Grid and Martingale EAs on Gold: How Long Does an Account Survive? A 22-Year Simulation on Real Prices
- We simulated an EA that "adds trades when price moves against it" on real hourly gold prices from 2004 to 2025, restarting it at the beginning of every month: 261 start dates.
- Martingale version (each added trade 1.5× larger): 52.9% of accounts were wiped out within 6 months and 81.6% within 2 years. Median account life: 5.2 months.
- Fixed-lot grid version lasted longer, but 73.6% were still wiped out within 2 years.
- In the years it survived, the numbers look spectacular ($1,000 became $28,844 in 2024). That is the screenshot used to sell these EAs. But in 14 of 22 years the account went to zero.
Why these EAs look too good to be true
If you have seen a gold EA advertised as "wins 85–95% of trades", with a profit curve that climbs in a near-perfect line and almost never dips, there is a very good chance it runs on one of these two mechanisms:
- Grid (averaging): when price moves against the position, the EA opens more trades in the same direction to pull the average entry price closer. A small bounce is then enough to close the whole basket in profit.
- Martingale: like a grid, but each added trade is larger than the last (for example 1.5× or 2×), so an even smaller bounce closes the basket in profit.
Both share one property: they never accept a loss. A wrong trade is not cut. It is held, and more trades are stacked on top of it. The result is a very high win rate and a smooth Balance curve, because all the losses are hidden inside open trades, which do not appear in Balance until they are closed.
So the right question is not "how much does it make?" but "how long does it last before price simply doesn't come back?" That is what we measured.
How we tested it
We wrote a simulator that behaves like the grid and martingale EAs commonly sold today, and ran it on real hourly XAUUSD prices from 2004 to the end of 2025. Every assumption is listed here so anyone can reproduce it:
| Starting capital | $1,000 · leverage 1:500 · 1 lot of gold = 100 oz |
| First trade | 0.01 lot, in the direction of the previous hourly candle |
| Adding trades | Every time price moves $5 against the most recent trade |
| Size of added trades | Martingale: ×1.5 the previous trade (rounded up to 0.01, as real EAs do) · Grid: fixed 0.01 |
| Take profit | Close the whole basket when price is $3 beyond the basket's average entry |
| Costs | $0.30 spread per trade opened |
| Account blown | When equity falls below 50% of used margin (the stop-out level most brokers use) |
| Start dates | The first trading hour of every month, Jan 2004 – Dec 2025: 261 starts, each run for up to 24 months |
We use many start dates because the outcome of this kind of EA depends heavily on when you switch it on. Start just before a long slide and the account dies fast. Start during a choppy range and it grows beautifully. A single start date is easy to cherry-pick.
Result 1: half the accounts were gone within 6 months
| Wiped out within | Martingale ×1.5 | Fixed-lot grid |
|---|---|---|
| 3 months | 39.8% | 26.4% |
| 6 months | 52.9% | 42.5% |
| 12 months | 66.3% | 56.7% |
| 24 months | 81.6% | 73.6% |
| Median account life | 5.2 months | 8.9 months |
Put simply: if you buy a martingale EA and switch it on at a random date, you are more likely to lose the account within six months than to keep it. Hold it for two years and fewer than one in five survive.
The fixed-lot grid looks a little safer because the trades don't grow, but the end result is similar. When price runs far enough in one direction, the open trades eat all the margin anyway.
Result 2: the surviving years are the advert, the other years are the reality
Now look at it the way EA sellers usually do: start in January and see how much is left in December (martingale ×1.5, a fresh $1,000 each year):
In the years it survives, the numbers are thrilling: $1,000 became $26,401 in 2011, $20,082 in 2023 and $28,844 in 2024. Post a screenshot of 2024 and everyone wants one.
But across 22 years, only 8 years survived to December. In the other 14 years the account went to zero, several in the first weeks: 6 January 2010, 12 January 2012, 5 January 2020. The worst was 2015: just before stop-out, the open position had grown to 47.83 lots on an account that started with $1,000.
Note that someone who started at the beginning of 2024 and made 28×, then kept running into 2025, would face the same market as someone who started in January 2025, an account that was wiped out on 31 January 2025. Profit that has not been withdrawn can disappear in a month.
Why it breaks: the arithmetic of never accepting a loss
Here is how fast the position grows when price moves against it, adding every $5 with a 1.5× multiplier:
| Adverse move | Open trades | Latest trade | Total lots |
|---|---|---|---|
| $0 | 1 | 0.01 | 0.01 |
| $20 | 5 | 0.08 | 0.19 |
| $35 | 8 | 0.27 | 0.76 |
| $45 | 10 | 0.62 | 1.79 |
| $55 | 12 | 1.40 | 4.12 |
After a $45 adverse move, the position is 179 times the first trade, and every further $1 against costs about $179 on a $1,000 account.
And gold keeps getting more volatile. In the same data, the average daily high-to-low range was $13 in 2019, $22 in 2022, $29 in 2024, $53 in 2025 and $114 in 2026 (to September). A $45–55 adverse run that used to be rare can now happen within a single day. An EA tuned when gold moved $13 a day is many times more fragile when gold moves over $50 a day.
That is how an EA can "win almost every trade" and still wipe out the account: a thousand wins of a few dollars each, and one loss that takes everything.
It really does look good on the way up, and that is the trap
In the monthly-start simulation, the median martingale account reached about $3,712 (3.7× its starting capital) at some point before the run ended. That is when the owner is most confident, may add more money or recommend it to friends, and when the performance screenshots are taken.
A smooth equity curve does not mean low risk. Sometimes it means the risk has not been counted yet.
How to tell whether an EA you are about to buy is a grid or martingale
You don't need the source code. The Strategy Tester report or the live account statistics are enough:
- Equity drawdown much larger than balance drawdown, for example 4% balance DD but 27% equity DD. Losing trades are being held open.
- An unusually high win rate (75–95%) but an average loss larger than the average win.
- Several trades open at once in the same direction, evenly spaced, with sizes growing step by step.
- No stop loss, or one so far away it might as well not exist.
- A short test period, or only strong uptrend years, with nothing like 2008, 2013, 2020 or 2022.
- A strongly negative Z-score in the report (−10 or below): trades are not independent, they open and close in baskets.
Each point is explained in detail with a real report in How to Read an EA Backtest Report. If you have a report file, try our free EA checker.
What an EA built to last looks like
No EA wins every trade and nobody can guarantee profit. But EAs designed to survive tend to share these traits:
- Every trade has a stop loss from the moment it opens, so the maximum loss per trade is known in advance.
- No adding to losing trades and no increasing size after a loss.
- Position size is calculated from account equity, not from the urge to win it back.
- The backtest is long enough to include every market regime (20+ years) and is profitable in both the first and second half of the data, not just in a chosen window.
An EA that survives will have losing months and visible drawdowns. That looks less pretty than a grid curve, but it is what fully counted risk looks like. Our full checklist is on What Makes a Good EA.
Limitations of this test
- We used hourly candles, so moves inside the hour are not visible. On each candle we checked the adverse side before the profitable side, which is slightly harsher than reality.
- Real EAs use different settings, so we varied the step ($3, $5, $10, $20) and the lot multiplier (1.0, 1.3, 1.5, 2.0): 16 combinations in total. Every combination lost 70% to 98% of accounts within two years. The worst was a $3 step with a 2.0 multiplier (98.1%); the "best" still lost 70.1%. Settings change how fast it breaks, not whether it breaks.
- Overnight swap and slippage are not included; in real life they make results worse.
- Past performance does not guarantee future results, for this type of EA or any other.
This article is for educational purposes and is not investment advice. Trading gold and CFDs carries a high level of risk and you may lose all of your capital.
