
Technical analysis can be discretionary or systematic. The same distinction applies to automated trading: putting a strategy inside a piece of software does not tell us whether the underlying strategy is any good. An Expert Advisor, or EA, is simply a program used with MetaTrader to automate trading-related decisions and actions. It can monitor market conditions, generate signals, manage positions and, when automated trading is enabled, place or modify orders according to programmed rules. MetaTrader itself describes Expert Advisors as trading robots designed to perform trading operations.
That definition is much less exciting than the way Expert Advisors have often been marketed. Over the years, retail traders have been offered thousands of EAs promising to remove emotion, identify opportunities automatically and generate returns while their owners sleep. Some are legitimate pieces of automation built around clearly specified strategies. Others rely on fragile trading logic, aggressive position sizing or backtests that look considerably better than the strategy behaves in live markets.
We spent more than a decade working in the retail FX industry, including in brokerage management. From the broker side, we saw some of these systems being used across real client accounts. Certain EAs could generate remarkably convincing trading histories for months before a change in market conditions exposed how much risk had been accumulating underneath.
That experience is the reason for this article. This is not an argument that automated trading is inherently bad. Automation is a tool. The more useful question is:
What exactly is being automated, how was the strategy tested, and what risks are hidden behind the performance that the trader sees?
Those questions were relevant during the MetaTrader 4 Expert Advisor boom. They are just as relevant now that similar ideas are increasingly described as trading bots, algorithmic trading systems and, more recently, AI trading bots.
A poorly designed trading robot isn’t necessarily fraudulent. Our guide to trading scams explains how fabricated performance and deceptive sales claims differ from legitimate strategy failures.
What Is an Expert Advisor in Forex?
An Expert Advisor is a program that operates within the MetaTrader ecosystem. Developers can use MQL4 or MQL5 to define rules governing when the program should analyze conditions, open trades, close trades, modify orders or manage risk. The EA itself is not a trading strategy. It is the software that implements one. That distinction matters.
A developer could program an EA to follow a moving-average strategy, trade breakouts, rebalance positions, exploit a statistical signal, average into losing trades or simply execute a predefined risk-management process. The quality of the resulting system depends on the logic it implements and how that logic behaves under real market conditions.
MetaTrader gives users extensive tools for testing these systems. Its Strategy Tester can run an EA against historical data, vary parameters during optimization and perform forward testing on a separate portion of the sample. The platform can also simulate execution delays and use real tick data to make testing conditions more realistic.
These capabilities are useful. They do not automatically make a strategy valid. A sophisticated backtesting platform can test a good hypothesis or a bad one equally efficiently. That is a theme we will return to later.
From Expert Advisors to AI Trading Bots: What Actually Changed?
The term Expert Advisor is closely associated with MetaTrader, but the underlying idea is much broader: encode a set of trading decisions in software and allow a computer to execute them. Today, the same basic concept is often described using terms such as trading bot, automated trading system or algorithmic trading. More recently, the word AI has become part of the vocabulary as well.
These terms should not be treated as interchangeable in a technical sense. An automated system may be completely rule-based and contain no artificial intelligence at all. Another system may use machine learning as part of its forecasting process. A third may use a large language model to analyze information or participate in decisions. And increasingly, AI may not be part of the trading strategy itself—it may simply have been used to write the code. That last development deserves particular attention.
MetaTrader already made EA creation accessible before modern generative AI. Its MQL Wizard, for example, can generate an Expert Advisor without requiring the user to write the program manually: users select trading signals, money-management rules and other predefined components, and the platform generates the code. Large language models and coding agents take that accessibility much further.
Instead of selecting components from a predefined menu, someone can describe a trading idea in ordinary language and ask an AI coding tool to implement it. The barrier between having an idea and having executable software has therefore become much smaller. That can be enormously useful.
It can also create a new mismatch between the ability to build a trading system and the ability to evaluate one. A program may compile correctly and still implement something different from what its user intended. Position sizing, signal timing, order-management logic, indicator calculations and seemingly minor implementation choices can materially alter how a strategy behaves.
More fundamentally, the software may implement the requested idea perfectly while the idea itself has no durable statistical edge. This is where the current enthusiasm around AI trading starts to resemble an older story.
Regulators have already warned investors about marketing that uses artificial intelligence to lend credibility to automated trading products. The CFTC notes that promoters have used claims about AI-created algorithms and trading bots to advertise unrealistic or guaranteed returns, while emphasizing that AI does not provide an ability to predict sudden market changes or turn automated strategies into guaranteed money-making systems.
The important point is not that AI cannot be useful in trading. It plainly can be used for data analysis, model development, automation and other legitimate applications. The point is much simpler:
AI has made trading strategies easier to automate. It has not made them easier to validate.
Or, put another way:
The programming barrier has fallen dramatically. The knowledge barrier has not.
Whether the program is called an Expert Advisor, trading robot or AI trading bot, the same questions remain:
What exactly causes the system to enter and exit a trade?
What determines position size?
What happens when a trade moves against it?
How was the strategy developed?
How much parameter searching or optimization was performed?
Was it tested on genuinely unseen data?
Were spreads, commissions, slippage and execution realistically modeled?
How large can its drawdowns become?
What happens when the market behaves differently from the period used to develop it?
A new label does not answer any of those questions.
What We Saw From Inside a Forex Brokerage
This article originally grew out of a list of Expert Advisors that had circulated through retail forex communities and that we could identify from trading activity.
Many MetaTrader EAs leave identifying information in the comments attached to trades. In a brokerage environment, this can make it possible to recognize groups of accounts apparently using the same software or strategy.
That provided an unusual perspective. Instead of seeing only a developer’s backtest or a trader’s selected account statement, we could observe what happened when similar automated systems appeared across multiple client accounts.
Some of the systems discussed below also appeared in public retail-trading communities and EA databases. Our historical list should therefore be understood for what it is: a collection of systems that attracted attention and, in some cases, were also identifiable in brokerage trading data—not a controlled statistical sample of all Expert Advisors ever used by retail traders.
That distinction is important. We cannot use these observations to claim that a specific percentage of all EAs fail. Nor does the failure of one family of automated strategies prove that algorithmic trading itself does not work.
What these cases can show much more effectively is how particular trading and position-sizing structures can create an illusion of stability before producing severe losses.
And that brings us to two concepts that appeared repeatedly in the systems we encountered:
grid trading and Martingale-style position sizing.
Grid Trading and Martingale Are Not the Same Thing
Two approaches appeared repeatedly among the Expert Advisors we encountered: grid trading and Martingale-style position sizing. They are often mentioned together, but they describe different things.
A grid strategy places or adds orders at predefined price intervals. For example, an EA might open additional buy positions every time EUR/USD falls another specified number of pips. Some grids use equal position sizes. Others vary the size of subsequent trades.
A Martingale-style strategy, by contrast, changes the amount at risk after adverse outcomes. In its classic form, the position size is doubled after each loss so that a subsequent win can recover previous losses. Retail trading systems often use the term more loosely for strategies that progressively increase exposure after the market moves against them. The two approaches can therefore exist separately.
A grid does not have to increase position size, and increasing position size after losses does not necessarily require a grid. But they are frequently combined.
A recent study of an automated FX grid system makes the same distinction: the grid determines where additional orders are placed, while a separate Martingale component increases position sizes at adverse price levels. The authors also highlight the resulting vulnerability to severe drawdowns during sustained directional moves.
That combination is particularly important for understanding many of the EAs we observed. Imagine an EA buys one unit of EUR/USD. Price falls, so it buys again. It falls further, and the EA buys again. If each new trade has the same size, the system is averaging into the position using a grid.
If each new trade is larger than the previous one, the system is also introducing Martingale-style scaling. Suppose the multiplier is 1.5. Position sizes might progress like this:
1.00 → 1.50 → 2.25 → 3.38 → 5.06
After only five entries, the system is carrying more than 13 times the initial position size in cumulative exposure. With classic doubling, the progression is even more aggressive:
1 → 2 → 4 → 8 → 16
Five entries create total exposure of 31 units from an initial position of just one. This is why looking only at the first trade tells us very little about the risk of such a system. The important question is what happens after the market moves against it.
Why These EAs Can Look So Profitable
The appeal of these systems is easy to understand. Suppose an EA buys EUR/USD and the price falls. Instead of accepting the loss, it opens another buy position at a lower price. The average entry price of the combined position moves down.
If the market subsequently rebounds—even without returning to the original entry price—the basket may be able to close near break-even or with a small overall profit. Repeat that process successfully many times and the trading history can look remarkably attractive.
The EA may record:
a very high percentage of profitable trading cycles;
long sequences of winning trades;
relatively smooth growth in realized account balance;
few obvious losing trades;
apparently rapid recovery after adverse price moves.
That can create a powerful impression of consistency. But the apparent consistency can come from postponing losses rather than eliminating them. Instead of closing a losing trade, the system adds another position.
If price turns back, the accumulated basket may recover and close profitably. If price does not turn back, however, exposure continues to grow while the open loss becomes larger. This creates an asymmetric payoff structure:
many relatively small realized gains can coexist with the possibility of a much larger eventual loss.
The danger is therefore not necessarily visible in the system’s win rate. An EA could win on most completed trading cycles and still carry substantial downside risk if the occasional losing sequence is large enough.
This is one reason a claim such as “95% winning trades” tells us surprisingly little about whether an automated strategy is actually attractive. We need to know how much it wins when it wins, how much it loses when it loses, and how much risk it carries before those outcomes are realized.
The Difference Between Balance and Equity Matters
This is where another distinction becomes crucial: account balance and account equity are not the same thing. In MetaTrader, the balance does not include the profit or loss on positions that remain open. Equity does. So imagine an account starts with $10,000.
The EA repeatedly closes small profitable baskets and the account balance gradually rises to $11,000. At first glance, that looks excellent.
But suppose the EA is simultaneously holding a group of losing positions with a combined floating loss of $3,000. The balance can still show approximately:
$11,000
while the account’s equity is closer to:
$8,000
before allowing for other account adjustments. The smooth balance curve and the economic reality of the account are telling very different stories. This matters particularly for strategies that are reluctant to realize losses.
An EA can continue reporting completed winning trades while an increasingly large losing basket remains open. The realized history looks stable because the most damaging part of the strategy has not yet passed through the balance. The risk is sitting in floating P/L.
MetaTrader’s own reporting distinguishes the two measures for exactly this reason: balance excludes currently open positions, whereas equity incorporates their current profit or loss.
For anyone evaluating an EA, a smooth balance curve should therefore never be examined in isolation. Look at the equity curve as well. A large and persistent gap between balance and equity can reveal risks that the realized trade history obscures.
Recent research examining a grid system combined with Martingale-style scaling illustrates how dramatic that difference can become. In one particular experimental system, maximum equity drawdown approached 80%, while maximum balance drawdown was only about 14%. That single study should not be generalized to all grid strategies, but it provides a useful illustration of how realized-balance statistics can dramatically understate floating risk.
The Strategy Works—Until the Recovery Does Not Arrive
None of this means that price must move in one direction forever for the system to fail. A grid merely needs an adverse move that is sufficiently large, sufficiently persistent, or sufficiently expensive to survive.
As positions accumulate, several constraints become increasingly important:
Capital is finite.
A trader cannot continue adding positions indefinitely.
Margin is finite.
Leveraged positions consume margin, and falling equity reduces the cushion available to maintain them.
Position size can grow quickly.
Martingale-style multipliers can make exposure rise much faster than the original trade suggests.
Markets do not have to revert on the trader’s schedule.
A currency can continue moving long after a strategy expects a reversal.
MetaTrader calculates margin level using the relationship between account equity and required margin. When equity deteriorates sufficiently relative to margin requirements, the account can ultimately enter margin-call or stop-out territory under the broker’s applicable trading conditions.
This is why saying that these systems “profit in sideways markets and lose in trending markets” is directionally useful but incomplete.
The deeper problem is path dependence.
Their survival depends not just on where the market eventually ends up, but on how far it moves against the accumulated position before any recovery occurs.
A currency can eventually return to the EA’s preferred level and the trader can still be forced out before that happens.
That explains one of the most deceptive characteristics of this family of strategies:
They can appear unusually reliable precisely because losses are repeatedly deferred—until one adverse sequence becomes too large to defer any longer.
The correct question is therefore not:
How often does the EA win?
It is:
What happens to the account when the recovery it depends on does not arrive soon enough?
What We Actually Saw in Client Accounts
The risks described above were not theoretical to us.
During my years working inside retail FX brokerages, we could sometimes identify particular Expert Advisors across client accounts because the software left recognizable comments on the trades it opened.
MetaTrader allows orders to contain a comment field, and MQL exposes that information programmatically. Some EA developers used these comments to identify their software or particular types of trades. In practice, that meant a distinctive string attached to an order could sometimes act almost like a fingerprint. (mql5.com)
From the broker side, we could search trading records for these recurring strings and examine accounts apparently running the same Expert Advisor.
This gave us a perspective that an individual trader normally does not have.
A trader sees one account.
A developer may present one backtest or selected live account.
A broker can potentially see the same recognizable trading logic appearing across many unrelated accounts.
That does not turn brokerage observations into a controlled academic study. We did not randomly sample all Expert Advisors, assign traders to strategies or construct a formal survival database. So we would not use these observations to estimate what percentage of all EAs succeed or fail.
But they were extremely useful for seeing how particular strategies failed in practice.
When the Same EA Appears Across Multiple Accounts
One pattern was particularly memorable.
An Expert Advisor would become popular in retail trading communities. Clients would begin running it on their accounts, sometimes after seeing attractive historical results or recommendations from other traders.
In some cases, clients contacted brokerage support for help installing or configuring the software. The person using an automated strategy therefore did not necessarily understand its code—or even all of the trading logic being executed on the account.
For a while, everything could look fine.
The EA would accumulate profitable trades. Account balances would rise. Traders would share their results. More people would start using the same system.
Then market conditions changed.
Because many accounts were running similar logic, we could see similar problems emerge across otherwise unrelated traders: positions accumulating in the same direction, exposure increasing as the market moved against them, floating losses expanding and, eventually, accounts suffering severe drawdowns or reaching their margin limits.
We witnessed two particularly striking episodes involving EAs identifiable in our records.
The first involved a system we knew as FfirstTea.
Several years later came a much larger episode involving an EA from the Phoenix family.
These were not simply cases of one trader making an unfortunate decision. What caught our attention was seeing recognizable automated behavior across multiple client accounts and then watching similar weaknesses surface when the strategies encountered conditions they could no longer absorb.
Phoenix: A Clear Example of the Mechanism
Phoenix is particularly useful because we retained enough historical information to reconstruct how one version of the strategy behaved.
In our separate Phoenix Expert Advisor case study, we plotted trades from a historical account directly onto EUR/USD charts.
The pattern was revealing.
When the initial trade moved against the strategy, the EA continued opening positions in the same direction. Position sizes increased as the adverse move continued. When the market eventually retraced, the basket could sometimes recover and close at a profit.
For substantial periods, this produced exactly the behavior we described earlier:
adverse move → additional exposure → eventual retracement → basket closes → balance recovers
And sometimes the recovery really did arrive.
That is part of what made the system convincing.
A strategy like this does not have to fail immediately. In fact, if it failed after three trades nobody would be particularly interested in it. Its appeal comes from surviving repeated episodes that appear to demonstrate that the recovery mechanism “works.”
The historical Phoenix account survived several alarming periods in which positions remained underwater for extended stretches before the market finally moved back in its favor.
Eventually, however, repeated recovery is not enough.
If one adverse sequence becomes larger than the capital and margin available to withstand it, the previous successful recoveries do not protect the account.
That is the central weakness of this type of strategy.
It does not need to be wrong very often.
It needs to be wrong too far, for too long, once.
Survivorship Can Make the Story Even More Convincing
There is another reason systems like these can become popular.
While an EA survives, its successful history remains visible.
Every recovered basket becomes another example its users can point to:
“Look—it went into drawdown before and came back.”
But an earlier recovery does not establish that the next adverse move will remain within the account’s ability to survive.
This creates an important distinction between two questions:
Has the strategy recovered from large drawdowns before?
and:
Is there a defensible limit to how large its future drawdown can become?
Those are not the same question.
A strategy that repeatedly adds risk during losses can produce an impressive survival history right up until the point at which it no longer survives.
That is why the individual EAs in our historical database are most useful as case studies in strategy structure and risk, not as a blacklist proving that every program carrying a particular name must fail.
The names may change. Developers may release new versions. Completely unrelated products can even reuse the same name years later.
The behavior is what matters.
If an Expert Advisor repeatedly increases exposure as losses grow, postpones realization of those losses and depends on an eventual market retracement to escape the position, the relevant question is not how successful its previous recoveries looked.
It is:
How much adverse movement can the account survive before the strategy runs out of capital or margin?
Why EA Backtests Can Look Better Than Reality
One of the main selling tools for an Expert Advisor has always been the backtest.
The developer runs the EA over several years of historical prices and presents the resulting equity curve:
steady growth, limited drawdowns, hundreds or thousands of trades and perhaps an impressive risk-adjusted return.
Backtesting is not the problem.
Any systematic trading strategy should be tested before real capital is placed behind it. MetaTrader includes a sophisticated Strategy Tester precisely for this purpose.
The problem is treating a successful backtest as proof that the strategy has discovered a durable edge.
It has not.
A backtest tells us how a particular set of rules would have behaved under a particular historical simulation. Whether those results contain information about future performance is a separate question.
Optimization Can Manufacture an Excellent Historical Strategy
Expert Advisors often contain adjustable parameters.
A moving-average system might allow the user to change the lookback periods. A grid EA may have settings for the distance between entries, maximum number of positions, lot-size multiplier, take-profit level and dozens of other variables.
MetaTrader can automatically test many combinations of these parameters and identify those that produced the strongest historical results.
That is useful—but it introduces a serious statistical problem.
Suppose we test enough combinations of:
indicators;
parameters;
currency pairs;
timeframes;
entry rules;
exit rules;
stop levels;
position-sizing rules.
Eventually, some combination will look unusually good simply because it happened to fit the historical sample well.
The more alternatives we try, the greater this risk becomes.
This is backtest overfitting.
Research on investment backtests has examined this problem formally. Bailey and colleagues show why selecting strategies from repeated historical trials can produce apparently exceptional results that subsequently deteriorate out of sample. Related work on the Deflated Sharpe Ratio addresses the same underlying problem: once many strategy variations have been tested, the best reported result may partly reflect selection bias rather than genuine forecasting ability.
This is particularly relevant to automated trading because computers make experimentation cheap.
An optimization engine can test thousands of parameter combinations much faster than a human could evaluate them manually.
But computational power does not create information.
It can just as easily make it faster to discover historical coincidences.
The Best Backtest May Be the One We Should Distrust Most
Imagine that an EA developer tests 5,000 different configurations and shows you the one with the smoothest equity curve.
You are not really looking at one experiment.
You are looking at the winner of 5,000 experiments.
That distinction matters.
If the developer does not disclose how many rules and parameter combinations were tested before arriving at the published version, an apparently remarkable result becomes much harder to evaluate.
This does not mean optimization is inherently bad.
Parameters often need to be calibrated.
The problem arises when the same historical data are repeatedly used both to discover the strategy and to prove that it works.
A system can gradually become tailored to the noise, quirks and market regimes contained in that particular sample.
It then encounters new data that were not available during development—and the performance disappears.
Forward Testing Is a Stronger Test
A better approach is to separate the data used to develop the strategy from the data used to evaluate it.
MetaTrader provides a forward-testing function for exactly this purpose.
During optimization, one part of the historical period is used to select parameters. The chosen configurations are then evaluated on a separate forward period that was not used for the original optimization.
MetaTrader explicitly describes this feature as a way to help avoid parameter fitting or over-optimization.
That makes the second period considerably more informative.
If an EA performs beautifully during optimization and then immediately deteriorates on the forward sample, the original result was probably less robust than it appeared.
Forward performance does not prove that a strategy will continue working forever either.
It simply asks a more demanding question:
Does the relationship survive data that were not used to select the strategy?
For an EA being marketed on the strength of a historical backtest, that is a much more useful question than asking how impressive the optimized curve looks.
Execution Assumptions Matter Too
Even a strategy that survives out-of-sample testing can look better in simulation than in actual trading if the simulation assumes execution conditions that are too favorable.
MetaTrader’s Strategy Tester illustrates why this matters.
It allows users to test under different tick-generation and execution settings. One setting executes orders without delay at requested prices—conditions MetaTrader itself describes as ideal. Other modes can introduce fixed or random execution delays. Testing can also use actual historical tick data supplied by a broker rather than relying on coarser simulated price paths.
Those differences can matter enormously for some strategies.
A system that trades infrequently and holds positions for weeks may be relatively insensitive to a few milliseconds of execution delay.
A short-term EA that attempts to capture tiny price movements may be far more sensitive to:
spreads;
commissions;
slippage;
execution delays;
requotes;
tick-level price movement.
If the expected profit per trade is very small, seemingly minor differences between the simulated and live environment can consume much of the apparent edge.
So when somebody presents an impressive EA backtest, the question should not simply be:
What was the return?
It should also be:
Under what assumptions was that return produced?
Historical Performance Is Still Historical
There is one limitation no Strategy Tester can eliminate.
The test uses the past.
Even a properly designed EA can experience performance decay because market behavior changes. Liquidity changes. Volatility regimes change. Trading costs change. Other participants adapt. Relationships that existed during one period may weaken or disappear during another.
This means we should think about EA testing as a sequence rather than a single backtest:
Define the strategy → test it historically → evaluate it on unseen data → include realistic trading conditions → stress-test its assumptions → observe how it behaves going forward.
No individual step proves that the EA possesses a permanent edge.
But each additional test makes it harder for a fragile strategy to masquerade as a robust one.
And this becomes especially important with the AI-generated trading systems discussed earlier.
If an LLM can produce dozens of strategy variations in minutes and an optimizer can test thousands of parameter combinations immediately afterward, the number of hypotheses a trader can search has exploded.
That makes testing easier.
It also makes overfitting easier.
The ability to generate more strategies therefore makes statistical discipline more important, not less.
A beautiful backtest should be the beginning of the investigation—not the end of it.
Historical Expert Advisors We Identified in Client Accounts
The list below did not come from internet rankings, reviews or a collection of Expert Advisors we simply disliked.
It came from actual brokerage trading records.
During my time working at retail FX brokers, clients used many different Expert Advisors. Some could be identified because the orders they generated contained recognizable comments or other recurring identifiers.
MetaTrader allows orders to carry a comment field, and MQL also provides identifiers specifically associated with Expert Advisor activity. In practice, these fields can make it possible to recognize trading activity generated by a particular EA across different client accounts.
That is how the table below was built.
How the List Was Compiled
The process had two parts.
First, we needed to identify the trading software.
Public EA databases, trading forums and other sources were useful here because they sometimes documented the commercial name of an Expert Advisor together with comments or strings left in its trades. Once we knew what to look for, those identifiers could be searched for in brokerage trading records.
For example, if an EA inserted a distinctive word into its order comments, a database query could search for that string even when the full comment contained additional information such as take-profit or stop-loss markers.
Second—and much more importantly—we examined what happened to the client accounts using those systems.
An EA was not added to this list merely because somebody complained about it online. The systems below were used by clients of the brokerages where we worked, and the accounts we observed running them subsequently suffered losses.
In a number of cases, the pattern was particularly clear. The same recognizable EA appeared across multiple unrelated accounts. Its trading logic could be observed over time, and eventually the weaknesses of the strategy surfaced in real trading.
That is what makes this list different from a collection of bad online reviews.
It is based on observed client trading activity.
What the List Does—and Does Not—Prove
There is still an important limitation.
This was not a controlled study of the entire Expert Advisor market.
We did not randomly select EAs, measure every user of every system or calculate a formal population-wide failure rate. The list is therefore not evidence that a certain percentage of all Expert Advisors fail.
Nor should the names below automatically be interpreted as warnings about every product currently sold under the same or a similar name. EAs can have different versions, developers can change their strategies, and names can be reused years later.
Instead, each entry should be read historically:
This was an identifiable Expert Advisor that we observed being used in brokerage client accounts, and those observed accounts subsequently suffered losses while using it.
The table is therefore most useful for two purposes.
First, it documents a part of the retail FX ecosystem that is normally difficult to observe from the outside.
Second, it allows us to look beyond marketing claims and ask what characteristics repeatedly appeared in automated strategies that ultimately caused problems for real accounts.
In many of the systems we could analyze, the answer brought us back to the same mechanisms discussed earlier: grid-style averaging, increasing exposure after adverse moves, large floating drawdowns and dependence on the market eventually retracing before available capital or margin was exhausted.
Historical EA Database
The table below contains the identifying trade-comment patterns we used where available, the Expert Advisor name we associated with them, and our classification of the strategy based on the trading behavior we observed.
The % symbols shown around some comment strings are not part of the EA’s commercial name. They were used as wildcards in database searches so that the query could identify the relevant text even when additional characters appeared before or after it.
The classifications should also be understood as descriptions of the historical versions we observed, not necessarily every subsequent version carrying the same name.
Two cases from this database stood out because of their scale: FfirstTea and, several years later, Phoenix.
With both systems, we saw losses appear across multiple client accounts rather than as a single isolated trader failure. Phoenix was particularly useful because we later reconstructed one historical account in much greater detail, allowing us to see how its increasing exposure developed trade by trade.
That case provides a much clearer picture of what a seemingly successful Expert Advisor can look like before its risk finally becomes visible.
Historical Expert Advisors Observed in Client Accounts
These Expert Advisors were identified in historical brokerage trading records. Inclusion reflects systems used by clients whose observed accounts subsequently suffered losses while running them. The table is historical and should not be interpreted as a judgment about unrelated or later products using similar names.
| Partial trade comment / identifier | Expert Advisor | Observed strategy |
|---|---|---|
%AKIB FX% | AKIB FX | Grid |
%Algo 2.0% | Algo 2.0 | Grid |
%AUTOPROFIT%SYSTEM% | FX AUTOPROFIT SYSTEM | Martingale |
%AX Trade% | AX Trade | Grid |
%BBSO% | BBSO | Martingale |
%BenderFX_AgressiveFibo% | BenderFX_AgressiveFibo | Martingale |
%Bubak% | Bubak | Martingale |
%Corti% | Corti 1.42_cortiea.com | Grid |
%Dark Venus% | Dark Venus | Martingale |
%Dash Colour Pro Ai Robot% | Dash Colour Pro Ai Robot | Grid |
%EA Antivirus% | EA Antivirus | Martingale |
%EA LUCRE% | EA LUCRE | Grid |
%EA Pongrac Franjo% | EA Pongrac Franjo | Martingale |
%EA%Gold%Stuff% | EA Gold Stuff | Martingale |
%EasyMoney% | EasyMoney EA | Martingale |
%Erizo Pro% | Erizo Pro | Martingale |
%EvolutionFX% | EvolutionFX | Unknown |
%firstea% | Fiirstea | Grid |
%FlexHybrid% | FlexHybrid | Martingale |
%Fxcore100_EA% | Fxcore100_EA | Grid |
%fxcore100% | www.fxcore100.com | Grid |
%Fxcore% | Similar to Fxcore | Grid |
%FXStabilizer% | FXStabilizer | Grid |
%Gear-A% | Gear-A | Martingale |
%Global*AF-TrendKiller% | Global*AF-Scalper.Ltd | Unknown |
%GPSForexRobot% | GPSForexRobot | Unknown |
%grid% | Name unknown | Grid |
%GSEA% | GSEA | Grid |
%Happy Algorithm% | Happy Algorithm | Unknown |
%Idea Gold% | Idea Gold EA | Martingale |
%Ilan% | Ilan | Martingale |
%Imperio Algorithm% | Imperio Algorithm | Martingale |
%IWCFX INCOME EXCLUSIVE% | IWCFX INCOME EXCLUSIVE | Martingale |
%IX Inversors% | IX Inversors | Martingale |
%Jum%StoCh% | Jum+StoCh | Martingale |
%Junzhang% | Junzhang | Martingale |
%LevelOne% | LevelOne | Martingale |
%Life Changer EA% | Life Changer EA | Martingale |
%MikeMoney EA% | MikeMoney EA | Martingale |
%MukBang% | MukBang | Martingale |
%Neo The Legend% | EANeo The Legend | Martingale |
%odin% | Odin | Grid |
%Ohanda Pro EA% | Ohanda Pro EA | Grid |
%Ordem_Tendencia% | Ordem_Tendencia_E_01_N07 | Unknown |
%OsloFX% | OsloFX | Martingale |
%Phoenix% | Phoenix | Martingale |
%PipskomeEA% | PipskomeEA-X2021 | Scalper |
%ProTradeHub% | T&K - ProTradeHub | Unknown |
%Rapea% | Rapea | Martingale |
%runwiseFX% | runwiseFX | Martingale |
%SanityScalper% | SanityScalper | Grid |
%scalp% | has SCALP in the comment | Scalper |
%SGD% | SGD | Grid |
%Signalspro007 EA% | Signalspro007 EA | Grid |
%Solo Cello% | Solo Violin | Martingale |
%Solo Piano% | Solo Violin | Martingale |
%Solo Violin% | Solo Violin | Martingale |
%StepTrade% | StepTrade | Unknown |
%The King EA% | The King EA | Grid |
%TheSkilledTrader% | TheSkilledTrader | Unknown |
%TILTBOYZ% | https://t.me/TILTBOYZ | Martingale |
%Trade Explorer% | Trade Explorer | Martingale |
%Tran Thanh Nguyen% | Tran Thanh Nguyen | Grid |
%Trio Pro% | Trio Pro HL | Unknown |
%TSNEWTickSniper% | TSNEWTickSniper | Unknown |
%Ubat% | Ubat | Grid |
%Unstoppable% | Unstoppable | Martingale |
%Very Well Trade% | Very Well Trade | Martingale |
%VPFXS% | VPFXS | Grid |
%WINNER%FX%TSR% | WINNER-FX | Grid |
%www.bdprofit.com% | www.bdprofit.com | Unknown |
%www.mforex.pro% | www.mforex.pro | Unknown |
%Ranger%EA% | Ranger EA | Unknown |
%BUY COUNTER%' OR COMMENT LIKE '%BUY TREND%' OR COMMENT LIKE '%SELL COUNTER%' OR COMMENT LIKE '%SELL TREND% | TSR-TREND-COUNTER, 4 strategies in one | Unknown |
Phoenix — When Repeated Recoveries Look Like Proof
A historical brokerage case showing how repeated recovery from large drawdowns can make a Martingale-style Expert Advisor appear more robust than it really is.
One Expert Advisor in our historical database deserves a closer look: Phoenix.
Phoenix became widely discussed in retail trading communities around 2019–2022. We could also identify it in brokerage trading records through recognizable trade comments associated with the system.
More importantly, we were able to reconstruct the behavior of one historical Phoenix account in considerably greater detail.
The example is useful because it shows why Martingale-style Expert Advisors can remain convincing for much longer than their underlying risk might suggest.
Phoenix Did Not Fail Immediately
In June and July 2020, Phoenix demonstrated the behavior that made systems like this attractive.
The EA opened a position. When the market moved against it, additional positions were opened in the same direction, with exposure increasing as the adverse movement continued.
Eventually, EUR/USD reversed.
That reversal allowed the group of positions to recover and close.
From the trader's perspective, the lesson could easily have appeared to be:
And in that particular sequence, it did.
The problem is that surviving one adverse move tells us very little about whether the same mechanism can survive the next one.
Then the Adverse Moves Became Larger
Later in 2020, Phoenix accumulated sell positions while EUR/USD continued rising.
The system was effectively waiting for the market to reverse far enough for its increasingly unfavorable basket of trades to recover.
Eventually, it received the reversal it needed.
Then something similar happened again in early 2021.
This time Phoenix accumulated buy positions while EUR/USD was falling.
By February 2021, the account had reached the point where available resources were severely constrained and the system could no longer continue adding exposure in the way it had previously.
Once again, however, the market eventually reversed.
By late March, the account escaped what had looked like a potentially terminal sequence.
We plotted the account's historical trades against EUR/USD in our dedicated Phoenix Expert Advisor case study .
The repeated recoveries are precisely what make this example so instructive.
Each escape can strengthen the belief that the strategy has been validated.
But what has actually been demonstrated?
Not necessarily that Phoenix possessed a reliable forecasting edge.
It demonstrated that, during those particular episodes, the market retraced before the account exhausted its ability to remain in the trade .
Those are very different statements.
Margin Creates a Hard Limit
A strategy can wait for a reversal only while the account remains capable of supporting its open positions.
MetaTrader defines margin level as the relationship between account equity and required margin:
Brokers define margin-call and stop-out thresholds. Once the relevant conditions are reached, positions can ultimately be closed forcibly under the broker's trading rules.
This creates a hard constraint that is sometimes overlooked when people say:
Even if the market does eventually return, the account has to survive the path taken before that happens.
Phoenix survived several severe episodes because the reversal arrived in time.
Eventually, one did not.
The Final Failure
The historical Phoenix account continued trading after its earlier recoveries.
Its balance history contained exactly the pattern we would expect from this type of strategy: periods of apparent progress interrupted by severe drawdowns, followed in some cases by dramatic recoveries.
Until the recovery mechanism finally failed.
By late 2021, losses had overwhelmed the account we reconstructed. From the brokerage side, we subsequently saw the broader Phoenix episode deteriorate across client accounts using recognizable versions of the system.
That distinction matters.
We are not claiming that every product ever called “Phoenix” failed, nor that every account using every version produced an identical result.
We are describing the historical Phoenix systems and accounts we observed.
The Real Lesson From Phoenix
It would be easy to summarize Phoenix by saying:
That is true but incomplete.
The more interesting lesson is how easily repeated survival can be mistaken for evidence of robustness.
Consider what the trader experienced:
Large drawdown → recovery → another drawdown → another recovery
How large the next adverse move can become before the account can no longer survive it
Each successful escape added another apparent data point in favor of the system.
But if the strategy depends on repeatedly adding exposure and waiting for a sufficiently large retracement, earlier recoveries do not place an upper bound on the next drawdown.
A strategy can therefore build a long and persuasive history before encountering the one sequence it cannot survive.
That is why Phoenix is such a useful example.
The system did not look dangerous because it failed constantly. It looked convincing because it survived repeatedly—until it didn't.
How to Evaluate an Expert Advisor Before Using It
Evaluating an Expert Advisor should begin with the strategy—not with the backtest return.
A polished equity curve, high win rate or impressive monthly return tells us what the system produced under a particular set of conditions. It does not tell us why it produced those results or what happens when those conditions change.
Before trusting an EA, we would want answers to the following questions.
Understand What the EA Actually Does
Start with the trading logic.
What causes it to enter a position?
What causes it to exit?
Does it use a stop-loss?
Can it hold multiple positions simultaneously?
Does it add to losing trades?
Does position size change after losses?
Is there a maximum number of trades or a maximum amount of exposure?
You do not necessarily need to understand every line of MQL code. But you should understand the economic behavior of the system well enough to explain what it will do when the market moves sharply against it.
If the only explanation is that the algorithm is “proprietary,” “AI-powered” or too sophisticated to explain, that leaves a major part of the risk assessment unanswered.
Examine Position Sizing Before Looking at Win Rate
For many of the problematic EAs we observed at brokerages, the most important feature was not the entry signal.
It was what happened after the first trade lost.
An EA that opens a 0.1-lot position and stops out is fundamentally different from one that responds to the same loss by opening 0.2, then 0.4, then 0.8 lots.
So do not ask only:
How often does it win?
Ask:
How does exposure change when it loses?
Look specifically for:
Martingale-style lot increases;
grid entries against an adverse move;
averaging down or averaging up;
unlimited or unusually deep trade baskets;
absence of a hard loss limit;
position sizing tied to account balance without a realistic cap.
A high historical win rate can coexist with a very unfavorable loss distribution.
Look at Equity, Not Just Balance
A steadily rising balance curve can be misleading if large losing positions remain open.
For any EA that holds multiple positions or averages into trades, examine:
maximum equity drawdown;
maximum balance drawdown;
the gap between the two;
largest floating loss;
maximum simultaneous exposure;
longest period spent in drawdown.
If balance drawdown looks small while equity drawdown is much larger, ask why.
That difference may indicate that losses were being carried rather than realized.
This is particularly important for grid and Martingale-style systems, where apparently smooth realized returns can coexist with substantial floating risk.
Separate Optimization From Evaluation
An optimized backtest is not the same thing as an independent test.
MetaTrader allows developers to run an EA repeatedly using different parameter combinations and select those producing the strongest historical results. It also provides forward testing, where selected parameters are subsequently evaluated on a separate portion of the data specifically to help detect over-optimization. (metatrader5.com)
So ask:
Was the same historical period used to design, optimize and evaluate the strategy?
If so, the evidence is much weaker.
A more convincing process looks like:
Develop → optimize → freeze the rules → test on unseen data
Even then, good forward performance does not prove that the strategy will remain profitable indefinitely. It simply gives us stronger evidence than a result produced on the same data used to develop the system.
Check How Realistic the Backtest Was
Not all MetaTrader backtests use the same assumptions.
The Strategy Tester can use real historical ticks, model execution delays and simulate different trading conditions. MetaTrader also provides a No Delay mode in which orders are filled at requested prices without requotes—conditions the platform itself describes as ideal. (metatrader5.com)
That matters.
Ask whether the test included realistic:
spreads;
commissions;
swaps where relevant;
execution delays;
slippage;
tick-level data;
margin requirements.
The shorter the holding period and the smaller the expected profit per trade, the more these assumptions tend to matter.
A strategy earning three pips on an average trade has much less room for execution friction than one targeting movements of several hundred pips.
Test Whether the Result Is Robust
A durable strategy should not normally require one magical parameter combination to survive.
Suppose an EA looks exceptional with:
47-period lookback + 23-pip grid spacing + 1.37 lot multiplier
but performs poorly when those settings are changed slightly.
That should raise questions.
Robustness testing asks whether broadly similar parameter choices produce broadly similar behavior.
You can also test the strategy across:
different historical periods;
different volatility environments;
different currency pairs where the economic logic should apply;
somewhat higher trading costs;
execution delays;
alternative but reasonable parameter settings.
The purpose is not to find another perfect configuration.
It is to determine whether the strategy’s apparent edge survives reasonable changes in assumptions.
Treat AI-Generated Trading Code Like Any Other Unverified Code
The arrival of coding agents adds one more question:
Who actually understands the program?
If an LLM generated the EA, obtaining working code is only the beginning.
The user should still be able to verify that:
the program implements the intended entry and exit rules;
position sizing works as expected;
stop-loss and risk limits cannot be bypassed unexpectedly;
trade-management logic behaves correctly when several positions are open;
the backtest does not accidentally use unavailable future information;
errors and unusual market conditions are handled sensibly.
This is especially important when the person prompting the AI cannot independently read the code it produces.
The software may compile.
It may run.
It may even generate a spectacular backtest.
None of those things establish that the underlying trading hypothesis is valid.
Regulators have already warned about automated systems marketed with claims that AI can generate extraordinary or guaranteed returns. The CFTC’s guidance is straightforward: AI does not give a trading bot the ability to predict the future or sudden market changes, and costs such as fees and spreads still need to be considered. (cftc.gov)
The label has changed.
The validation problem has not.
What Would Make an EA More Convincing?
No single statistic can establish that an Expert Advisor has a durable edge.
But we would place considerably more weight on a system where:
the trading logic can be explained clearly;
the risk and position-sizing rules are explicit;
losses are not simply hidden in growing floating positions;
backtest assumptions are realistic;
optimization and evaluation periods are separated;
results remain reasonably stable across alternative parameters;
meaningful out-of-sample or forward results exist;
trading costs do not eliminate the apparent advantage;
the developer is transparent about drawdowns and failure conditions, not just returns.
None of these guarantees future profitability.
What they do is make it harder for overfitting, hidden leverage or deferred losses to masquerade as a trading edge.
That is ultimately the standard we would apply to an old MetaTrader Expert Advisor, a modern trading bot or an AI-generated strategy alike:
Understand the logic. Understand the risk. Then test whether the evidence survives conditions the strategy was not designed around.
The Bottom Line: Expert Advisors Are Tools, Not Trading Edges
So, are Expert Advisors bad?
No.
An Expert Advisor is software. It automates a set of trading rules. Whether the resulting system is useful, fragile or outright dangerous depends on what has been programmed into it and how convincingly that strategy has been tested.
That distinction is important because the original version of this article was much more categorical. Years of seeing problematic EAs inside retail brokerages made me deeply skeptical of them—and for good reason. We repeatedly encountered systems that looked remarkably successful until their underlying risk finally surfaced.
But those observations do not prove that automated trading itself does not work.
What they show much more clearly is how certain strategies can create misleading impressions of reliability.
In our historical brokerage records, many of the problematic EAs used grid trading, Martingale-style position sizing or some combination of the two. They could produce high win rates, smooth increases in realized balance and repeated recoveries from drawdowns.
Those characteristics looked reassuring.
Sometimes they were exactly what made the eventual failure so difficult to anticipate.
The Phoenix case illustrates the problem particularly well. The system survived several severe adverse moves. Each recovery provided another reason to believe the mechanism worked.
Until the account encountered a move it could no longer survive.
That is the broader lesson worth carrying forward.
A trading system should not be judged by how convincing its best historical period looks. It should be judged by whether we can understand its logic, quantify its risk and find evidence that its apparent edge survives tests it was not specifically designed to pass.
That means looking beyond:
headline returns;
win rates;
optimized backtests;
smooth balance curves;
impressive screenshots;
claims that the algorithm is proprietary or powered by AI.
Instead, ask how exposure changes during losses. Examine equity as well as balance. Separate optimization from evaluation. Use unseen data. Include realistic trading costs and execution assumptions. Test whether performance survives reasonable changes in parameters and market conditions.
And understand what the program is actually doing with your money.
AI Changes the Tools, Not the Standard of Evidence
The same principle applies to the new generation of AI trading bots.
Large language models and coding agents have made it dramatically easier to turn a trading idea into executable software. That is a genuine technological advance.
But generating code is not the same thing as discovering a profitable trading strategy.
A trader can now automate an idea without necessarily understanding the code produced, the assumptions embedded in it or the statistical evidence required to determine whether the idea contains an edge.
In that respect, the old Expert Advisor era and today’s AI trading boom have more in common than their marketing suggests.
The terminology changes.
The technology improves.
The fundamental questions remain:
What is the strategy actually doing?
Where is the supposed edge?
How was it tested?
What happens when it is wrong?
And how much can the account lose before the strategy gets another chance to be right?
An EA can automate a sound strategy just as efficiently as it can automate a terrible one.
AI can generate sophisticated code just as easily as it can generate sophisticated-looking nonsense.
Automation provides consistency and speed.
It does not provide validity.
That is why we would not ask whether Expert Advisors—or AI trading bots—“work” as a category.
The better question is:
Does this particular strategy have a defensible edge, and is the evidence strong enough to justify putting real capital behind it?
Everything else is implementation.
Disclaimer: This article is for educational purposes only and does not constitute financial or investment advice. Trading forex, cryptocurrencies, and other financial instruments involves a high level of risk and may not be suitable for all investors. Always conduct your own due diligence and consult with a licensed financial advisor before making any investment decisions.
Frequently Asked Questions
Common questions about Expert Advisors, automated forex trading and evaluating trading bots.
What is an Expert Advisor in forex?
An Expert Advisor (EA) is a program used with MetaTrader to automate trading-related tasks. Depending on its programming and permissions, an EA can analyze market conditions, generate signals, open and close positions, and manage existing trades according to predefined rules.
The EA itself is not the trading edge. It is the software that implements a strategy. Whether that strategy is useful depends on its underlying logic, risk management and evidence.
Can Expert Advisors actually make money?
An Expert Advisor can implement a profitable strategy, but automation does not make a strategy profitable by itself.
A convincing evaluation requires more than an attractive historical return. The strategy should be tested on data that was not used to optimize it, under realistic trading costs and execution assumptions, and across conditions different from those used to develop it.
Past profitability—whether from a backtest or a live account—does not guarantee that the same relationship will persist.
What is the difference between an Expert Advisor and an AI trading bot?
Expert Advisor is MetaTrader terminology for an automated trading program. An EA can be completely rule-based and contain no artificial intelligence at all.
An AI trading bot may use machine learning or a large language model as part of its analysis or decision process. Alternatively, AI may simply have been used to generate the program's code while the actual trading logic remains rule-based.
Either way, adding AI does not establish that the strategy has a statistical edge. The CFTC has specifically warned against claims that AI-powered trading systems can produce unrealistic or guaranteed returns.
Is a high win rate a sign that an Expert Advisor is safe?
No. Win rate tells us how frequently trades or trading cycles end profitably. It does not tell us how much the strategy loses when it is wrong.
This is particularly important for grid and Martingale-style systems. An EA can produce many small winning cycles while accumulating much larger floating losses during adverse market moves.
Evaluate the size of wins and losses, maximum equity drawdown, simultaneous exposure, position-sizing rules and the gap between balance and equity—not the win rate in isolation.
How should an Expert Advisor be tested?
Start by defining exactly what the strategy does and how it manages risk. Then test it historically using realistic spreads, costs and execution assumptions.
Optimization and evaluation should be separated. MetaTrader provides forward testing specifically so selected optimization results can be tested on a different historical period, helping reveal strategies that were fitted too closely to the original sample.
A stronger evaluation also examines alternative parameters, different market periods, equity drawdowns and performance under less favorable assumptions. A backtest should be the beginning of the investigation, not proof that the EA will remain profitable.
References
Selected platform documentation, regulatory guidance and academic research used in this article.
- MetaQuotes Platform documentation
Expert Advisors and Custom Indicators
Official MetaTrader 5 documentation on Expert Advisors, automated trading permissions and trading robots.
- MQL5 Reference Platform documentation
Trade Request Structure (MqlTradeRequest)
Documents the order-comment field and the Expert Advisor magic identifier used to identify and organize automated trading orders.
- MetaQuotes Platform documentation
Official documentation covering historical EA testing, real tick data, execution settings, commissions, margin assumptions and forward-test periods.
- MetaQuotes Platform documentation
Official documentation on testing multiple parameter combinations and comparing optimization results with forward performance.
- MetaQuotes Platform documentation
MetaTrader documentation defining equity, margin, free margin and margin level, including the relationship Equity ÷ Margin × 100.
- Bailey, Borwein, López de Prado & Zhu Academic research
The Probability of Backtest Overfitting
Journal of Computational Finance, 2017. Develops a framework for estimating the probability that an investment backtest has been overfit through repeated strategy selection.
- Bailey & López de Prado Academic research
The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality
Journal of Portfolio Management, 2014, 40(5), 94–107. Examines performance inflation caused by multiple testing and strategy selection.
- Cheevirot, Smanchat & Nuchitprasitchai Academic research
Algorithms, 2026, 19(6), 442. Distinguishes grid execution from Martingale position scaling and illustrates the potentially large divergence between balance and equity drawdown.
- U.S. Commodity Futures Trading Commission Regulatory guidance
AI Won't Turn Trading Bots into Money Machines
CFTC Customer Advisory, January 2024. Warns against claims that AI-powered trading algorithms or bots can produce unrealistic or guaranteed returns and emphasizes trading costs and underlying risks.
This is a selected reference list rather than an exhaustive bibliography. First-hand brokerage observations and the historical Expert Advisor database discussed in the article are original Finansified material and are identified as such in the text.
Ready to analyze the macro behind your pairs like institutions do?
Explore how inflation, growth, policy rates, liquidity, confidence, and Taylor Rule benchmarks shape economies behind FX pairs and influence their behaviour. The BETA is live NOW!



