Is AI Killing Your Trading Edge? Do This Instead
Adrian Reid shows why generic AI prompts produce losing trading systems and walks through a disciplined workflow instead: define rules precisely, backtest with real data, run robustness checks across parameters and markets, then use AI as a research assistant, not a decision maker. He demonstrates the setup with Claude, RealTest, and Norgate data.
Chapters
Most traders are using AI completely wrong and it's costing them money because right now people are typing in things like, "Give me a profitable trading system" into ChatGPT and getting a confident answer and then risking real money on what AI tells them to do. And here's the really dangerous part. AI sounds so confident delivering bogus answers that even smart traders are getting misled. But if you're serious about systematic trading and you're a busy analytical professional, what I'm about to show you is going to transform the way you trade and the way you use AI. I'm going to teach you in this video not to use AI to replace your judgment or to generate some magical indicators, but to compress years of research into months and weeks. And how to do this while still validating everything objectively with real solid backtesting. I'll also show you the exact workflow I use to turn a real market observation into a solid backtested trading system, accelerated using AI, but backtested correctly using a proven process. So this is AI in systematic trading, how to use artificial intelligence to build, test, and improve your trading systems faster than ever. The goal of this video is to show you exactly how to use AI to build and test trading systems safely and efficiently. So here's the dirty little secret of AI. Have you ever done something like this? Typed into ChatGPT, give me a profitable stock trading system for US stocks. If you do this, you'll quickly get a very confident answer. And I did this exercise. And here on the right of the screen are the rules that ChatGPT spat out. A system for US stocks. Here's the entry strategy. It focuses on liquid stocks with a price greater than $5. Stocks above a 50-day moving average. The stock shows upward momentum according to the RSI. The price pulls back to support. Volumes above average on the breakout. Exit strategy, profit target of 2 or 3 times your risk. Stop loss, trailing stop, risk management. All of the different components that you'd expect. And it sounds very plausible. It sounds great. It has some conventional wisdom in there. Things that people would be looking for like liquidity and price filters and momentum filters and so on. But here's the trick. Would you actually trade it with real money? Should you actually trade it with real money? And this is the problem. I backtested this trading strategy exactly as given by ChatGPT, and guess what? It failed abysmally. And so in this video, I'm gonna show you why it failed and what you can do about it. So this system that ChatGPT so confidently presented with us is an absolute losing monster. It lost 6.7% per year. It basically lost throughout the entire backtest. By the end of the backtest, it was down 96% and only 29% of the trades were winners. So this system confidently given to us from AI was an absolute disaster. So let's talk about why AI-generated systems are almost certainly gonna be unprofitable. AI has been trained on thousands of forum posts, blog articles, textbooks, including Reddit, X threads, and social media. And the problem is, Most of this training data describes approaches that don't work, aren't tested, or are overfitted to historical data. The broad internet does not know how to trade well. There's far more people talking about trading who aren't making money than people who are, particularly when you take into account social media, Reddit threads, and things like that. So AI can't distinguish between a good idea and a plausible-sounding idea that's repeated many times. What AI does in its training is pattern match. It looks for common themes. It looks for things that are repeated. It looks for things that are cited by reputable sources. Now, none of these things correlate with a profitable strategy. They all just relate to how frequently and who says the things. So the AI training data pattern matches on text, not actual market data. So AI hasn't done any backtesting when it presents you with these rules. And the confidence of the language that AI models use has zero correlation with the quality of the output. So with all of this conflicting information, if you digested every single piece of information on the internet and all of the social posts about trading, do you think you could take all of that and synthesize it immediately, instantly into a profitable trading strategy? Of course not. There's so many conflicts, and that's why AI natively, when asked for a trading system, doesn't give you a profitable result. So what are we gonna do about this? The only escape is to use testable rules-based systems. So here's the one thing you have to do before you ever act on any AI trading idea. You've gotta test it objectively. If you can't test it objectively, you can't know if it works. And AI ideas have gotta be expressed in complete objective rules before they have any value. So it's no good to just take bits and pieces. You've gotta convert that into a complete testable trading system and then backtest it correctly. Otherwise you've got no business trading it 'cause you've got no idea whether it's actually profitable. Because as we've said before, AI training data is not proven backtested systems. It's just the text that it finds online. You need every entry and exit rule and position sizing rule defined objectively and precisely, and then you can code it, run a backtest against historical data, and find out whether you've actually got a profitable idea. So here's a simple framework you can use to make sure your idea is complete before you start trying to test it. You've gotta have every component to make the system work. So we need the universe and the data. We need to specify what universe of tickers is this gonna trade on. Also what bar size and what timeframe are we gonna backtest this over? We need to know what the filters are, market regime filter, liquidity, volatility, price filter, and any other filters. We need the entry rules, which might consist of constituency within a particular index, maybe stocks in the NASDAQ 100, for example, the setup, the trigger, the order type, the trade direction, and the re-entry rules. You need position management, which includes position sizing, position score, max number of positions, and any pyramiding rules. And you need exits, which includes stop loss, trailing stop, time stop, exit rules, and exit order types. And then finally, portfolio and backtest settings, starting capital, commissions, slippage, and what you're benchmarking the system's performance against. Now, commission and slippage is something that a lot of traders leave out, and that's a huge mistake. You've gotta include it, otherwise you're gonna be driven to create a strategy which is very short-term, low expectancy, and has a laser-sharp equity curve. But as soon as you add real commissions and slippage, it's gonna fall apart. So of course, this framework is very comprehensive. I'm not suggesting you need every single one of these in every single system. That's certainly gonna give you too many rules and you're gonna end up with an overfit mess. But this is a framework for you to make sure that you've got everything you need for your particular type of strategy. Okay, so we've got our ESCAPE, which is testable rules-based systems.. But even with backtesting, there's a bit of a problem you've still got to be aware of. That problem is garbage in, garbage out. So here's the problem. If AI's training data included an overfit system that backtests amazingly well and a really confident write-up caused AI to pick those rules up and present them to you as a trading idea, the backtest could look extremely profitable. The trouble is, if that is overfit, then in the future it's likely to fall apart and stop working. This means we need to do a lot more than just a simple backtest to check that a system generated by AI is actually profitable. We need to check parameters for robustness, we need to check all of the rules for significance, we need to check that the edge is stable over time, and we need to validate on in-sample and out-of-sample data, and ideally on related markets to make sure that the system holds up. Once we do all of that and then incubate the system, maybe we can start thinking about trading it. But before that, it's extremely risky to consider trading an idea that AI has generated for you. So this comprehensive process is what's going to protect you from losses caused by AI-generated systems. So here's the right mental model. AI is a research assistant and coding tool, and you are the responsible trader. You have to combine these two things. So you still need a process, you still need backtesting, you still need judgment. You're not abdicating everything to AI. Quite the opposite. You're actually using AI to accelerate your efforts. So AI doesn't replace the trader's edge, it compresses the time cycle. The system development process, that doesn't change. It's exactly the same system development process you would use if you were doing it manually. But if you don't have a system development process, then that's the first thing you need to learn to master systematic trading. And that's exactly what I teach in the Trader Success System if you need to get a head start on that. A trader who used to be able to test one idea a week can now test 20 ideas a week. And over a year, this compounds into a massive research advantage. But it only works if your testing process is sound. If you don't have a sound process underpinning all of the AI acceleration, all you do is crash the car faster. So you really do need to understand the principles and processes of systematic trading to drive AI correctly to get a system that will work. So here's the contrast when you actually get this right. The old way on the left is a trader trying to figure out coding, trying to understand the formulas, and manually doing all of the coding. The Claude way is far simpler and far more relaxed if you do it correctly. When you use Claude to do your coding and accelerate the process for you, interacting with your backtesting platform through Claude, then all of a sudden things get a lot faster, a lot easier, and a lot more relaxed. So here's what AI is extremely good at. It's good at coding trading systems, but only once you teach it the right syntax with skill files and framework to generate a system that works. It's good at backtesting trading system ideas, but only once you define the backtesting process rigorously so AI knows exactly what steps to follow and how to make each of the backtesting decisions. I recommend using your AI through the command line interface with RealTest so that you're backtesting it correctly. We know that RealTest is now a standing backtesting platform, and by using RealTest to do the tests and Claude or Codex to do the coding and interaction with RealTest, then you can really accelerate things. AI is also extremely good at brainstorming ideas, but only once you give it sufficient context to get ideas that are relevant for you in the system that you're working on. AI is also great at analysis and filtering once you've set objective criteria on how to make each filtering decision. AI is also extremely good at documenting your journey, and this is one of my favorite aspects of this because AI can record every single test and result and conclusion that it draws and brainstormed ideas all into the markdown files when it's working through the process, which creates a hugely rich source of information for you to come back to later on and cross-reference with other strategies you're working on. AI is great at learning from past backtests and system experiments, particularly once you link it up to the documentation that we mentioned previously. It's also great at following a consistent process. Most traders jump from here to there all over the place. They go down rabbit holes, they make mistakes, they loop over optimizations multiple times. Whereas if you program the process and steps into AI, it's going to run through the process the same way every single time, and you're going to get much more consistent results much more quickly. And then finally, it's really great at brainstorming new ideas. At the end of every system run, I ask What did we learn in this process? What ideas can we generate as a result of those learnings? I ask it to give me 3 to 5 new system ideas based on the learnings in the backtest we just ran. And that gives us a huge variety of ideas to test. So here's how this approach and this mental model compounds into a massive research advantage. The old way, doing things manually, maybe you generated and coded an idea each week. And if you're lucky, 10% of those might end up in incubation. Over a year, that means 50 ideas tested and maybe 5 ideas into incubation. What AI does is multiply your efforts dramatically. So now you can test 20 ideas or more each week, which results in 3 systems going into incubation on average. And over a year, that means over 1,000 ideas tested and over 150 systems into incubation. So it really dramatically accelerates your progress, gives you much more options, and allows you to build a much stronger portfolio. Not because AI is doing all the work for you, because AI is speeding everything up so you can get through more ideas, more iterations, and build more resilience and diversity into your portfolio. But before we go any further, there's a couple of things you need to understand that you should not be doing with AI, and here they are. Don't ask AI for the best trading system. There is no best trading system. It'll confidently give you something mediocre or probably even worse. You can't trust AI's opinion on whether a system works unless you backtest it. So yes, generate the code with AI. Yes, brainstorm ideas with AI, but you can't take its word for it. You have to backtest it. And AI coupled with RealTest is a powerful way to do that. You can't skip your validation process because AI sounds confident. You have to do every single step in the process rigorously. You also shouldn't use AI-generated parameter values without testing them. AI will choose parameter values based on its training data, based on all of the references on the internet and on social media. So you need to validate those by running a parameter sweep and optimizing correctly to make sure the parameter values you choose are stable and robust. You know, just to hint as well, I would say don't share proprietary signals with free public AI tools. Use paid versions and have the privacy turned on so it doesn't use your backtest, your results in its training data. You wanna keep your ideas to yourself. Also, don't ask it to backtest your trading ideas unless you've set up the right infrastructure. If you tell Claude or ChatGPT to backtest your trading system, it might do something. It might install some Python libraries and it might try and run a backtest if you haven't defined it, but you have no idea whether that's gonna be valid. What you need to do is instead connect your AI system to a backtester like RealTest using the command line interface so the backtests are run correctly. The AI can then drive the backtest and make sure the results are valid. So you regenerate the coding in your AI platform. I'm using Claude. You send it to Realtest using the command line interface. Realtest sends the results back. Your AI interprets them and then generates the results and next steps. Then finally, I would just say, don't let AI trade for you. There's a huge difference between automation and AI. AI is for doing work. It's for generating ideas. It's for doing analysis. But executing your trading is a task that requires 100% repeatable process. That's a job for automation, not for AI. AI is better for thinking analysis. Automation is a script that allows you to place your trades mechanically the same way every single time. Don't use AI to place your trades for you. So here's the AI system setup that I'm using. In the center of all of it, I've got RealTest to do my backtesting. I use Norgate Data to ensure that I've got high-quality data that's reliable for all my backtesting and trading decisions. Now RealTest does the backtesting, but Claude is doing all of the coding and controlling the process. I use the command line interface to connect Claude to RealTest. So Claude does the code, sends it to RealTest, validates it, imports the data, runs the backtest or optimization, and then the results come back to Claude. In order for this to work, you need a couple of things. You need some skill files. You need the coding syntax, which is supplied with RealTest. I've got that plus my own supplementary syntax, which defines how I like to lay out a system system, how I like to name everything, how I like all the comments done, and the components I do and don't want in a system. You also need the backtesting process fully defined. In the Trader Success System, you learn the Expert's Guide to Backtesting, which is a step-by-step process for testing and optimizing a trading system and avoiding overfitting at the same time. It generates reliable systems that you have confidence in. You need a process like that to explain to Claude what the steps are so that it knows how to run a backtest. Natively out of the box, Claude does not know how to backtest. Neither does ChatGPT. None of them do. All they know is what's in their training data. And yeah, there might be some interesting useful information about how to run a backtest in the training data, but it's not gonna be your process. It's not gonna be the right consistent process. So you need to define it for yourself. You also need skill files to explain how to analyze and make decisions so that Claude makes the right decisions and does the right analysis on the backtesting results. And you need skill files telling Claude how to document the test results that it gets so that you can come back to them later and check its work. And then finally, you need some documentation. So the output from Claude gives you the different code versions, the analysis results, backtesting lessons, an idea log for further testing, and a journal to capture all of the things that you tested and the results that you found. When you've got this complete setup, it makes AI a super powerful tool. Then obviously then you need to execute your trading systems. I use Interactive Brokers with an automation, not with an AI. If you want an example of automation, go to quantconsole.com. It's the automation that I've been using for years to automate all of my trading decisions. Okay, so let's do some examples. Here's the ideal move hypothesis creation step. So what I did is I found this chart, which is actually the Canadian stock index. And I thought there was some interesting price action here. So I asked Claude to look at this chart. I just pasted the chart into my Claude chat and asked it to look at this chart and describe the market behavior that you see. And here's the results it came back with. I'm not gonna read through all of this, but basically it describes step by step exactly what's happening over the course of this chart. And the behavior I was particularly interested in is this behavior here. So you see that there was some choppy sideways action, sudden correction, some volatility, and then that volatility dissipated and a smooth consistent uptrend emerged. This is the move that I really wanna catch in a trading system idea. So then I've said to Claude, great, I wanna create a trading system that's completely objective and backtestable that enters after a violent pullback is over and the market returns to low volatility rising conditions like we saw in April to August 2025 on this chart. It's not a mean reversion strategy, it's a trend following strategy, and it enters once the panic is over and uptrend that's low volatility has commenced. I want to hold the positions while the uptrend's in place, close it when the market is no longer trending smoothly like we saw in August to October '25. Give me 5 setup, entry, and exit rules that I can test to combine these systems into a complete backtest. Now here's the key: I didn't ask Claude for one system or the best system. I asked it for 5 different alternatives. And this is important because again, Claude doesn't know what the best one is. ChatGPT doesn't know what the best one is. What I've done is give it some context. I gave it a chart, I asked it to explain it and analyze it. Then I asked it to zero in on that particular market behavior and generate some rules that would catch that behavior. And those rules are what we're gonna use to backtest this system. So again, you can see here now Claude's come back with 5 complete rules. Here's the first one. It's an ATR compression breakout. You can pause the video and read this. I'm not gonna go through all of this right now. There's system 2, which is a Bollinger Band width squeeze with a trend filter. System 3 was an ADX trend filter with volatility normalization. System 4 was a linear regression slope with an R-squared filter. And system 5 was a Keltner Channel system. So each of those systems are complete systems that we could backtest. So what I want to do is demonstrate how we now use these ideas with RealTest and with Claude to backtest them a little bit more fully. So I've said, use the RealTest coder skill with System 1, the idea I just came up with, to generate a complete RealTest system script. Ask me whatever additional questions you need to complete this system code. And this is really important because if you don't tell your AI to ask you questions to close the gaps. It's going to make assumptions for itself, and they may or may not be right. So using the framework that I showed you earlier, Claude has now come back to me and asked me questions to fill in the gaps in the system idea that I've given it. So here's what it said: before I code this up, I need to clarify a few things. Basically, which instruments, what position sizing, what data source, what date range, what additional exit rules if any, and what re-entry rules if any. Again, I'm not gonna go through all of this. You can pause the video and read the individual questions if you like. So this has given me an opportunity to reply to Claude and give real trader feedback based on real experience, which is gonna make the system so much better. So I've replied, I've given the feedback. I said, let's trade it with ETFs. I wanna trade it with a certain universe. I use the US sector ETFs in this case, equal position sizing, Norgate data, date range earliest to latest. No time stop. Let's have a trailing stop between 10% and 30% in steps of 5%, and the system should re-enter as soon as the entry conditions are met. That's then allowed Claude to go and code up that system with my complete set of assumptions, which is great. I don't want AI to just blindly assume things based on all of its training data and come back with system code. I want to have a dialogue with it, and that's what's going to make it so much more powerful. So now we've taken this code and we've backtested it on the real market, and you can see that it actually works. The red is the benchmark. So the benchmark is much more volatile. It does have higher returns, but our system has lower returns but a much smoother profile. So you can see here, 2.8% is not exactly exciting, but it is profitable and it might give us something to work with. It's got quite a number of trades, 505 trades. So we've got something to work with here. It's also got a degree of diversification because it's not completely correlated to buy and hold. So something we can work with. So now what I'm going to do is ask Claude to adjust the code so I can run an optimization, and I want to adjust it so it's got an in-sample data period. So I'm going to go from January 2000 to January 2020, leave the rest of the data out, and save this as an in-sample code. At the same time, what it's going to do is implement some parameter ranges to optimize for each of the different parameters. Once it comes back with this code, I can then optimize this according to my usual Expert's Guide to Backtesting testing process, which is what I cover in the Trader Success System. Some of these steps can be automated or enabled by AI, or you can just run it manually. If you don't have that process, then all you can do is ask Claude to set up this optimization code for you, get the code, run your optimization the way you normally would, and then analyze the results. So I've had Claude run the optimization and even select the parameter results for me, and you can see that it's drastically improved. So the in-sample system performance is about the same in terms of return as the benchmark, but much lower drawdowns and volatility. So you can see this is quite a bit more promising. At this point, we wanna brainstorm improvement ideas. And this is something that AI is actually really good at. If you take some images of the best trades and the worst trades and you give them to your AI model and ask it for feedback on rules or ideas that could use to eliminate some of those bad trades or capture more of those good trades, it can give you some additional filters and ideas that you can test and code them up and put them into the system for you. So I've asked it here to analyze these 5 images, which are the worst trades, and come up with some improvement hypotheses. Again, it's done the analysis on those 5 different trades and it's generated improvement hypotheses for each one. Things like a broad market trend filter, requiring the moving trend moving average to be rising, requiring a deeper volatility compression, requiring a minimum recovery from a panic low, or requiring sustained volatility compression. All very sensible ideas that I may not have thought of myself. Some of them I could have, some of them I would have missed. But this point is Claude has done this brainstorming super fast. It took only a couple of seconds, and I could have even fed in far more images and got more ideas to test. So we've got the ideas. Then what we're going to do is ask RealTest to code up these ideas. So I've asked it to code it using the RealTest switch code, which is something that I use to allow me to test multiple hypotheses in a single optimization run. If you're a RealTest user, this is super powerful. What I've done is create a single filter rule here and added that single filter rule to my entry. But that filter rule is a switch. It's a switch that is based on this parameter value here. And depending on what value this has, it's going to use Rule 1, Rule 2, Rule 3, Rule 4, and Rule 5. So this is like an optimization that every step in the optimization, it switches from one filter rule to the next. So you can see up here we've got the 5 filter rules, we've got the switch rule that combines them together, and it only implements one of them at a time depending on what value this filter select rule has in the optimization. So now we can put this into RealTest, optimize on this one parameter value called filter select, and get 5 rows plus the baseline and compare them all. And here's what that looks like. We've got the 6 different rows, and the one that has some potential here is the ATR compression breakout. It improves the rate of return from the baseline. It goes from 3.6 to 4.1. The drawdown drops from 25 to 11. 11. The MAR ratio goes from 0.14 to 0.37. Number of trades drops a little bit. Win rate goes up. So it improves things a lot. The other ones don't really make a meaningful difference, so I'm going to discard them at this point. Now we've got that rule, we've added it to the system. We want to compare in-sample and out-of-sample and see whether our system holds up. Of course, there's a few other steps that you probably want to do in here. Certainly there's more to it behind the scenes that I do, but I don't wanna make this video an hour long. I'm just gonna cut to the chase right now. So this is our in-sample performance. Then we've got our out-of-sample performance. Now out-of-sample, what's interesting is this system is much worse than buy and hold in the S&P 500 index. It does, however, have much lower volatility. It does avoid the drawdowns. And if we look at the performance statistics, it's actually not too bad. There's a little bit of degradation on return, a little bit of degradation on drawdown, a little bit of degradation on MAR, but it's actually holding up. Now, would I trade this system? No, of course not. It's not good enough to trade, but just to demonstrate the process, it's actually pretty helpful. So here's what we just accomplished. We went from chart observation to a complete validated trading system. We started with the market observation, which was a pattern on the chart. We gave it to Claude, it analyzed the market, it generated 5 testable hypotheses. We took one of those, we ran it through the process, we coded it up, We optimized the system to check that the parameters were stable and choose areas of good performance. We then analyzed some of the worst trades, come up with improvement ideas. We tested each of those improvement ideas, found one that worked, and then tested in and out of sample to validate whether or not that passed. The total time to do all of that took minutes, not days. So the process is sound. AI made it fast, and we get a much better edge as a result. That combination of sound process plus AI enablement really gives us an edge as traders. So here's 6 principles that I want you to take away with you to use AI productively and safely in your trading. First, always test before trusting. No exceptions. Second, express every idea as a complete objective set of rules before testing it. Don't let AI guess and fill in the gaps for you. Thirdly, separate in-sample and out-of-sample data to make sure you can validate your rules in your testing correctly. Fourth, you're the trader. Make sure you use AI as your assistant and you drive the process using AI, not the other way around. Fifth, don't share proprietary rules with AI models. And sixth, AI gives you speed, your process gives you the edge. So now we're wrapping up. I want to give you a 5-step action plan so you can take this away and start using this for yourself. Step 1: Identify a market behavior that you want to capture and profit from. Describe it in plain English and find some chart examples that you can use to share with AI. Use AI to generate 3 to 5 testable rule sets based on that description and based on those charts. Then step 3, use AI to code the most promising rule set into your backtesting platform. I'm using RealTest. I suggest you use the same, but if you've got a different one, then you can try that. Backtest on in-sample data only. Run your process, Optimize, fine-tune, and validate on out-of-sample data at the very end. Don't cheat and burn your out-of-sample data early in the process. The best way to start is to start simply. Don't try and incorporate AI and run it, the whole process autonomously on day one. That is possible. I am doing it. Except what I want you to do is just take the first step. Use AI to generate some ideas. Use it to write the code for the system. Use it to edit the code for the system and fill in the blanks manually so that you can piece together that process. And then as you get more proficient, as you understand the technical setup and you come to trust your processes and build better skill files, then you can start to automate more and more and more of the process. And if you want the exact prompts that I use to generate this system so that you can copy them and generate the same sort of results for yourself, all you need to do is scan this QR code or go to enlightenedstocktrading.com www.adrianreid.com/free and get my free Trader Acceleration Bundle. There's an AI prompt toolkit in there for systematic traders with all of the prompts and examples that I used in this presentation to help you get started with AI in your backtesting. There's also a huge number of resources in the Trader Acceleration Bundle that are gonna help you accelerate your trading journey even faster. There's 2 powerful trading eBooks, 5 cheat sheets in total, 3 different courses courses to help you get started with systematic trading, a quiz to help you find your gaps, and a complete mean reversion trading system as an example to help you get started. So go to enlightenedstocktrading.com/free to get the Trader Acceleration Bundle now. And if you're new to systematic trading, I'd suggest what you do next is watch this video because this is the one that's gonna lay the foundation and put you on the right path to consistent, profitable systematic trading. So watch this video next make sure you subscribe to the channel, give this video a thumbs up, and let me know in the comments if you've got any questions. My name is Adrian Reid. This is Enlightened Stock Trading, and I'll see you in the next video. Bye for now.
