MarinShort answer: We built a live AI trading portfolio by exporting Myfxbook statements from three vendor accounts running Architect Algo, uploading them into Claude along with one-minute OHLC data for every traded symbol, and asking it to assemble a combination whose drawdowns would not peak at the same time. The result runs seven symbols across two strategies on a $1,000 BlackBull Markets account, and after roughly two and a half months it showed an 8.5% gain, a 2% average monthly pace, and a maximum floating loss of 12%. Running the OpenAI side of the robot cost $4 across that entire period, which surprised me more than the returns did.
The word AI turns up on nearly every new expert advisor uploaded to the MQL5 marketplace these days. Most of the time nobody explains what it actually does. So let me show you a real one, what the AI layer contributes, and what it costs to keep running.
What We Tested, and What We Have Not
Being precise about scope matters here, because “we used AI” is exactly the kind of claim that gets stretched.
The account started trading at the end of April. As of July 10 it had been live for close to two and a half months. One account, $1,000, seven symbols, two strategies, one broker.
Things I cannot tell you:
- How this behaves across a full market cycle. Two and a half months is a snapshot.
- Whether the portfolio Claude proposed outperforms one I would have picked by hand. We did not run a control.
- How the AI confirmation layer performs during a genuine volatility event.
- What the OpenAI cost looks like at higher trade frequency or on a larger portfolio.
Point 2 nags at me. The method felt rigorous while we were doing it, and I still think it was better than guessing, but “better than guessing” is a lower bar than “proven.”
Live Results So Far
| Metric | Result |
| Account size | $1,000 |
| Broker | BlackBull Markets |
| Live since | End of April |
| Duration at review | About two and a half months |
| Total gain | 8.5% |
| Average monthly pace | Around 2% |
| Maximum floating loss | 12% |
| Symbols traded | 7 |
| Strategies running | 2 (Vigorous AI and Sharpshooter AI) |
| OpenAI API spend | $4 |
Balance and equity tracked closely together until the end of May. Then they separated: equity started falling while the balance kept climbing.
That divergence is the whole story of this account, so it deserves unpacking.
Reading the Gap Between Balance and Equity
A rising balance with falling equity means the robot is still closing trades profitably while a set of open positions moves against it. Both things are true simultaneously, which confuses people who look at only one line.
More recently the equity has been moving back toward the balance, which is what recovery looks like on this kind of system.
Was I worried? Not especially, and I want to explain why rather than just asserting it. We are running seven symbols on a $1,000 account, which is a lot of concurrent exposure for that size. Floating loss is the expected condition, not the exception. The question I care about is whether the balance keeps rising through it, because that tells me the strategy is still finding and closing winners while the open basket waits.
Here is the honest counterweight, though. “Floating loss is normal” is also the sentence every trader tells themselves right before a basket stops recovering. The reason I accept it here is that the vendor’s own accounts show the same pattern, not because the phrase sounds reassuring.
Our most recent closed trades included a basket of AUD/NZD positions that shut in profit. Open at the time of review: AUD/NZD again, NZD/USD, and EUR/JPY, most of them recently opened, with one running about two weeks. That aging profile is acceptable to me, mainly because other symbols keep contributing while those positions sit.
What the Vendor’s Own Accounts Show
Before committing capital I went through the tracked accounts published on the Responsible Forex Trading site, where Architect Algo lives.
| Vendor account | Size | Notes |
| Vigorous OpenAI Version | $3,000, started October 2025 | Around 35% profit, 7 assets traded |
| Vigorous AI | $500 | Smaller account running the same approach |
| Sharpshooter OpenAI | Not stated | Second strategy in the family |
| Architect Vigorous AI 18 Pairs | $1,000 | Tripled to roughly $3,000, deeper drawdown, 18 symbols |
| Sharpshooter AI Low Risk | Not stated | Conservative variant |
Most of them sat in floating loss at the time I looked. Not one or two: most.
That consistency is what reassured me. When a pattern appears across the developer’s accounts and ours simultaneously, it is a characteristic of the system rather than something going wrong on our machine. The account trading 18 pairs showed the deepest drawdown of the group, which fits the logic exactly: more concurrent symbols, more simultaneous open exposure.
Worth saying plainly, since the vendor says it themselves on their blog: AI does not eliminate losses, and no trading system does. I appreciate a developer who writes that down.
The Problem That Forced This Method
Here is the constraint that shaped everything.
Architect Algo’s AI layer reads market sentiment in real time. Real-time sentiment cannot be reconstructed historically, which means this robot cannot be backtested in any meaningful way. The Strategy Tester has nothing to replay.
For a systematic trader that is genuinely awkward. Backtesting is how we normally decide which symbols and settings belong together. Remove it and you are left with either guesswork or live trading with real money while you figure it out.
So we tried something else: use the live data that already exists, from the developer’s own accounts, and reason over it properly.
How We Built the Portfolio With Claude
This is the part I think is actually worth copying, so I will give the steps as we performed them.
Step 1: Export the raw statements: We pulled balance statements from Myfxbook for three of the AI accounts, two running Vigorous and one running Sharpshooter AI.
Step 2: Define what we wanted: The goal was not maximum return. It was a combination where the symbols and strategies would not sit in floating loss together, or at minimum would not hit their worst drawdowns at the same moment. Correlation of pain, rather than correlation of returns.
Step 3: Upload and prompt: Statements went into Claude with that objective spelled out.
Step 4: Realize the first attempt was insufficient: This is the useful bit. Balance statements alone were not enough, because the model could see when equity dipped but had no market context explaining why. Without price data it could not identify where the peak drawdowns actually occurred or whether two strategies were suffering from the same underlying move.
Step 5: Add the missing layer: We uploaded one-minute OHLC data for every symbol appearing in those statements. That changed the quality of the analysis substantially. Suddenly the model could map drawdown periods onto real price behavior.
If you want that data yourself, the Historical Data app on the Algo Trading Space site under Software provides it. Click through to the raw data and download OHLC for whichever timeframe you need. We built that tool for our own work, and it turned out to be the piece that made this whole exercise function.
Step 6: Read what came back. Claude produced a portfolio dashboard, which honestly exceeded what I expected from the conversation. Top eight combinations, switchable, each with its own character:
| Portfolio variant | Character |
| Lowest Drawdown | Most conservative combination, no gold exposure |
| Smooth Equity | Prioritizes the shape of the curve over raw return |
| Diverse | Widest spread across symbols and strategies |
| Aggressive | Highest return profile, includes gold |
Switching between them changed which robots appeared in the selection. Gold showed up in the Aggressive variant and disappeared from the Lowest Drawdown one, which is roughly what you would expect and served as a decent sanity check that the analysis was not arbitrary.
Step 7: Choose. We ended up with a combination of Vigorous and Sharpshooter running on different symbols. Seven in total, on a thousand dollars.
What the AI Layer Does While Trading
Portfolio construction used Claude. The robot itself uses a separate integration, and the two should not be confused.
To switch on the AI functionality you supply your own API key, from either OpenAI or Claude. The expert advisor then communicates with those servers during operation.
Looking at the Experts tab logs in MetaTrader, here is the actual sequence:
- The EA reads live news from Investing.com
- It compiles a market sentiment document from what it finds
- Before entering a trade or adding to a position, it asks the model for confirmation
- It receives approval or refusal, and acts accordingly
So the AI is not generating signals. It is filtering them. The underlying logic still comes from Vigorous and Sharpshooter, two strategies that regular followers of our channel have known for years. What changed is that a sentiment check now sits between signal and execution.
I find that framing more honest than “AI trading robot,” which implies the model is doing the trading. It is not. It is voting on whether the existing strategy should proceed.
What It Actually Cost
This is the number people ask about first, usually because they assume it will be prohibitive.
Over more than two months of continuous operation, the API key I created for this portfolio had spent $4.
| Item | Cost |
| OpenAI API spend, two-plus months | $4 |
| As a share of the $1,000 account | Under 0.5% |
| Monthly equivalent | Under $2 |
That is less than a coffee per month, and considerably less than the VPS hosting the platform runs on.
Two caveats before anyone budgets around that figure. Cost scales with how often the robot queries the model, so a higher-frequency setup or a larger portfolio would spend more. And API pricing is set by the provider, not by the EA, so it can change without warning.
Still, the assumption that AI integration makes a strategy expensive to run did not survive contact with the invoice.
The Broker Setup
We run this on BlackBull Markets, and we have traded with them for two to three years across various accounts.
What I like about their structure:
- No minimum deposit on the Standard account, which suits testing at small size
- Leverage up to 1:500
- No commission on Standard, with $3 per lot on the Prime raw spread account
- Prime conditions are strong if you are trading frequently enough to justify the commission model
One practical detail that matters more than it sounds: their trading servers sit in London. If you run this or anything like it, pick hosting close to that location. A server in New York talking to a broker in London gives back latency for no reason.
What Could Go Wrong
Seven symbols on a thousand dollars is the risk I would watch, not the AI component.
- Concurrent drawdowns aligning anyway: The entire portfolio method assumes the historical drawdown timing tells you something about future timing. Correlations change, particularly during risk-off events when everything moves together regardless of what the past suggested.
- The account being too small for the exposure: A 12% floating loss on $1,000 is $120. On a heavier allocation the same percentage feels very different, and the temptation to intervene grows.
- API key issues: If the key expires, hits a spending cap, or the provider has an outage, the confirmation layer stops responding. Worth understanding how your EA behaves in that situation before it happens.
- Over-trusting the dashboard: Claude produced an analysis, not a prediction. It organized information we supplied about the past. That is valuable and it is not foresight, and I would rather say so than let the word AI do work it has not earned.
- Vendor account selection bias: We chose three accounts from the developer’s published set. Published accounts are, by nature, the ones a developer chooses to publish.
Would I Do It Again?
Yes, with the method more than with the specific outcome.
Using an LLM to organize live equity data against price history solved a real problem: how to build a sensible portfolio when the underlying system cannot be backtested. That approach transfers to other robots, other strategies, other constraints. The $4 running cost and the 8.5% gain are pleasant, though two and a half months tells me relatively little, and I would treat anyone claiming otherwise with suspicion.
What I would change next time is adding a comparison. Building one portfolio through the AI analysis and one through my own judgment, then running both, would tell us something the current setup cannot.
If you want the underlying robot’s specifications, our Architect Algo page covers those separately.
Frequently Asked Questions
Do I need an API key from both OpenAI and Claude?
No. The expert advisor connects to one provider, and you choose which. We used an OpenAI key for the live trading integration in this account, while Claude handled portfolio construction as a separate exercise entirely outside the platform. Generate the key from the provider’s developer console, add billing details, and paste it into the EA settings. Keep the key private, since anyone holding it can spend against your account balance with that provider.
Can AI trading robots be backtested like normal expert advisors?
Not when the AI component reads live market sentiment, because that information cannot be reconstructed for a historical period. The Strategy Tester has no sentiment data to replay, so any backtest would assess the underlying strategy without its filter layer. Traders assessing this category should weight verified live results and forward testing considerably more heavily than simulated performance, and treat any AI robot advertising long backtests with appropriate skepticism about what was actually tested.
What happens if my OpenAI account runs out of credit mid-trade?
Behavior depends on how the developer coded the fallback. Some systems pause new entries and manage existing positions with their standard logic; others may continue trading without the confirmation step. Neither is inherently wrong, though the difference matters enormously to your risk. Ask the vendor directly before going live, set a billing alert on the API account, and test the scenario deliberately on a demo account rather than finding out during a live position.
Which historical data do I need for this kind of portfolio analysis?
We used one-minute OHLC data covering every symbol appearing in the account statements, which gave the model enough resolution to place drawdown periods against actual price behavior. Lower resolution data may work for slower strategies, though it will blur the timing of short drawdowns. The Historical Data app on the Algo Trading Space site provides downloadable raw OHLC across symbols and timeframes, which is what we pulled from for this particular analysis.
Is a $1,000 account big enough to run seven symbols?
It worked here, with a 12% maximum floating loss, though I would call it the upper limit rather than a recommendation. Each additional symbol adds concurrent open exposure, and the vendor account running 18 pairs showed a noticeably deeper drawdown. Traders uncomfortable watching a double-digit percentage of their balance sit underwater should run fewer symbols or a larger balance. Size for the drawdown you can tolerate rather than the return you want.
How is this different from asking ChatGPT for trade ideas?
Substantially. We did not ask any model to predict prices or suggest entries. The task was organizing data we already possessed: real equity curves from live accounts, mapped against real price history, to find combinations whose worst periods did not overlap. That is an analysis problem rather than a forecasting one, which plays to what language models actually do well. Asking for market predictions is a different activity with a much weaker track record.
The Account Tracker used throughout this article is currently available to VIP Club members, alongside live trading results, early insight, and priority support. We are preparing a standalone subscription for traders who want the tracker without full VIP access. If you would rather have a portfolio built and configured for you, our premium setup covers strategy selection, broker account preparation, VPS hosting, and MetaTrader configuration.
Disclosure: Algo Trading Space may earn a commission from some links on this page. It does not change our figures, our results, or our conclusions.
Risk warning: Trading leveraged products carries a high risk of losing money rapidly. AI integration does not remove risk, prevent losses, or guarantee results. The figures here come from a single $1,000 account traded for roughly two and a half months, which is too short a period to establish reliability. Past performance does not indicate future results. Trade only with capital you can afford to lose. Nothing on this page constitutes investment advice.
