---
title: 'I Gave AI 6 Years of Data. This Is What It Found.'
source: 'https://youtube.com/watch?v=HCiMznnYMiI'
video_id: 'HCiMznnYMiI'
date: 2026-08-24
duration_sec: 874
channel: 'The Moving Average'
---

# I Gave AI 6 Years of Data. This Is What It Found.

> Source: [I Gave AI 6 Years of Data. This Is What It Found.](https://youtube.com/watch?v=HCiMznnYMiI)

## Summary

This video documents an experiment where the creator used AI to analyze six years of EUR/USD market data to discover a trading strategy. The process involved iterative hypothesis testing, where AI generated Python code to sift through vast datasets, leading to a specific strategy based on Tokyo session range breakouts and volume-based mean reversion. The creator emphasizes that this is an experiment, not a finished system, and provides all the files for viewers to test and continue the research themselves.

### Key Points

- **AI as a Research Partner** [00:03] — Instead of asking AI for a direct trading strategy, the creator gave it years of market data and asked it to help search for patterns, using AI to write Python code for analysis.
- **The Data Problem** [01:09] — Markets generate enormous data (candles, volume, time, volatility), making manual testing impossible. AI can analyze these large datasets efficiently.
- **First Idea: Candle DNA** [01:53] — The first hypothesis was that specific candle formations precede large moves. Initial results looked promising, but when turned into strict trading rules, they failed with a 29% win rate and 0.54 profit factor.
- **Iterative Testing and Failure** [03:02] — Many ideas were tested (candle sequences, continuations, stop-losses, sessions) and most failed on unseen data, which increased trust in the process as AI wasn't just confirming ideas.
- **The London Open Strategy** [03:43] — The creator revisited a personal strategy: trading the London open based on the Tokyo session range, using pending orders to let the market choose direction.
- **Initial Strategy Failure** [05:04] — The original London open strategy lost money after trading costs, but analyzing the failure led to a new idea: trading failed breakouts of the Tokyo range with high relative volume.
- **The Winning Conditions** [05:45] — The final strategy: establish a narrow Tokyo range (max ~21.72 pips), wait for a breakout with relative volume >1.6, then trade the mean reversion back into the range, with take-profit at the range's midpoint.
- **Independent Verification** [06:57] — The strategy was converted to Pine Script for TradingView testing. Initial results didn't match Python, but the issue was time zones; after fixing, trade numbers exploded.
- **TradingView Results** [07:36] — TradingView backtest showed 232 trades, 75.86% win rate, and a profit factor of 1.26.
- **MT5 Verification** [08:03] — An MT5 Expert Advisor was built and tested on broker data, yielding 103 trades, 72.82% win rate, and a profit factor of 1.3, confirming the TradingView results.
- **Overfitting Danger** [09:13] — Changing the session timing (3:00-11:00) produced 85.19% win rate with a 2.69 profit factor, but only 27 trades, highlighting the ease of overfitting to past data.
- **AI as a Research Partner** [10:20] — The creator's key insight: use AI to build a research laboratory, not to get direct trade signals. AI helps test hypotheses on massive datasets.
- **Open Source Experiment** [11:53] — The creator provides all files (Python research, TradingView strategy, MT5 EA) for free, encouraging viewers to test, modify, and continue the research.
- **Clear Disclaimer** [12:58] — This is not a finished trading system. It's historical backtesting with potential differences between platforms. Never deploy real capital without extensive forward testing.

### Conclusion

The video concludes that AI is a powerful research partner for trading, but backtested results must be treated with caution. The creator hands over the experiment to the community, encouraging collaborative testing and further development.

## Transcript

market data and basically said, "Figure this out with me." I think I may have found one of the most useful ways that traders can use AI. And no, I didn't ask Chatt where Eurusd is going tomorrow. I didn't ask for the best indicator
settings. And I definitely didn't ask it to give me a 90% win rate strategy. Instead, I gave AI years of actual market data and basically said, "Figure over the next few days is honestly pretty interesting because most of what
we tried failed. But eventually we ended up building an automation that had some results that do have potential. And by the end of this video, I'm going to give you the actual files, the Python research, the trading view strategy, the
MT5 expert advisor. I'm not selling it and I'm not telling you to put real money into this. I'm literally giving you the experiment so you can download it, give it to your own AI, test it yourself, and see if you can continue
where I left off. But first, let me explain how I got here so that you don't redo the work that I've already done. Because I think the process is actually more interesting than the strategy itself. This whole thing started with a
question. Markets generate an absolute ridiculous amount of data. Every candle has an open, high, low, and close. You have volume, you have time of day, you have volatility, previous days, highs and lows, sessions, specific sequences
of different types of candle formations, and millions of combinations of all of these things. So, as a human being, there's no realistic way I'm going to manually test every iteration of those combinations. But AI can write Python
code, which I can't. Python can analyze enormous data sets. So I thought to myself, what if instead of asking AI to give me a strategy, I give it the data and ask it to help me search for one. So I exported EuroUSD historical data and
we started the experiment. The first idea, it called it candle DNA. The basic theory of it was that maybe there is a specific recipe that happens in candle formations before a large move occurs. things like the candle's body size,
things like the candle's body size, range, volume, ATR, and where the candle sits relative to previous highs and lows, sequences of bullish and bearish candles. And initially, some of the results looked really interesting. We
found certain situations where specific characteristics of these candle formations happened two or two and a half times more frequently than other formations, pertaining specifically to what happened before these big moves.
And this is where it would have been really easy just making a YouTube video and saying AI found some secret pattern. Except when we actually took those patterns and turned them into strict trading rules, if and then statements
like coding works, they lost money a lot. One version had around a 29% win rate and a profit factor of 0.54. Garbage. So, we killed it. And this happened over and over and over again.
We tested candle sequences. We tested continuations. We tested different stop- losses, different targets, different holding periods. We looked at previous days highs and lows, session highs and lows, Tokyo session, London session,
specific 15minute time buckets like the orb strategy, relative volume, volatility. Basically, every time we found something interesting to look at, the next question was, does this actually survive when tested on the data
we haven't seen? And most of the time the answer was no. Which weirdly enough made me trust the process even more because the AI wasn't just being a yes man and sitting there telling me that every idea was amazing. Most of my ideas
were getting murdered by the data. And then I remembered a strategy that I made years ago. Trading the London open based on the range of Tokyo session. The original concept was extremely simple. Before London opens, you establish a
range. You put a buy above the range and a sell below the range. And instead of trying to predict which direction the market's going to go, you basically said, "I don't know. Let the market move. Liquidity is coming into the
market because London's about to open. And if price moves, I'll let the market choose the direction for me based on those two pending orders. I've traded these variations before and I've showed it on live streams and I've actually had
decent results. You guys trading along with me have seen that. But it was always a coin flip. I wasn't predicting what way the move was going to go. I was setting up myself regardless of which direction it went. Now, as this strategy
grew, I started thinking to myself, should it be a 30 minute range? Should it be a 60-minute or 90 minute? Where should the stop loss go? Halfway in of the range? Now, should it have specific risk-to-reward ratio like a
1:2? Should the trade enter on close of candle or as it breaks the range? And those were the perfect questions to ask AI to sift out of the historical data. So we built another Python research program. And the original strategy
sucked. You couldn't get consistent results without some biases or skipping days. So after costs of trading, basically the spread and commissions of a trade, it lost money. Without that, it made money. And just like that, another
brilliant idea of mine got crushed by data. But while analyzing why it failed, Instead of trading the initial breakout, what happens if price breaks the Tokyo range with unusually high relative volume, fails to continue, and then
comes back inside the range? Now, we're not predicting a breakout. We're waiting for the market to show us a breakout and then essentially betting that a failed breakout will mean revert back into the range. And that's where things started
to get interesting. Eventually, me and the AI arrived at a very simple set of conditions. You establish the Tokyo range. We only care about relatively narrow Tokyo sessions. The current maximum was around 21.72 pips, the size
of the range. Then London opens. We wait for price to break out of either side of that range, but we don't automatically trade it. We look at the relative volume of that breakout. The current threshold is 1.6 times relative volume. If the
breakout qualifies, then we wait. If price breaks up above that range high, looking for a short, basically trading the fake out. And if price breaks below the range, then comes back into the boundary, we're looking for a long. The
boundary, we're looking for a long. The takerit is then calculated as the middle point of that overnight range. And that's basically it. No 17 indicators. No MACD combined with RSI combined with a 200 moving average combined with the
a 200 moving average combined with the phase of the moon. Range breakout volume failure mean reversion. Now we had another problem. Python telling me something works and that the data makes sense isn't enough. So I asked the AI to
convert the strategy into a Pine script so we could independently test it inside of Trading View. And this just got me deeper down in the rabbit hole. The first Trading View results didn't match Python at all. We were getting far fewer
trades and it looked like the strategy wasn't producing, which usually happens freshly coded. So instead of changing parameters until Trading View magically where the two systems had disagreement. And eventually we found it and it was
the time zones. They change multiple times per year and they're different corrected the trading view strategy to use the local server time based on London open and suddenly the numbers of trades exploded. So now we're testing
approximately the same market session. And this is where we got some results. Here is the trading view screenshot. Across the full deep back test we had Across the full deep back test we had 232 trades, 176 winners, and a 75.86%
win rate and a profit factor of approximately 1.26. Now 75% sounds amazing, but don't get carried away just yet. A back test is a back test and trading view is still trading view. Can we reproduce this somewhere completely
different? So I had AI build a strategy again, this time an actual MT5 expert advisor, the EA that goes on to your MT5 platform. And then I ran it through MetaTrader Strategy Tester using historical broker data. And this is
where I expected everything to fall apart. Instead, we got this 103 trades, apart. Instead, we got this 103 trades, 75 winners, 28 losers, 72.82% 82% win 75 winners, 28 losers, 72.82% 82% win rate, profit factor 1.3. Now, that got
my attention because Trading View and MT5 were not producing identical trades. I mean, they are different feeds and different execution engines with working through some differences in session timing. But look at the actual
statistical results. Trading View about 76% winners and a 1.26 profit factor. 76% winners and a 1.26 profit factor. MT5 about 73% winners and 1.3 profit factor. That doesn't prove that we found some magical trading strategy. But now
we have two independent environments pointing towards something similar. It's like a peer-reviewed medical research study. One test was made and then the other confirmed it. That's also where you guys come in. More confirmation
tipping point where the experiment became way more interesting for me. Then I changed one thing, the session timing. Instead of calculating the range from 2:00 a.m. until 10:00 a.m. broker time, I moved it 1 hour later. 3:00 a.m. until
11:00 a.m. Now, you can see what happened right here on this screenshot. 27 trades, 23 winners, four losers, 85.19% 85.19% win rate. Profit factor 2.69. And this
is exactly where you need to understand how easy it is to yourself with back testing. because I could stop the video right here. Put up on the screen 85% win rate bot generates me millions of dollars. Sell the EA for $5,000
and conveniently forget to mention it's only 27 trades. That's why I keep reiterating this was an experiment. This is the data that I got from the experiment. I'm giving you the data so that you can also continue the
experiment on your side and see the process that I went through. Maybe you'll get some better results. This project changed the way I think about AI and trading. Before I was trying to have the AI write code for me based on my
specific trading rules. And I don't think that's the best use of AI. Like you can't just say, "Hey chat GPT, should I buy EuroUSD?" AI does not have a crystal ball. I think what's much more interesting instead of having it assume
is using AI as a research partner using data sets that are so large it would take you a lifetime to sift through manually. So now I have an idea and AI helps me turn that into a code. Python tests years of data and we look at what
failed. We formulate another hypothesis, test again, then if something survives, translate it into a trading view pine script, test it again, then translate it into MT5, test it again, and eventually forward test it on a demo account, of
different relationship between AI and trading. AI isn't giving me trades. It's helping me build a laboratory with the most qualified research team ever. And this experiment for now because I could spend another week optimizing this.
Maybe get the win rate from like 75% to 77%. Maybe we find another filter. Maybe changing the range gives us another few%. But there's also a huge danger here. The more parameters I optimize against, the easier it becomes to
accidentally build something that's absolutely incredible only on past data and then completely useless on trading future price action. So instead, I'm going to do something different. I'm going to give you this project. I've put
we've been working on, the Trading View strategy, and the MT5 EA. Download them. Look through them. Run your EA through your own strategy tester on your own
broker on your own terminal. Change brokers. Change pairs. Try GBPUSD. Try USD JPY. Try New York instead of London session. Give the Python research to session. Give the Python research to ChatGpt Claw, Gemini, whatever you use.
Tell it the information that we've discovered already and then ask what should we test next? Because imagine if hundreds or thousands of us start independently testing the same idea. Maybe everyone comes back and says,
"Arty, it's garbage. Perfect. We learn something that this doesn't work." Or maybe people start finding the same behavior on different brokers, different years, and different markets. Maybe this is good for oil. I don't know. But it's
investigating. But I need to make one thing extremely clear. I've already said it multiple times in this video. This is not a finished trading system. I'm not telling you that this is profitable. I'm not telling you to put this on a live
account. And I'm definitely not telling you that we created an 85% win rate money printer. These are historical back tests with historical data. There are differences between Python results, Trading View results, and MT5 results
that need to be investigated. And the highest win rate MT5 variation only produced 27 trades across a massive data set. So treat this exactly how I'm treating it, as research. This was a super fun experiment for me because it
gave me a new perspective on how I can work together with AI. I'm nowhere near being done playing with this. And the more data I have across every different asset and I build this supercomputer of information, the higher likelihood I
information, the higher likelihood I have of finding out exactly what these huge financial institutions, what brokers and hedge funds are all doing because every major player is using algorithmic trading. Now we have the
tools at our disposal and this is an opportunity for you that you should get on now. So, all you got to do, go to my website, TMAF FX. You're going to find a link somewhere on the home screen. Just click that link. It'll give you a whole
dropbox of the files. Have fun playing, but don't get overly excited about results. Always forward test on demo accounts and never deploy real capital
until it's absolutely proven over a year of forward testing on a demo account that it produces decent results. This is your money. This is your risk. This is your money. This is your risk. This is me handing the baton to you.