Start with the caveat, because it outranks every number below. Gold Atlas is a non deterministic expert. It uses a vendor randomizer, which means two runs on identical tick data produce different trades. Our catalog carries that as a standing note for this EA, and it changes what a backtest is. Everything that follows describes one draw. Run it again and the profit, the drawdown and the trade count all move. That is not a criticism of the audit method, it is a property of the product, and it is the single most important thing a buyer should know before reading any performance claim about it.
With that said, the draw we measured is informative. Over 23 years of gold ticks the EA took 10,999 trades and finished at +$8,239.71 on a 100,000 USD measurement frame with fixed 0.01 lots. The deepest reconstructed drawdown was 0.46% closed and 0.49% floating. No account death at any size we tested.
The exit mix has exactly one entry in it. All 10,999 trades closed on a stop loss. Take profit exits 0. Before you read that as a disaster, look at the second line. 4,495 of those stop exits closed in profit, which is 40.9%. Our engine reads this as a trailing stop signature, and the inputs agree that the exit logic is a trail rather than a fixed target.
This shapes the whole statistics profile. The win rate sits in the low forties across every sub strategy, from 37.7% to 48.4%, and the strategy still makes money, which is the normal arithmetic of a trend follower that cuts losers early and lets the trail run. It also means the reported win rate tells you almost nothing here. What matters is the distribution of the winners, and that distribution is the audit's red flag.
The EA traded on 4,636 days. Eighty percent of the profit was made on 54 of them, which is 1.2% of the trading days. Our concentration dimension rates that red, and the detail line is blunt about why. Miss a handful of days and the edge is gone.
The three best days in the run are 2026-04-28 at +$276.36, 2026-03-18 at +$272.25 and 2026-08-05 at +$226.69. The three worst days are also all in 2026. That clustering has a mechanical cause worth stating, because it is easy to misread as the strategy suddenly working. At a fixed 0.01 lots, one dollar of gold movement is worth the same in 2003 and in 2026, but gold itself went from 346.51 to 4,385.68 across our window, a move of 1,165.67%. The same percentage swing is simply worth more than twelve times as many dollars at the end of the window as at the start. Any fixed lot backtest over this span will look like it discovered its edge recently.
Gold Atlas ships with five entry models and the vendor default enables all of them, with a comment in the inputs recommending exactly that for diversification. Our engine splits the deals by the EA's own entry comments, and the five sub strategies sort themselves into a clean ladder.
Model 1 is the busiest with 4,946 trades and made +$619.16, which is about 13 cents per trade. Model 5 is the quietest with 859 trades and made +$2,279.99, which is about $2.65 per trade. In between the pattern is monotone. Fewer trades, longer median holding time, higher win rate, more profit. Model 1 holds a median 1.5 hours, model 5 holds a median 13.2 hours.
The five model results add up to $8,239.71, exactly the net result of the run, so this is a decomposition and not an estimate. What it suggests is that a buyer who switches models off is making a much larger decision than the input names imply, and that the default recommendation to run all five is carrying 4,946 trades worth of costs for 7.5% of the profit. We are not recommending a configuration. We are saying that the diversification argument in the vendor's own input comment is not visible in the deal data of this draw.
The break even shock, the extra cost per trade that takes the run to zero, is $0.75. At an extra $0.50 per trade the run still nets +$2,740.21. At $1.00 per trade it becomes -$2,759.29. The measured average position is 0.01 lots, so the whole result lives inside a spread assumption.
Actual measured costs were fine. Commission across the run was -$879.92 at our retrofit of $3.50 per lot per side, swap was -$708.23 across 2,188 overnight trades, and costs took 16.2% of gross profit, which our engine rates ok. The sensitivity is not about the costs we could measure. It is about the ones a tester cannot show.
The journal records 119 failed entry attempts and 586 failed modify attempts. The failed entries earn a caution in our data quality dimension, because entries that never filled in the simulation are entries whose live behaviour is unknown. Against 10,999 trades that is a small share, and it is a real one.
The second note is methodological and speaks for the audit rather than against the EA. Because the EA holds up to 5 positions at once, closes do not arrive in the order the positions were opened. In 4,692 of the 10,999 trades, roughly 43%, the position that closed was not the oldest one open. A reconstruction that pairs deals first in first out, which is what you must do without a tester journal, would have mispaired all of those. Our pairing came from the journal with a coverage of 1.0 and a confidence of exact, and the reconstructed final balance matches the report to the cent at $108,239.71. Any per trade statistic about a multi position EA that was computed without the journal should be treated as an estimate.
Gold Atlas was published on 2026-01-06 and last updated on 2026-02-14. Of the tested history, 8,278 days and 10,744 trades lie before the release and carry +$7,801.78 of the result. The window after the last vendor update, the only slice that is out of sample by construction, is 179 days long, contains 215 trades and nets +$146.95 at a profit factor of 1.07. Positive, thin, and far too short to conclude anything from. It is also, on a random drawing EA, one draw of one short window.
On prop rules the picture is the familiar split. Our sweep against IQ Capital Classic records 0 account deaths at every size from full down to one eighth and 0 breaches of the per position loss limit. The floor is never the problem. The consistency rule is, and it is breached in 10 years, which is what a profile that earns on 1.2% of its days does to a rule that caps the best day at 30% of annual profit. The modelled withdrawal rate at full size is 0.33% per year. Our prop fit verdict is red for the consistency reason, not for a survival reason.
The challenge model adds the number a funded seat actually turns on. Over 1,000 resampled paths the probability of failing on the loss floor is 0.0 and on a daily loss limit 0.0, while phase one is passed on 0.119 of paths and phase two on 0.989, which combines to 0.118. On 0.881 of paths phase one is still undecided when the data ends, because a path runs 6,068 days and the median passing path needs 5,712 of them. That prices the attempt. The audit expects 8.4 tries at phase one, 19 tries for a 90% chance of clearing it once, and puts the chance of five failures in a row at 0.531. Every sizing step below full size passes phase one with probability 0.0, because the target stays fixed while the trading that has to reach it shrinks.