Every week someone posts a gold EA with a win rate above 95% and an equity curve that goes up like a staircase. I wanted to know whether that is a strategy or an accounting trick, so over the last month I built three grid systems for XAUUSD and tested each one until it told me the truth. All three failed. The way they failed is the interesting part.
System one: a grid with a real stop loss
The first was a volatility-sized recovery grid, with one rule the popular ones do not have: a hard equity stop on the whole sequence, fixed the moment the sequence opens.
The loss cap worked. Across roughly 1,000 sequences, two configurations and two out-of-sample windows, the cap was never breached once — 21,846 assertions, zero violations. I was pleased with myself for about an hour.
Then I looked at the profit. Over four years the system lost 2,456 USD, and trading costs accounted for only 128 USD of that. The strategy itself lost the rest.
That is the finding, and it is not about my implementation. A grid makes money on paper because it converts a small probability of a very large loss into a high probability of a small gain. The expectancy is roughly flat; the distribution is what changes. The moment you cap the large loss, you have removed the only thing that was paying for all those small gains. A bounded grid and a paper-profitable grid are the same system with the tail cut off, and the tail was the business.
System two: the same idea, tuned properly
I assumed system one was just badly tuned, so I did the search properly: 3,840 parameter combinations, 43.8% profitable in sample, selection by plateau rather than by peak, and then two untouched out-of-sample windows. 44 of 47 candidates came out positive in both — against 25% expected by chance. There is real structure in gold breakout geometry, and that result says so.
Then I ran the one test that people skip, because it is slow and it is not flattering: the continuous path from 2013 to 2026, one account, no restarts.
- 844 trades
- 98.5% win rate
- Final result: −23.2%
Eight hundred and forty-three winners and one loser. The loser was a single sequence on 16 April 2025 that gave back 4,425 USD. The second worst loss in thirteen years was 132 USD.
I swept equity stops from 5% to 35% looking for a level that would have saved it. Every one of them lost money. There is no stop that fixes this, because the stop is not the problem — the shape of the payoff is.
That 98.5% is not a sign of quality. It is the signature of the problem.
System three: measuring someone else's grid
For the third I stopped building grids and built a tool that watches one: it recovers the grid geometry from whatever positions are already open and projects what the next few layers cost. Pointing it at a realistic gold grid gave the cleanest statement of the whole exercise:
92.3% of trades won. The account finished down 86.2%.
Both numbers describe the same trade history.
What I did instead
I went back to the boring thing: channel breakout with a hard stop loss on every single trade, four strategies on different channel lengths, no averaging, no recovery lots, one position per strategy at most.
The result is much less impressive to look at, and that is the point:
- Win rate 50.7% — almost exactly a coin flip
- Profit factor 1.75
- 2013–2026: +298%, CAGR 10.6%, max drawdown 9.3%
- 13 of 14 years positive; 2022 was −5.2%
- Largest single loss: a normal stop loss, like every other loss
Half the trades lose, and the losses are all roughly the same size. That is what a payoff looks like when nothing is hiding in the tail. It will never produce a 98% win rate, and a system that does should be asked where the losses went.
The test I wish more buyers would run
If you are evaluating any EA, including mine, the single most useful thing you can do is run the continuous path over the longest history your broker has, on one account, and look at the largest single loss rather than the final number. A system whose worst loss is 30 times its second worst is not a system with good risk control that had one bad day. It is a system whose entire result depends on a day that has not happened yet.
The second most useful thing: test long enough to matter. I measured every rolling window of my own backtest. A one-month test lands profitable 52% of the time, three months 68%, twelve months 88%. If you test one month and see red, you have sampled a coin flip, not evaluated an EA.
The breakout EA is on the Market as Gold Channel Rider (https://www.mql5.com/en/market/product/195647). The grid systems are not published, and after the numbers above I hope it is clear why.
