Gold EA Backtest vs Live: What the Tester Hides

You download a gold EA, open the MetaTrader Strategy Tester, select XAUUSD, and press start. Two minutes later you are looking at an equity curve that climbs from left to right, a profit factor above 2, and a drawdown you would happily accept.

So you attach it to a live account. Six weeks later it has given back most of those gains, and it is holding a position larger than anything the report showed.

Nothing broke. The backtest was a hypothesis. The demo account is the experiment — and you ran the experiment second, with real money.

Here is what makes gold different from the majors. XAUUSD carries a wider spread than EURUSD, USDJPY or GBPUSD, and that spread widens violently around news. Its tick size makes every point of slippage expensive. And it trends hard enough that a strategy catching two or three large moves can look brilliant on a chart of equity. Those three features are precisely the ones a Strategy Tester approximates instead of measures. The tester is not lying to you. It is answering a narrower question than the one you thought you asked.

The gap is not a technical footnote either. It is the difference between a tool that deserves capital and a tool that deserves a spreadsheet and six more weeks of patience.

How does the gap usually show up in practice?

It shows up as a strategy that produced a beautiful report and an ordinary live account, with no single dramatic event to blame. The losses are spread thinly across execution costs you never saw in the tester’s numbers.

Here is the pattern, written as a scenario because it repeats constantly.

Daniel is an illustrative composite, not a real customer, built to show a pattern that repeats often. He is a trader who spends a Saturday vetting a free gold robot he found in a forum thread. He reads the description, likes the logic, and runs the tester on XAUUSD on the M15 chart from January 2022 to December 2025. Modelling quality shows 90%. The report gives him a profit factor of 2.1, a maximum drawdown of 8%, and 640 trades. The curve is not a straight line, and that reassures him more than the totals do: it dips, it recovers, it looks like a real system rather than a fitted one. He scrolls the trade list, sees losses mixed with wins, and concludes the test is honest.

On Monday he opens a live account and trades the smallest lots he can. Week one is fine. Week three produces a loss roughly twice the size of the tester’s average loss. By week six he is down about 18%, and the EA is still holding a position through a news release. He closes it and tells himself the market changed.

The market did not change. The bill did.

The tester traded XAUUSD with a fixed spread taken from a quiet hour. His broker charged a floating spread that widened on the release. His stop was slipped. His broker charges commission per lot which never appears in the report. Swap, if his account pays it, was missing too. Individually these are rounding errors. Multiplied by 640 trades, they are his entire edge — and Daniel never saw a single line item that said so.

The real cost here is not one failed EA. It is the habit of accepting a detailed-looking report as evidence. The Strategy Tester is built to look authoritative: a modelling quality percentage, a trade-by-trade list, a drawdown chart, a Sharpe-like shape. It reads like proof. It is a simulation running on assumptions you did not choose, some of which you cannot see at all.

So flip the order of operations. The backtest is a hypothesis. The demo account is the experiment. The rest of this page is about running that experiment properly on gold: what the tester hides, how to read a walk-forward result, and what to check before a single real order is placed.

Why does a gold EA backtest so rarely survive live trading?

Because the tester resolves spread and timing in the EA’s favour, and on gold those two costs are large enough to consume the whole edge. A backtest tells you how a set of rules behaved under one set of assumptions. A live account tells you how the same rules behaved under your broker’s conditions. Those are different questions with different answers.

Start with the arithmetic. Expectancy is average win times win rate, minus average loss times loss rate, minus trading costs. Every strategy is some version of that sum. What separates a backtest from a live account is the last term, and on XAUUSD the last term is big.

A gold scalper targeting 60 points with a 30-point stop is working on a thin margin. Note that I am using “points” to mean the last decimal place in your broker’s quote, because gold pricing conventions differ between brokers. If the tester charges 15 points of spread and your live account charges 35, you have removed a third of the win on every trade. Apply that to 400 trades and a profit factor of 1.6 slides below 1.0. The strategy did not change. The cost per decision did.

There is a second reason, subtler and more dangerous. Gold profits do not arrive evenly. In most gold backtests a large share of the total comes from a small number of trades that caught a real directional move. Those are exactly the trades most sensitive to execution. They happen when volatility is highest, spread is widest, and slippage is most likely. Model those trades generously and your headline numbers flatter the whole system, because the system’s entire character is defined by them.

That is why a gold EA that fails live rarely fails slowly. It fails in the weeks when the market finally moves, which is the only time it was ever going to make money.

What exactly does the Strategy Tester hide about your gold data?

It hides the intrabar path. Between the open and the close of a single bar, the tester can only guess the order in which prices were touched, and that guess decides whether your stop or your target was hit first.

In MetaTrader 4, “Every tick” mode does not use real ticks. It builds ticks inside each M1 bar from that bar’s open, high, low and close. M1 history itself is often assembled by the terminal when the broker’s server does not hold it, so on older periods you can end up with a path generated from an H4 bar, presented as tick-level detail. MetaTrader 5 is better when your broker supplies genuine tick history, but tick archives are short at most brokers and thin on gold outside the active sessions.

The modelling quality figure tells you how much of the test was real. Ninety percent means roughly a tenth of the ticks were invented. Reaching the high nineties requires real tick data, and even then the tester has to decide the sequence of prices within the finest interval it holds.

That matters more on gold than on almost anything else. A news release can put 200 points of range into a single M15 bar, high and low far apart, stop and target both inside it. The tester picks a path. Your account, on the day, gets the one the market actually took.

Two habits cut this risk down.

• Test on “Open prices only” first. If the strategy collapses in that mode, it depends on intrabar detail — the least trustworthy data in the whole simulation. A ruleset that survives open-prices-only testing is robust to the parts of the model nobody can verify.
• Read the tester’s journal, not just its report. The log lines tell you when M1 data was missing and history was generated. Most traders never open that tab. It is where the simulation admits what it did.

If you want the mechanics of the terminal in more depth, our guide to online forex tester backtesting walks through the settings and what each one does to the result.

How does spread treatment change a gold backtest result?

It changes it more than any input parameter, because spread is charged on every trade while parameters only shape which trades you take. A tester run at a fixed 15-point gold spread is not a conservative estimate. It is an optimistic one.

There are four separate costs, and only some of them appear in a default test.

Cost	Where it hides	How to handle it
Floating spread	Replaced by a fixed number in the tester	Set the tester spread to your broker’s typical gold spread for the session you trade, then re-run
Commission	Absent in MT4’s report; a field in MT5’s tester settings	Enter your broker’s per-lot commission so it is deducted from every trade
Slippage	The tester fills at the price it asked for	Assume one to three points on entries and worse on stop exits, then subtract it manually
Swap and financing	Applied only if the symbol specification is correct	Confirm the swap values in your broker’s specification, or avoid holding overnight during the test window

Two further traps sit underneath that table.

The first is Bid and Ask. In MQL, Close[] is a Bid price. A buy executes at the Ask, which is Bid plus spread. An expert advisor that measures its entry distance against Close[] and forgets the spread is crediting itself a free entry on every buy. When you read source code, that is the first thing to look for.

The second is the execution model. The tester can fill the instant your condition is true. Live, your order travels to a server, and during a spike the price you get is not the price you asked for. Gold’s stop levels compound this: brokers enforce a minimum distance, and stops placed inside a spike get filled wherever the next available price is.

The practical test is blunt. Take the report you trust, widen the spread to your broker’s worst realistic gold spread during the session you trade, add commission, and run it again. If the edge survives, you are looking at something worth the next step. If it vanishes, the edge was a cost assumption.

Where does look-ahead bias come from in an indicator?

It comes from acting on information that was not available at the moment the decision was supposed to happen. Look-ahead bias is not exotic. It is usually one line of code reading the wrong bar.

Four versions show up again and again in gold tools.

• Reading the forming bar. An indicator that uses index 0 is using a bar that has not closed. In the test, the value at that moment is a fact from the data. In live trading, that value is still moving.
• Higher-timeframe candles that have not closed. A multi-timeframe filter pulling the current H1 candle onto an M5 chart repaints. The tester can see the H1 candle’s final form; your terminal cannot.
• Trading at the signal bar’s close. If the EA decides on the close of a bar and the tester fills it at that same close, the EA has used information from the end of the bar to trade at the end of the bar. A real order arrives a moment later, after the price has moved.
• Indicators that need the future. ZigZag and fractal-style tools confirm a pivot only after later bars arrive. Attached to a live chart they repaint; inside a tester they look prescient.

There is a fifth category that is less a bias and more a habit: strategies that quietly assume the tester’s fill. Nothing in the code looks wrong. The assumption lives in the difference between “condition true” and “order filled”.

This is the strongest practical argument for using tools whose source you can read. EA31337 Libre is a good example of an open-source MT4 and MT5 expert advisor whose code you can inspect line by line — and the same principle applies to any tool on this site. Search the source for how signals read bar indices, whether the Ask is used on buys, and whether any calculation touches index 0. Ten minutes of reading beats a hundred optimisations.

Which backtest artefacts should you audit before you believe a report?

Audit the artefacts that describe assumptions, not the ones that describe profit. Profit factor and total net gain tell you what the simulation produced. Modelling quality, spread, sample size and out-of-sample share tell you whether that production means anything.

Backtest artefact	What it hides	What to do instead
Modelling quality around 90%	A meaningful share of ticks were generated, so intrabar paths were invented	Demand real tick data, or accept that stop-versus-target order is a guess
“Every tick” mode on MetaTrader 4	Ticks reconstructed from M1 bars, sometimes from higher timeframes	Re-run on open prices only and compare the two results
A single fixed spread input	News-time spread widening, session-by-session variation	Test at your broker’s worst realistic gold spread
No commission line	Per-lot costs that hit every trade	Add commission to the test, or subtract it from the results yourself
Profit factor from under 200 trades	Randomness dressed up as edge — a handful of wins carry the total	Require at least 300 trades across the whole test
Optimisation on the full period	Parameters fitted to the same data they were judged on	Split in-sample and out-of-sample before you optimise anything
Equity curve dominated by one large win	A system whose character depends on one news event	Remove the biggest winner and check whether the edge survives
Maximum drawdown on closed equity only	Floating losses the EA was carrying at the worst moment	Look at the balance-plus-floating drawdown, or the lowest equity point
Backtest ending this year	Selection: a strategy tested only on the market that just happened	Repeat the test on an older window with a different gold regime

Two habits make this list usable. First, always ask what the report does not contain — commission, spread assumptions, sample size. Second, always re-run the decisive test yourself rather than accepting a screenshot. A backtest you produced is a backtest you can interrogate.

How do you run a walk-forward test that actually means something?

By deciding before you start that only the out-of-sample windows count. Walk-forward testing is not a different indicator or a magic option. It is a rule about evidence: parameters may be fitted on one stretch of history, and the strategy may only be judged on history it has never seen.

Here is a protocol you can run in a weekend on gold.

1. Cap the parameters before you look at results. Write down how many inputs the EA has and how many values each one can take. A tool with ten inputs and ten values each has an enormous space to search, and searching it long enough will find something that looks excellent by chance. Set your limit first, then respect it.
2. Split the history into rolling windows. A workable shape on gold is twelve months of in-sample optimisation followed by three months of forward trading, rolled forward one quarter at a time across five to eight years. Gold has ranged and trended in that period, so the out-of-sample windows see more than one kind of market.
3. Optimise only inside the in-sample window, then freeze. No peeking. If you produce a parameter set with the best in-sample result, use that set for the whole forward window and change nothing during it.
4. Concatenate the out-of-sample windows. This combined curve — never the in-sample one — is your hypothesis. It is the only result you are allowed to quote to yourself.
5. Count the trades. Judge no single forward window on fewer than about thirty trades, and aim for at least three hundred across the whole test. Below that you are reading noise with a chart attached.
6. Expect decay and measure it. Out-of-sample performance is normally worse than in-sample. A drop of thirty to sixty percent in profit factor is ordinary. A drop from 2.5 to 0.6 is not decay; it is the optimisation being caught.
7. Check parameter stability. Look at the optimisation surface, not just the single best row. If the optimum is a spike surrounded by poor results, it is fitted to noise. A broad plateau of decent results is a healthier sign than one spectacular peak.
8. Repeat with a different window length. Use nine months in-sample and three forward, then twelve and four. If the conclusion flips between the two, you do not have a conclusion.

MetaTrader 5 makes step two easier because its optimiser supports a forward period directly, which reserves part of the history for out-of-sample testing. Use it, and then repeat the exercise manually on a different period anyway.

What walk-forward testing does not do: it does not use real ticks, it does not remove the need to model your costs, and it cannot tell you about the future. It is a filter that removes a large class of self-deception. Treat it as the strongest filter available, not as proof.

What does a demo forward test measure that a backtest cannot?

It measures your broker’s actual fills, which is the one input no simulation can supply. The demo account is the experiment, and it only produces usable evidence if you set it up to be comparable to the thing you intend to trade.

The rules that make a demo test worth reading:

• Same broker, same account type, same symbol, same timeframe. A demo at a different broker with tighter gold spreads measures a market you will never trade.
• At least six to eight weeks and at least one hundred trades. Scalping gold EAs reach that trade count quickly. Swing EAs need calendar time more than trade count, so be patient with them.
• Log every trade, by hand or in a spreadsheet. Record the spread at entry, the requested price versus the filled price, commission, swap, the maximum adverse excursion — how far the trade went against you before closing — and the result.
• Compare distributions, not totals. Average win, average loss, win rate, profit factor, the size of the worst loss as a multiple of the average loss, and the equity curve’s shape. A live account that matches the backtest’s totals while taking twice the backtest’s average loss is a different system.
• Publish it somewhere verified. Myfxbook or an equivalent tracker timestamps the record, which removes the temptation to reinterpret a bad month later. It also stops you from quietly retesting until the story improves.

Then apply a decision rule before you start, so that the result cannot be argued with afterwards. If live expectancy comes in at less than half the backtest’s expectancy across at least one hundred demo trades, treat the hypothesis as rejected. Do not “give it more time” — that is how a losing strategy acquires a permanent place in your portfolio.

One honest caveat. Demo fills are usually better than live fills, because nobody is taking liquidity out of your order. A demo that merely matches the backtest is therefore inconclusive rather than encouraging. A demo that underperforms the backtest is genuinely informative, because you have found a real problem while it is still free to find.

Two more things worth doing before any live money moves. Pick your broker for execution quality rather than for a sign-up bonus — our best forex brokers comparison covers what to look for, and gold spread consistency belongs near the top of the list. Then size the position so that a normal gold drawdown is survivable. A twenty to thirty percent floating drawdown is unremarkable on an aggressive XAUUSD strategy. If that would break your account or your nerve, the lot size is wrong, not the market.

What are the warning signs of a curve-fitted gold EA?

The signs are mostly about what the seller will not show you. A fitted system is usually a real strategy with parameters tuned to one specific stretch of gold history, and the tuning is invisible by design.

Watch for these:

• No readable source code. You cannot audit what you cannot read. Compiled files reveal nothing about how signals read bars or whether spread is applied.
• A backtest that begins in a strong gold uptrend. Many gold tools are tested from 2023 onwards, when gold rose for reasons that may not repeat. Ask for a window containing a long range.
• Published “optimal settings” for a specific broker. Parameters tuned to one broker’s spread become wrong the moment the spread changes.
• Grid or martingale logic with no hard stop. The equity curve is smooth because losses are deferred into floating drawdown. It stays smooth until the one move that is not survivable — and that move has not been in the test window yet.
• Stops set in points with no volatility reference. A 30-point stop is wide in a quiet Asian session and impossibly tight on a news release. Strategies that ignore volatility usually hide their worst case in the tester’s invented intrabar path.
• Results shown only as screenshots. No report, no trade list, no way to reproduce it.
• No mention of the spread the test used. If the assumption is not stated, assume it was optimistic.

Stack that list against a tool before you spend time on it. Two or three hits and you are looking at marketing rather than a system. A tool that survives the list is worth testing — and you can also reduce the part of the portfolio that depends on someone else’s honesty by keeping the risk side mechanical, which is what a drawdown reduction tool is for.

What is in the download, and what does it not do?

It is the Trillex 10s EA, a single-file open-source MQL5 expert advisor for XAUUSD released under the MIT licence by NadirAliOfficial. You get readable source code, which is the point: the audit method in this post is only possible if you can open the logic and look at it. It also illustrates the thesis here better than any invented example, because the strategy generates signals on ten-second synthetic bars, and its repository notes describe repairs made so a synthetic backtest could survive contact with real ticks. We have not independently re-tested those repairs; we are reporting what the notes say.

That list is the whole argument in miniature. According to those notes, the live-ready build switched on a spread filter that had been disabled, added a hard stop loss where the default had been none, added a confirmation rule so the Parabolic SAR has to hold its new direction before a flip counts, and moved the synthetic bars to a mid-price so a widening spread is not misread as a price move. Every one of those changes closes a gap between a modelled tick and a real one, which is the gap that decides whether a gold backtest means anything.

Item	Detail
Name	Trillex 10s EA
Author	NadirAliOfficial
Licence	MIT
Cost	Free — no licence fee, no paywall
Platform	MetaTrader 5
Instrument	XAUUSD (gold), attached to an M1 chart
Approach	A three-EMA stack plus Parabolic SAR, with an optional ten-second synthetic bar mode, a spread filter and a fixed stop loss
Source	GitHub repository, linked in the download card above and in the sources below

Read the logic before you attach it. Three things to check with the checklist from earlier in this post: how the entry price is compared against the bid and the ask, whether the stop distance adapts to volatility or is a fixed number, and whether any signal is read from a bar that has not closed. Those three answers tell you more about the tool than any performance claim could.

What it does not do, stated plainly:

• It does not come with a verified live track record. We have not traded it for you, and the author does not publish an audited account.
• It does not promise profit. No automated system does. Trading gold with leverage carries a high risk of losing money, and past results — simulated or real — do not indicate future outcomes.
• It does not remove the work in this post. You still need your own backtest at your own costs, and your own demo forward test at your own broker.
• It does not come with support, updates or settings tuned to your broker. It is a small open-source project maintained by one author in his own time.
• It is not a rehosted commercial EA. It is an independent build with its own assumptions and its own failure modes.

If you want a second reference point on what a gold EA can realistically be, our round-up of MT4 gold expert advisors covers the trade-offs between scalping, breakout and mean-reversion designs. The free indicator library is also worth a look if you would rather test your own entries with a free MT4 and MT5 indicator before automating anything. And if you are still choosing a build, start from the gold EA category rather than from a forum thread.

What should you do next?

Change the order of operations, starting with the EA you already have. Re-run its backtest at your broker’s realistic gold spread with commission included. If the edge survives, put it on a demo at the same broker for eight weeks and log every trade. If live expectancy lands under half the backtest’s, stop.

That is the whole method. The backtest is a hypothesis. The demo account is the experiment. Nothing you download changes the sequence.

So do one concrete thing today: take the gold EA you were most tempted by, and give it the audit list rather than the benefit of the doubt. If you would rather start from something you can actually read, download the open-source gold EA from the card above, open the source in MetaEditor, and run your own forward test on demo data. It costs you an evening and it costs you nothing else.

Neither the file nor this page is financial advice, and neither can protect you from a losing month. What the method can do is keep the losses small, visible, and survivable while you find out whether there was ever an edge there — which is the only question automation can honestly answer.