Automated trading · Research
中文读者：黄金EA是什么？完整避坑指南 · 繁體中文讀者：黃金EA是什麼？完整避坑指南
What Is a Gold EA? How Automated Gold Trading Really Works
By Ken Chigbo · 18 years in financial markets · Last verified 8 August 2026
A gold EA is an automated program that buys and sells gold (XAU/USD) on MetaTrader according to rules written in advance. It executes those rules. It does not understand the market. Many products marketed as “AI gold EAs” are fixed-rule automated systems; the AI label alone does not establish that machine learning is being used.
What decides whether one is worth running is rarely the backtest curve. It is maximum drawdown, sensitivity to spread and slippage, and how much live trading history exists. This page explains what a gold EA actually is in technical terms, why backtests mislead, and the twelve things to check before risking money.
The short answer
| What it is | A program written in MQL4 or MQL5, attached to a chart, that places orders automatically |
| Where it runs | MetaTrader 4 or MetaTrader 5, on your machine or a VPS. It stops when the terminal stops |
| What decides results | Maximum drawdown, spread and slippage sensitivity, live sample size — not backtest return |
| Is it AI? | Not necessarily. Many EAs are fixed-rule automation; genuine machine learning should be supported by evidence about training data, features, validation and retraining |
| Biggest hidden risk | Martingale and grid position-sizing: long stretches of small gains, then one loss that ends the account |
1. What a gold EA actually is
“EA” stands for Expert Advisor — MetaQuotes’ term for an automated trading program on the MetaTrader platform. It is written in MQL4 (for MT4) or MQL5 (for MT5), compiled, and attached to a chart. From then on it reads incoming price data and places, modifies or closes orders according to its code.
Three consequences follow, and each one matters more than most marketing admits:
• It only runs while the terminal runs. Close MetaTrader, lose your connection, or let the machine sleep, and the EA stops managing open positions. This is why VPS hosting is standard rather than optional.
• It executes rules, it does not judge. An EA has no view on whether today’s US inflation print changes the picture. If the conditions in its code are met, it trades.
• It is bound to your broker’s prices. The same EA on two brokers produces different results, because spread, execution speed and slippage differ. An EA is never a standalone product; it is a program plus an execution environment.
2. Automation is not artificial intelligence
This is the single most abused word in the retail gold EA market, so it is worth being precise.
| Rule-based automation | Genuine machine learning | |
|---|---|---|
| Decision logic | Written by a human, fixed at compile time | Derived from training data by an algorithm |
| Changes over time? | No — identical input gives identical output, forever | Yes, if and when the model is retrained |
| What the seller can show | The rules, or at least their logic | Training data, features, retraining frequency, validation method |
| Honest description | “Automated” / “systematic” / “rule-based” | “Machine learning”, with the above disclosed |
The test is simple. Ask what the model was trained on, which features it uses, and how often it is retrained. A system genuinely using machine learning can answer. If the answer is vague, treat it as a fixed-rule EA and judge it on drawdown and live sample size like any other.
There is a legitimate middle case, and it should be described precisely: AI tools can assist a human in developing and tuning a rule set, which is then fixed and executed unchanged. That is a statement about how the system was built, not about what it does while trading. It is not the same claim as “the AI trades for you”, and the two should never be blurred.
3. Why backtests mislead
A backtest replays historical prices through the EA’s logic. It is genuinely useful for finding obvious flaws, and genuinely dangerous as a sales document. Four reasons, in order of how often they cause real damage.
Modelling quality. MetaTrader’s Strategy Tester reports how faithfully it reconstructed price movement inside each bar. MT5 supports every-tick testing based on real ticks, which is the highest fidelity available; results derived by interpolating from one-minute open-high-low-close bars are far coarser. For a strategy holding positions for minutes, that difference alone can invert the result. MetaQuotes documents the available modes and their limitations in the MT5 Strategy Tester documentation.
Spread assumptions. Many backtests run on a fixed spread. Live gold spreads widen — often several times over — around the US open, high-impact data, and the daily rollover. A scalping strategy tested at a constant spread and traded through a data release is not the same strategy.
No slippage or latency. The tester fills at the modelled price. Live, your order travels to the broker, and the fill can be worse — particularly in the fast conditions a gold strategy tends to trade in. Backtests systematically flatter strategies that trade often.
Curve fitting. Optimisation searches parameter combinations for the best historical result. Search hard enough and you will always find settings that performed beautifully on that specific history and carry no predictive value. MetaQuotes’ own optimisation documentation describes the mechanism; forward testing on data the optimisation never saw is the standard defence.
4. The statistics that matter, and what they hide
Maximum drawdown is the largest peak-to-trough fall in account equity. It is the number that decides whether you can actually hold the position through a bad run. A system with a 45% historical drawdown will, at some point, ask you to watch nearly half the account disappear and keep going. Most people cannot, which is why they abandon systems at the worst moment.
Profit factor is gross profit divided by gross loss. Above 1.0 is profitable in the tested period. It says nothing about how the profit arrived — a profit factor of 1.6 built from thousands of small wins and one catastrophic loss is a different animal from 1.6 built evenly.
Win rate is the most misleading figure in the market. A 95% win rate is trivial to manufacture: take small profits, never cut losses, and let losers run. The number that matters alongside it is average win versus average loss. High win rate plus a much larger average loss is the signature of a system that has not yet met its bad day.
Sample size. Thirty trades tell you almost nothing. A gold strategy needs to have traded through several distinct regimes — a Fed hiking cycle and a cutting cycle, a risk-off spike, a quiet range — before its statistics mean much. Ask how many live trades, over what period, not how many appear in the backtest.
5. Martingale and grid — the risk that hides in plain sight
Some retail gold EAs use one of two position-sizing structures worth checking carefully: Martingale or grid.
Martingale increases position size after a loss, so that one eventual win recovers the sequence. Grid opens additional positions at intervals as price moves against the original entry. Both produce long, smooth, extremely attractive equity curves — because both convert frequent small profits into a single concentrated risk that has not happened yet.
Gold can be especially unforgiving to both when price trends strongly without meaningful retracement — exactly the condition that ends a martingale sequence. The tell is visible in the statistics: a very high win rate, a shallow reported drawdown, an equity curve with almost no volatility, and an average loss far larger than the average win. If a seller will not disclose the position-sizing logic, that omission is itself the answer.
6. An EA is not a copy-trading strategy
These are routinely confused, and the difference determines who controls what.
| Running an EA | Copy trading | |
|---|---|---|
| Where the logic sits | On your terminal | On someone else’s account; you receive the resulting trades |
| Needs to stay online | Yes — hence the VPS | No — the broker’s infrastructure mirrors trades |
| Can you see the rules? | Sometimes, if the source is provided | Usually not — you see results, not logic |
| Who holds the money | You | You — in your own account at the broker |
| Main risk added | Technical failure, wrong settings | Dependence on another party’s ongoing decisions |
Both are exposed to the same underlying market risk. Copy trading removes the hosting problem and adds a different one: you are relying on a third party’s continued discipline, and you generally cannot inspect what they are doing until after they have done it.
7. The twelve-point due-diligence framework
This is the checklist I apply before taking any automated gold system seriously. A system that cannot answer points 1 to 4 does not reach point 5.
1. Live track record. How long, how many trades, on a real funded account — not a demo, not a backtest.
2. Maximum drawdown, stated plainly. Peak-to-trough, as a percentage, on that live record.
3. Position-sizing logic. Fixed, percentage-based, martingale or grid. If it is not disclosed, assume the worst.
4. Average win versus average loss, alongside win rate. Never the win rate on its own.
5. Regime coverage. Has it traded through a hiking cycle, a cutting cycle, and a risk-off spike?
6. Backtest modelling quality, and whether the tester used real ticks or interpolated bars.
7. Spread assumption in testing, against the spread you will actually pay.
8. Slippage and latency — whether they were modelled at all.
9. Broker dependency. Which broker produced the record, and how different your execution is likely to be.
10. Behaviour around high-impact data. Does it trade through releases, or stand aside?
11. Independent verification. A third-party-tracked record beats a screenshot, always.
12. What the seller will not say. The gap in the disclosure is usually the risk.
8. What you need to run one
MetaTrader 4 or 5; a broker account that permits automated trading on gold; and in practice a VPS so the terminal stays online continuously. Beyond that, confirm the broker allows your intended holding period — some restrict very short-term trading — and check the gold contract specification, because lot size, tick value and swap terms vary between brokers and change what any given percentage return means in money.
Sources
Technical claims about MetaTrader testing and optimisation are drawn from MetaQuotes’ own documentation. Regulatory context is drawn from FCA primary publications.
• MetaQuotes — MT5 Strategy Tester: testing modes and modelling
• MetaQuotes — MT5 strategy optimisation
• MetaQuotes — MT4 Strategy Tester
• MQL5 documentation — program testing at runtime
• FCA PS19/18 — permanent measures for retail contracts for difference
• FCA FG24/1 — finalised guidance on financial promotions on social media
• LBMA — precious metal price data
Related reading
• Trade gold the macro way — the framework this sits inside
• Choosing a broker for automated trading — execution, VPS and platform considerations
• How copy trading works in practice
Applying this to a live strategy
If you want to see the framework above applied to a real gold system rather than a hypothetical one: KenMacro Gold Scalper publishes its reference-account track record, observed maximum drawdown, profit factor and trade count, with the reference period stated — the same figures this page argues you should demand of anything automated. Availability of any broker connection depends on your jurisdiction and is determined by the provider.
Risk disclosure
This page is educational research, not investment advice, and not a recommendation to buy, sell or use any particular product or service. Trading leveraged products such as CFDs on gold carries a high risk of losing money rapidly. Automation does not reduce market risk — it removes hesitation, which can increase the speed at which losses accumulate.
Past performance is not a reliable indicator of future results. Availability of any product mentioned varies by jurisdiction and is determined by the provider, not by this website. Consider taking independent advice if you are unsure.