Aurum Nexus Matrix was published on 11 August 2026 on the MQL5 market. At the time of this test the listing showed 250 dollars, ten activations, twelve downloads of the demo version, and no reviews at all. Not a single one. That is normal for a thirteen day old product, but it is information: nobody else had assessed it publicly.

The listing describes a system designed for gold only, with dedicated execution profiles per broker, one for IC Markets and one for RoboForex. On top of that come an optional artificial intelligence layer, three risk modes, a news filter and an integrated dashboard. The wording is measured and promises no profit, which on this market is already a good surprise.

Two passages deserve to be read word for word, because they come from the developer and because they answer questions every buyer asks.

On loss protection, the listing states that the maximum drawdown protection is an account level protection, and that it should not be confused with a guaranteed fixed stop loss on every individual trade or basket. Section 9 shows what that means concretely in the order log.

On the recovery mechanism, the listing says it is not an unrestricted grid and that it does not guarantee that every basket recovers. It adds a direct warning: when the mechanism activates, total market exposure may increase. Finally, on the artificial intelligence, it specifies that this is an additional management layer and that it does not replace the entry strategy.

None of the conditions below come from me. They were written by the developer before the test began, which is the best guarantee of a fair test: when someone supplies the ground themselves, they cannot contest it afterwards.

The settings file was loaded exactly as delivered, without a single parameter modified. I re-downloaded it from his message to compare checksums: identical byte for byte. Two lines of that file come back throughout this page, the risk mode and the maximum drawdown threshold.

The version tested is 1.00, the one on sale at the time of the test. A version 1.14 became the version on sale during the testing period. Every figure on this page relates to 1.00.

If you want to reproduce this protocol on your own robots, the full method is in our guide on how to backtest an EA in MetaTrader 5, and report reading in how to read a backtest report.

The robot ran for nine days on an IC Markets demo account with the recommended configuration. Result: two trades, two winners, plus 23.38 dollars, that is plus 2.34 percent of the account. Holding time: seventeen seconds and twenty-three seconds.

Let us be clear about what those figures are worth. Two trades prove nothing statistically. Nobody can conclude anything about the profitability of a system from two operations, and I will not do so.

What this live test does establish is useful in another way: the installation works, the observed behaviour matches the product description, and the announced frequency is correct, around two positions per week. That is the role of a short forward test: verify conformity, not measure performance.

The developer states that his robot was developed, optimised and tested on IC Markets from 2023 onward. That is his official reference period, so that is where we start.

From January 2023 to August 2026, with his recommended configuration, the MetaTrader report gives:

On the period he claims, this robot makes money. That has to be said clearly, without reservation, because it is measured.

One nuance applies to the balance curve. Maximum drawdown reaches 46.65 percent. Concretely, starting with a thousand dollars, you had to watch your account fall below six hundred before it recovered. Many traders cut out before that. It is the whole difference between a backtest, which runs to the end mechanically, and a human in front of a screen.

Here is the central experiment of this test, and it is entirely simple. I change neither the robot, nor the settings, nor the broker, nor the leverage, nor the balance. I only change the date I start on.

On the five year window the developer raises a fair reservation, and I pass it on as it stands: he considers the IC Markets data before 2023 less reliable, because that is not what he built his system on. I therefore do not base my conclusion on the five year window.

But look at the last two rows. His window is longer than the three year test: it contains the same thing, plus seven months. And yet it earns more than a thousand dollars extra. A longer period earning more than a shorter one contained inside it means one precise stretch carries the entire result.

Those seven months of difference are the first seven months of 2023. I tested them on their own, under the same conditions.

A profit factor of 4.13 with 8 percent drawdown is an exceptional result. On that stretch, this robot is excellent. And the three years that follow produce zero.

One mechanical element amplifies the effect: position size is indexed on the account balance. A strong start therefore increases capital, which increases lot size, which inflates everything after it. Those seven months do not explain a hundred percent of the gap, but they are its engine.

And January 2023 is, word for word, the start of the period on which the developer states he developed and optimised his system. A system that wins exactly on its optimisation window and produces zero right after has a name, and it is not performance.

Every test above beyond four and a half months uses one-minute bars, because IC Markets only provides real ticks from 1 April 2026. This is not a choice, it is a data constraint, and it has to be stated.

On that four and a half month window in 100 percent real ticks, with the drawdown threshold set to the robot’s own default value, the result is positive:

That is the best figure in the whole file, and it comes from the best data. One reservation only, but it matters: 43 trades is a small sample.

A technical remark belongs here, and it works in the product’s favour. On the one window where both data models exist, I compared them: real ticks give minus 78.93 dollars, one-minute bars give minus 208.05. With the other configuration, minus 235.54 against minus 636. Twice out of twice, the bar model is harsher. The long window results shown above are therefore more likely pessimistic than flattering.

The developer’s point about tick volume

After the tests were finished, the developer added a technical element that belongs here. He states that his robot uses ten entry strategies, and that a trade is only executed once the conditions are confirmed by tick volume. He concludes that a backtest may not reflect the real behaviour of the system, and that a one month forward test would be more representative.

He is right for the one-minute bar tests. In that mode tick volume is modelled, not real. It is also consistent with what the measurements show: that model came out harsher than real ticks, twice out of twice.

It does not apply to the real-tick tests. Those run on the broker’s own recorded tick stream, volume included, and MetaTrader reports « History Quality: 100% real ticks » on those reports. On that window the entry logic is therefore evaluated on exactly the data it was designed for.

The developer adds a nuance that I pass on, because it is correct: a historical real-tick test remains a historical test. It replays the broker’s recorded stream, but it does not reproduce the tick sequence, the spread or the execution conditions the robot will meet in real time. That is exactly why a forward test still matters, and I have started a one month run on a demo account with his recommended configuration. The results will be published here.

Followed to its conclusion, the remark points to the four and a half month real-tick window as the reference measurement. It is also the smallest in the file: forty-three trades. It gives plus 128.84 dollars with the drawdown threshold corrected, and minus 78.93 with the full recommended configuration.

The first settings file I received, for a thousand dollar account, set the maximum drawdown threshold to 50. The default value written in the robot’s own file is 15.

I changed that single number, nothing else, and replayed the five months available in real ticks.

The reading is clear. Losing months are divided by a factor of three to seventeen, while winning months lose only a third, and that third matches exactly the ratio between the two risk modes. Drawdown collapses on every window, going from 70 percent to 22 percent on real ticks.

This is not hindsight optimisation: a single alternative value was tested, the one written as the default in the developer’s own file. Sweeping ten, twenty, thirty and keeping the best would have been curve fitting, and it would have proved nothing.

Be careful not to read that fifteen the wrong way, and the developer insisted on this point himself. Fifteen percent is an internal protection threshold, not a promise that the report will show fifteen percent maximum loss. The measurements confirm it: 17 percent on one window, 22 on another, then 47 and 74 over five years. The mechanism triggers, it does not cap the loss at the displayed value.

Back to the sentence from the product page quoted in section 1. In the order table of the MetaTrader reports, the S / L and T / P columns are empty. On every order, of every test, without exception. The two orders from the live account were checked in the broker’s order book, with the same result.

This is neither a flaw of the test build nor a configuration error: it is a deliberate design choice, position management happening at basket level, internally to the robot. The product page now describes it correctly.

The practical consequence must be understood before any purchase. If your terminal closes, nothing on the server side will close your positions. A power cut, a Windows restart or a lost connection leave the exposure open with no net. That is a serious reason to run this type of system on virtual hosting, and it is the non negotiable point of this test.

An analysis that only criticises is worthless, and several points deserve to be credited to this product.

It is not a martingale. Verified in the volume column of the reports: added positions are smaller than the initial entry, never larger. That is the opposite of the pattern that destroys accounts.

The artificial intelligence is real. A model is embedded in the file, with twenty inputs and one output. Honest nuance, and the developer states it himself on his page: this layer works on management and position exit, not on the entry decision.

Live behaviour matches the description, and the announced frequency is accurate.

The product page is measured. No promise of profit, an explicit warning about increased exposure during recovery, and a clear statement that the drawdown protection is not a guaranteed stop loss. That is more honest than the average on this market.

And on the period he claims, the robot wins. That measurement does not move.

This verdict is neither a buy recommendation nor a rejection. It states the conditions under which the result holds, and those under which it breaks.

Where it holds: over the last two years, with the drawdown threshold at fifteen, the system is positive, including on the best data available. There is no hidden martingale, and the product page describes the behaviour correctly.

Where it breaks: as soon as the start date moves. Seven months of shift are enough to wipe out the entire gain. The result you get therefore depends first on the day you begin, and nobody controls that parameter.

The non negotiable point: no stop is placed at your broker. A trader who cannot live with that characteristic does not need to examine the rest.

The lesson goes beyond this product, and that is what to remember. Take any robot, run it on three different start dates, and see whether the result holds. A result that does not survive a shift of a few months is not a strategy, it is a coincidence, well told.