Everything I've Learned Trading With Claude In 18 Minutes

0:00 I spent the past 2 years using Claude to build trading systems both for myself and for my clients. So in this video, I'll go over everything I learned, so this way you'll know exactly what Claude is actually good at. You'll know its strengths and weaknesses and how to properly adapt it into your own trading strategies. Now, I've studied math and econ at UCLA and worked in investment banking at Raymond James for over 3 years, so I can come at this from a quantitative finance background. This way, I can go over what actually works, what doesn't, the mistakes everyone makes when using Claude, and how to properly use Claude for your own trading. And as always, this isn't financial advice and I'm not guaranteeing any profits in this video, just simply showing you how to leverage Claude in your own trading. Now, to start, let me get this out of the way, Claude is not going to find you a profitable strategy by you just asking Claude, "Hey, what's a profitable strategy?" Claude also won't predict price for you, so make sure to get that out of your head as well. To start, you need to understand the limits of what AI models like Claude can actually do. And the one concept that's the most important here is the idea of deterministic versus non-deterministic systems. These are the two main paths you can use Claude for in trading.

1:08 Deterministic just means systems with fixed rules, while non-deterministic systems don't have fixed rules. Claude can basically do both of these, but there's a lot more nuance to this, so let me go ahead and explain both in detail. So starting off, a deterministic system has logic that never changes. What this means is that if you were to run the same exact system with the same data again and again, you'll always get the same result. And the main use case here is that we can use Claude to basically build out deterministic systems for us. So this could be an algo that's basically an automated trading bot. You can also have a signal engine.

1:39 And a lot of the stuff that you've seen on my channel that I've covered like walk forward back testing, regime detection models, these are all basically deterministic systems. And because these systems have fixed logic, that means that we're able to back test the results and see how it actually performed historically. The key here is that Claude can do almost all the technical coding needed to actually build out these deterministic systems without requiring an entire team of developers. And just as important, if not more important, is that you can actually spin up new strategies in a few hours and basically test multiple ones throughout the days every single day.

2:11 Now, non-deterministic systems, on the other hand, is more so like using Claude as an analyst or an advisor, rather than a builder that basically writes code for you. So, instead of writing code, this is where you would ask Claude to basically, you know, review setup, analyze earnings reports for you. It can even give opinions on your portfolios or even give long-term trade ideas as well. And you can basically do all of this in the Claude web app. If you just go to Claude.ai and ask these questions, they'll be able to give you some outputs. But the key difference here, and what's most important to understand, is that each time you ask a question, the answer will be slightly different.

2:45 This is just because of the way that LLMs work. Claude isn't running some fixed formula inside a deterministic system. And this is super important to know because these non-deterministic outputs are able to be back-tested. There's absolutely no way to go, you know, 6 months in the past and determine what Claude would have said to you if you asked for a trade idea, or if you asked it to, you know, review an earnings call. You'll never have that level of visibility as to what Claude would have given you as an answer. Now, this doesn't mean that non-deterministic systems are useless. You just have to know how to combine or use either or depending on your goals and what you want out of this trading system or framework. So, starting with the deterministic side, this is where Claude is most useful and powerful for traders.

3:27 You can use Claude code to essentially build out full automated trading systems. And with the proper structure, you'll have something that's rule-based, back-testable, and repeatable. And the key thing that differentiates a good system from a bad system comes from a few components. So, first, the rules have to be specific enough so that there's no ambiguity. So, a quantitative deterministic system would have something like buy when the 12 minus one month momentum score is above the 80th percentile in its sector. Now, second, you need proper validation, and this is where Claude really separates from something like TradingView. So, TradingView can backtest a very simple RSI crossover strategy or a moving average strategy. This is all fine for simple stuff, but the minute you need something a little bit more complex and accurate, it falls apart. So, things like walk-forward validations, Monte Carlo simulations, and sensitivity analyses, you can't do any of these in TradingView. But, with Claude, you can build all these validation layers by having it write Python code for you. And this way, you can actually stress test your strategy before risking any money at all. I think this is the biggest part where most people go wrong because typically people just look at a successful backtest on TradingView and then assume that their specific strategy is going to work, when in most cases, once you bring it to the live market, it completely falls apart. And then, beyond just backtesting, Claude can basically just build more complex components that TradingView simply can't. So, you have things like portfolio-level systems with sector neutrality, beta adjustments, factor risk decomposition, mean-variance optimization, transaction cost modeling, and again, these all have defined rules with data, so that each time you run the system, the logic never changes. So, here's an example I've done for one of my clients. This is a cointegrated pair system we built for commodities. The rules are completely deterministic. So, it scans for pairs that are cointegrated, monitors the spread between them, enters when the Z-score exceeds a threshold, and then exits when it reverts. Every single step is defined, every step is backtestable, and the system runs the same way every single time. So, the main point of these deterministic systems is that you can truly validate your idea before ever putting any money into the system. Now, moving on to the non-deterministic side, this is where most people get Claude wrong. So, non-deterministic is when you're using Claude more like a research analyst.

5:44 You're basically asking it to think and interpret an output, and every single time produces an answer, the output is always going to be slightly different since it's reasoning all over again. Now, because of this, like I mentioned earlier, this isn't able to be back-tested at all. You're never going to reliably get the same exact output from Claude if you were to ask it the same question. Because of this, you should never build your core trading strategy primarily on these non-deterministic Claude outputs. But, as a supplementary tool, it's extremely effective. Now, if you want to use Claude in the most simple way for non-deterministic outputs, the number one most important thing is to always give enough context. So, this includes things like your trading style, your current positions, your thesis, your time horizon, and even your risk tolerance. So, the more specific you are, the better the output's going to be. So, what I mean by this is if you just ask Claude what to buy, it's going to give you generic garbage. But, if you were to say that you trade momentum on large-cap tech with two to three-week holding periods, and even say the specific positions that you're currently long on, and you're worried about sector concentration, now Claude actually has context and can give you something useful. And in my opinion, the absolute best way to use non-deterministic outputs from Claude is as a reviewer and not a generator. I would still come up with your own trade ideas first, then I would ask Claude to actually poke holes into it. This way you can see what's the bear case, what you're currently missing, and what risks you're under-weighting. And this is actually how the best analysts at funds perform.

7:10 They constantly stress-test their own thesis. Now, another thing that's super helpful from non-deterministic Claude is research. Because LLMs are able to ingest so much data, you're able to feed Claude earnings transcripts, uh 10-Qs, 10-Ks, and other documents at scale. You can see which companies have the strongest revenue trajectory from the latest batch of reports, and you can also just flag anything concerning on any companies you're looking at based on the risk factors. Now, here's another example of a project I made. This is where you can have Claude essentially analyze options chains for you. Now again, this is non-deterministic, so it shouldn't be part of your core strategy, but Claude is able to give you judgment on which strikes look interesting based on IV rank, skew, delta, any upcoming catalysts, and it's a good way to just quickly scan a chain and get a second opinion on what stands out. And then the final thing I want to touch on between the non-deterministic and deterministic is pricing. So, when you're building deterministic systems, you can use things like Claude Code, which is basically just Claude's coding agent.

8:09 And then with this, all you really need is monthly subscription. So, Claude, there's no ongoing cost once the actual deterministic system is built. But then in those same exact systems, if you want to add non-deterministic outputs from Claude, it's going to cost per usage. So, basically, each time you want Claude to non-deterministically analyze a trade for you or some reports, it's going to cost a few cents every single time. Now, let's go over where Claude fails in the non-deterministic mode. So, first and most importantly, don't ever ask Claude for price predictions. It absolutely cannot predict prices, and it will actually give you a pretty confident-sounding answer, but I would avoid it as it's going to be completely random. And then next, whether using the Claude web app or you're adding in a non-deterministic version of Claude inside your system, don't ever send an image of a chart and have Claude essentially analyze it for you. The image analysis for trading patterns are still not reliable. So, for example, it could tell you that, you know, a head and shoulders pattern is there when it really isn't. It's much easier to bake in technical indicators like this inside the deterministic systems, since again, this is logic that is fixed and repeatable. And then finally, don't take Claude's trade advice without having your own thesis first. So, Claude will generate plausible-sounding trade ideas all day long, but most of them are going to be generic. So, your own market understanding combined with the non-deterministic analysis from Claude is where the system is really going to shine. Now, let me show you an actual example of how to combine both deterministic and non-deterministic outputs into one system. And before we start, if you want to learn exactly how to build both these deterministic and non-deterministic systems with Claude, feel free to check out my community in the description. It's currently the largest trading community out right now focused entirely on AI. So, all our members there are testing and building AI trading systems across every asset class, so stocks, options, crypto, and even futures. And if you already have an idea of what you want built, I also offer one-to-one custom builds in the description as well on my website. Now here, what I have is a three-layer trading system with the first two layers fully deterministic, so always the same logic. And then finally, we have the third layer where Claude will essentially act as an analyst on top.

10:19 So, the first layer here is a macro deployment gate. So here, before you even look at individual stocks, the system checks six macro signals to decide whether you should be actively trading right now. First here, it's pulling the VIX level and this is where it sits inside the one-year range. The VIX term structure here tells you if the options market is calm or stressed. Market breadth here is what percentage of S&P 500 stocks are above their 200-day moving averages. Credit spreads between high-yield treasuries, put-call sentiment, and then finally factor crowding. So, each one of these gets scored between 0 to 100 and then blended into a single deployment score. So, all this means is that if this composite score here is above 70, it means full deploy. 40 to 70 means you would reduce your sizing. So here, we have around 60% of normal sizing. And then if the overall macro score was below 40, we just wouldn't take a trade at all. And then again, all of this is completely deterministic, same data in, same data out. This is basically just Claude writing code for us here. So with all these scores, how we basically read this is that the VIX term structure is 99.9, which is pretty calm. Credit spreads are around 79, which is healthy. But then factor crowding is around 10, which is dragging the whole thing down a little bit. Basically, the correlation between momentum and value factor returns is at negative .69. So, basically, there's just some risk building underneath the whole system. And then, this is why even though VIX is low, the credit is fine, but the factor structure is a little bit suspicious. We don't have full deployment, and then we're only going to deploy at 60% of normal sizing. Then, scrolling down, we can actually see a historical overlay on the spy chart characterized by zones over time. So, these green zones here are when you would typically deploy more capital, and then the red zones are obviously when you shouldn't trade at all. So, again, this is the first deterministic layer.

12:08 And then, if we head over to the scanner, this basically gives us the second deterministic layer. All the first layer told us is that the market is okay. We should deploy still at around 60% of what we normally would. And then, these are the top individual ticker candidates that we should deploy into. And then, recall again, because this is deterministic, everything is quantitative, right? There's no subjectivity here or opinions. We're just simply looking at numbers, and then based on the market, we're determining which tickers are best suited in this environment. Now, this is all fine and dandy, but the way we would actually combine systems is that once we essentially filter out the market, filter out which tickers perhaps look good based solely on quantitative factors, we would then bring in Claude as an analyst that is going to essentially go through those tickers and then tell us what it actually thinks.

12:55 So, over here, the system is going to take each of the six candidates. We're sending them the last four quarters of financial data to Claude through the API, so they have revenue, cash flows, margins, debt, the whole picture, basically. And then, Claude essentially scores each one from zero to 10 on fundamental quality. Now, like I mentioned before, this is not backtestable, right? You won't know what Claude would think about these tickers a week from now or even 5 days before, right? So, each time you run this, you're going to get slightly different answers. But, because we're feeding all the financial data to Claude, we can at least hope that it's making a level-headed decision based on all the context it has. And then again, like I mentioned in the beginning of the video, this part does cost tokens each time you run it. Now, because the macro gate and scanner are all numbers focused, um it it's just code, right? You don't need to actually run any of these through Claude. But once you go to the analyst, each time you click run analysis, it's going to cost a few cents just because it's taking in all the data and then outputting something back to you. So, we're able to first see the quantitative score, which again is based on those two deterministic factors. But then we can also see a Claude score to get a blended score that gives you maybe a more holistic picture as to what tickers could be a better entry. So, some tickers like TXN could be ranked lower on the quant model, but then once you add more fundamental analysis, uh give the actual reports, have Claude analyze them, they could rise in rank. And like I mentioned earlier, I wouldn't ever just blindly trust Claude, right? We would want either a combination or have this act as a supplementary tool. And that's exactly what we're doing. So, we're not just taking the Claude score and saying, "Oh, it's like the perfect candidate." We would look at the quant scores, the Claude scores, blend them together, and you can even choose how much to weight the actual Claude score, because again, this is fairly subjective and you're relying on its intelligence to give you a score. And then if you were to scroll down, we would also know candidates that we should never really enter into, which is Ford. Um that's just because it scored the lowest on the quant models, but also Claude gave it a four out of 10. So, you can basically use both of these analyses to kind of come up with a more composite score that reflects both the quantitative and qualitative nature. Now, let me show you the actual prompts that I used to build this same exact system so that you can do the same. Now, one thing I want to point out is that if you take these exact prompts here and then run them through Claude yourself or Claude code, the actual code that it writes and compiles for you will be slightly different from mine. So, what this means is that maybe it can structure a function differently, uh uses a slightly different approach to the same calculation. But once that code is written, saved, and debugged, the system will run deterministically. So, you can think about it like this. We're using a non-deterministic tool, which is Claude or Claude code, to build a deterministic system. And then because of that, the system that we're building is going to be slightly different each time you build it just because Claude by itself is non-deterministic. So, this is the first prompt for the deployment gate.

15:39 Basically, it's just telling Claude code to pull six macro signals from Yahoo Finance, score each one from zero to 100 based on specific rules that I designed. We want to blend them together into a composite with these weights and then set the deployment zone based on the score. It also tells you to build a historical backtest so you can see how the gate would have performed over the last 2 years. And if you want to pause and screenshot this, you can copy it directly into again Claude or Claude code and build the same exact system.

16:04 So, now moving on to layer two, here we have the scanner. So, this will tell Claude or Claude code to basically pull the entire S&P 500 universe, score every single stock on five factors, and only show the candidates above a certain threshold depending on what the macro gate says. And then finally, we have layer three, which is the non-deterministic Claude analyst. This is the prompt that will basically tell the system to take each candidate, gather their last four quarters of financials, send it to Claude via API, and have Claude score it either 1 through 10 on earnings quality, growth, balance sheet health, margins, uh red flags. Then what we're going to do is blend that score with the quant score on a 60/40 split and give a re-rank. Now, if you enjoyed this video, make sure to leave a like, comment, and subscribe as it helps out the channel. I create weekly videos focused entirely on using AI for trading. And like I mentioned earlier, if you do want to go deeper, my community is in the description. All of our active members are all using Claude's build everything from fully autonomous trading bots to quant frameworks and systems to even analyst systems like these where they're having Claude essentially give qualitative analysis on specific tickers, setups, or systems. I have step-by-step guides for all of this and they're built so that even if you have zero coding experience or knowledge, you can follow along and build real trading systems that just weren't possible for retail traders before. I truly believe that AI is going to be the thing that closes the gap between retail traders and institutions.

17:31 And if you're not using it at least somewhere in your process, you're definitely going to fall behind people that are implementing it in some way, shape, or form.