. In this case, we’re using
3:53Claude, and we’re doing great. What
3:55Claude isn't good at—I learned this
3:57with the crypto experiment—is that,
3:58in general, it isn't great at day
3:59trading and guessing the right moment.
4:02So, trading with thousandths of a
4:04second when there are people with crazy
4:06servers doing things in milliseconds is
4:08not its job. So now I’ll show you
4:11exactly how I built it, and if you want
4:12to know how this experiment goes, and
4:14especially the second part with real
4:16money, remember to join the community
4:18and subscribe to the channel because I
4:19do experiments like these here every
4:21week. And now let’s see how it’s
4:23built. Okay. Let’s start this
4:25adventure, and I know what you’re
4:27thinking because it’s what I’m
4:28thinking right now, too. "Nice,
4:31Giuseppe, but I don’t know how to
4:33program." Perfect. Neither do I. And
4:35above all, I won't write a single line
4:37of code to realize this project, just
4:38as I didn't write any for the previous
4:40crypto project. So, this is the day we
4:43hire our personal trader, our super
4:46trading expert who will make us
4:48millionaires. Hopefully, we'll find out
4:51at the end of this experiment. Okay, to
4:53make it real, just like last time, we
4:54need two things. We need Claude Code,
4:57which builds our entire system without
4:58writing a single line of code,
4:59obviously. In the description, you'll
5:02find all the prompts, resources, and
5:04everything you need to do it, but watch
5:05the video first to understand how it
5:07works, in my opinion. And the second
5:09thing is a platform to trade on.
5:11Actually, not all platforms have this.
5:14You need a platform that has a
5:16connection or API, so it has the
5:17ability to allow Claude to communicate
5:19with the platform, or at least an MCP
5:21server. Without getting too technical,
5:24we need the two things to be able to
5:25talk to each other. For the crypto
5:27experiment, I obviously chose a crypto
5:29platform, but to operate on stocks and
5:31markets, we need something different. I
5:33evaluated several for this specific
5:35experiment, which will be part one, and
5:37I decided to use a platform called
5:39Alpaca. Yes, it’s a very nice name
5:42too, I must say, and the platform is
5:43very yellow, too. For the second part,
5:45I’m using a surprise one that I’ll
5:47tell you about later. And so, let's
5:49start building our super automated
5:51trading system. I chose Alpaca because
5:54it’s a platform designed specifically
5:56for the needs of those who need to
5:58integrate some type of application—in
6:00our case, AI—with the platform,
6:02unlike others that adapted later.
6:05That’s the first reason. The second
6:07reason is that you can open a demo
6:08account, so you can test the system for
6:10free, without documents and without any
6:12hassle. Alpaca also offers an Italian
6:14account with documents. Honestly,
6:17it’s in a declarative tax regime;
6:19I’ve never tried it with real, actual
6:20money. I’m testing it for this
6:22experiment, then for the experiment
6:24with my actual € 10,000, we’ll see
6:25another platform, but I’ll tell you
6:27about that later. Registration is
6:29really easy. You just go to "Create an
6:31account" and bam, in one second with
6:33just an email, we're in. We have our
6:36portfolio with our $ 100,000,
6:38unfortunately virtual—we can't
6:39withdraw them—but they allow us to do
6:41practically everything, meaning we can
6:43try out complete real trading with real
6:45stocks, everything real that we want.
6:48In short, this is the one to do as a
6:50first option. So we have our trading
6:52platform; if we click on API Docs, it
6:54shows us everything we need to know—
6:55or rather, we don't need to know
6:56anything because Claude Code will do it
6:58all, which I’d say is fantastic. It
7:01really is a great platform, aside from
7:02the way we are using it right now. To
7:05generate the keys that Claude will ask
7:06for, just go here to the bottom right
7:08to API Keys and generate new API keys.
7:10Claude will also suggest a system for
7:12not pasting them inside the chat to
7:13maintain security. So I’d say we’ve
7:16finished the setup of our trading
7:18platform, and I’m about to launch my
7:20own project on artificial intelligence
7:22and money. Absolutely insane. If you
7:25don't want to miss it, remember to
7:27subscribe to the channel and also to my
7:28newsletter in the description. Thanks
7:31to the experience gained with the
7:32crypto trader, the second thing we need
7:33is to be able to communicate with our
7:35system, with our Castagna Capital. "CC"
7:38is cute, but I need to register it.
7:39Castagna Capital. The best way to do
7:42this today is still Telegram, so we are
7:44essentially building a link between our
7:46Telegram—meaning our phone, the
7:47ability to dictate whatever we want—
7:49and our Castagna Capital, our trader.
7:52To do this, we are going to use
7:54BotFather with a very simple procedure.
7:58I’ll leave all the info for this for
7:59free in the description as well. The
8:01setup is really simple; just type "new
8:04bot," it gets created at the end, and
8:06we'll have the key to paste into Claude
8:08Code to make the connection. Claude
8:10does everything automatically. In any
8:12case, by using my guide and
8:13instructions, you just drop them in,
8:14and it already knows everything it
8:15needs to do. You practically don't even
8:17have to press a key on the keyboard.
8:19Perfect, we have our voice. To build it
8:22now, I chose a different path than the
8:24classic "I start writing prompts, I
8:26talk to it, or I tell it what I want to
8:28do." I decided to first create a kind
8:30of contract with it, an MD file where I
8:32give all the instructions for it to
8:34then build. This way, I avoid a whole
8:37part of the code iteration process that
8:39becomes a bit heavy or a bit long, or
8:41at least forces me to spend a lot of
8:42time in front of the computer. In the
8:45contract I created in a Claude chat,
8:47based on the conversation we had before
8:50—so based on the strategy we made
8:51before—all the things I want it to do
8:54are indicated: the goal, the
8:55constraints (so only paper trading),
8:57the alarm that starts without memory,
8:59which tools it must use, how it must
9:01use them, and the logic with which I
9:03want it to wake up. using Telegram, in
9:07short, everything it needs to build our
9:09system. The third element we need is
9:12obviously Cursor, so let's open a new
9:14terminal window. I personally use
9:17iTerm2 with Cursor. It’s an amazing
9:19Mac app, it replaces the terminal, it's
9:21totally free, and I’ll leave the link
9:23in the description; it works great.
9:25Actually, to see things graphically if
9:27you want to change something, it’s
9:28even better to use the Cursor app by
9:30going to the code section. But you can
9:32do the same thing with ChatGPT, Claude,
9:35Google, in short, whatever system you
9:37want to use, the important thing is
9:39that it has a system that creates code.
9:41I find Cursor works well for me with my
9:43Pro subscription, but ChatGPT or
9:45whatever else you want is fine too. My
9:47guide works for all AI systems. So I
9:51create a new folder, put our contract
9:53inside it, and tell it to read
9:54everything, show me the plan, and then
9:56start building. Okay, perfect. In about
10:01fifteen minutes it built our entire
10:03system, and here I make it do something
10:06few people do; I don't trust it, I make
10:08it break the system and ask it to do a
10:10stress test to see if it works or if
10:12something is wrong. So I technically
10:15ask it to play devil's advocate against
10:16what I’ve already built. I ask it to
10:18break it in at least six different ways
10:20and to really test them. Okay, it found
10:23three real holes. It corrects them and
10:27is effectively improving its own code.
10:29It's impressive stuff. I don't know
10:31where we'll end up. Done. It took me
10:33about half an hour, including
10:34registration and everything. I’d say
10:36that's a very acceptable amount of time
10:37. This is the day zero dashboard. We
10:40are ready to start. We have $ 100,000.
10:43The first wake-up call is set for
10:45tomorrow morning at 9:00, so our 24/7
10:46trader has been hired and is
10:48operational, and now I’m going to
10:49sleep; let's see what happens tomorrow.
10:52Okay, let's see how day 1 went. At 9:00
10:55in the morning, I received my
10:57notification, it had its plan, I’d
11:00say all in all it went well; after the
11:03Wall Street opening, it executed the
11:06plan and bought us 20 SPY, 10 QQQ, 25
11:09GLD for a total spend of $ 32,000, 33%,
11:12and held $ 67,000 still. It says haste
11:15is not a strategy. Well, the thing that
11:17seems strange to me is that I received
11:20this notification while I was out this
11:22afternoon visiting a client, and having
11:24done nothing, well, it’s strange,
11:26it’s really strange. Mid-session it
11:29checked again, all for nothing, then at
11:319:00 PM closing. So let's see the
11:33balance. Plus $ 76. Well, that's not
11:37bad. And here's the move I was waiting
11:41for; the last line tells us to look at
11:43Microsoft and Nvidia tomorrow morning,
11:45which closed better and is the
11:47candidate for a small fourth position.
11:49The cool thing about this system is
11:51that every time it wakes up, it
11:53remembers basically nothing. In this
11:56case, it literally doesn't remember
11:58yesterday's trades, it doesn't even
11:59remember being a trader, so before
12:01shutting down, it writes to its future
12:03self what it did and what to check
12:04first. It's basically an insane
12:06employee that creates its own memory
12:09time after time, so it never starts
12:11from scratch and can relearn from what
12:13it did before, using the best model we
12:16have available at that moment. So,
12:19crazy. So, first day, plus $ 76 and a
12:23crazy plan for Tuesday. Okay, let me
12:25show you what happened after 3 days. On
12:28Tuesday, it followed the plan, Nvidia
12:30closed very strong, so it decides to
12:32buy 30 shares and the account is at the
12:34weekly high, we're up $ 180. And so the
12:37trader, at 3:40 PM (I get the message
12:39on Telegram), does something very
12:42interesting. It says: "Hmm, I propose
12:44increasing the position by 90 shares,"
12:46so it wants to buy because the price
12:47dropped, but the program tells it no. I
12:50mean, this is the first time the system
12:52stops and doesn't reason based on
12:54emotions, because buying those shares
12:56would hit 21%, while the rule tells it:
12:58never more than 20%on a single stock.
13:00So the order doesn't go through. And
13:02then at 6:40 PM, it writes to me that
13:04the position was rejected by the
13:05guardrail because the rule is more
13:06important than its thesis and its
13:08prediction. So, very interesting. The
13:10cool thing is there's no "come on, this
13:13time is different, we can do it." If
13:16you give it a rule, it follows it,
13:17unlike us humans who, especially in
13:19trading, are definitely driven by
13:20emotions, and so the account for the
13:22first time on the third day closes in
13:24the red. Okay, let's see how this first
13:27week of automated trading and
13:28experiments went. Actually, it didn't
13:32go very well, we lost 0.12%, while the
13:36index was up 0.76%this week. They are
13:40incredible movements, but we lost
13:41against the market and it's the one
13:42that tells us first. It placed five
13:45orders with one stop, all rules
13:46followed, and there's a new rule for
13:48next week that tells us individual
13:49stocks should be bought on pullbacks
13:51and not at highs, and before every
13:52purchase, it writes the point where I
13:54would be wrong. Interesting as
13:57self-correction. And now I’ll give
14:00you my sincere and honest opinion on
14:02this system, on these systems in
14:03general, and on how online trading can
14:05actually work, even though I lost 0
14:06point something. However, the system
14:09works, and it works not because it
14:11understands the market better than you,
14:14me, or the pros, but because it has no
14:16fear, it has detachment; it will follow
14:18the rules we give it to the end and
14:21could even refine them or suggest new
14:23ones based on the ones we've provided.
14:25That is the fundamental difference
14:27between an automated system and a human
14:29being who invests. The best prospect I
14:31see for these systems, however, is not
14:33day trading, because you would need a
14:35system too fast that isn’t accessible
14:37to us, the poor retail consumers. The
14:40best function I imagine for these
14:42systems is investing based on news and
14:44what happens, but based on our rules
14:46and over a medium-to-long term, because
14:48day trading in the end, in my opinion,
14:50always remains a gamble. It’s a bit
14:52like playing at the casino. So,
14:53what’s the next step? Well, obviously
14:55to test it with my own real money in my
14:58trading account. And if you want to
15:00know how it goes with real money, join
15:01the community by clicking subscribe,
15:03and you can find the video here on the
15:05right. Thank you for watching and
15:07always unleash your potential. Bye

About this transcript

This page contains the full transcript of Ho Trasformato Claude in un Trader Automatico 24/7 by Giuseppe Castagna, generated from the public captions YouTube serves with the video. The transcript has 2,957 words across 434 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.