Black Friday: all 12 tools. Until 1 December.Black Friday $606.98 one by one

$606.98 one by one $49 once Save $557.98Get the SuiteGet it
ProEA Lab · MT5, TradingView and trading research · Evidence standards
Systems

Do Forex EAs Actually Work? Grid Risk, AI Bots and a Pre-Buy Test

An EA automates a set of rules; it doesn't supply an edge, and no representative dataset says what share of EAs work. Here is why EAs fail, why grid and martingale robots look safest just before they aren't, five checks for an AI trading bot, and a four-step test to run before you pay.

PL · 13 min readUpdated
A trader climbing a glowing emerald staircase of blocks and gold coins that ends at a cliff edge where the steps crumble into the void, with looming robots behind.

An EA can only work as well as the rules it runs. It executes them without fatigue or hesitation, but whether those rules make money after spread, slippage and swap is a separate question that automation doesn't answer. No representative dataset we know of says what share of forex EAs work, so the useful answer to "do forex EAs actually work?" is a test you run on the one in front of you.

Two kinds need extra suspicion: grid and martingale robots, whose smooth curves hide the loss they are built around, and "AI trading bots", where the word AI often stands in for evidence. Both get a section below. We sell one EA ourselves, and it is a grid system; that section says so plainly.

An EA doesn't add an edge. It automates whatever you hand it, edge or story.

Is there data on how many forex EAs work?

No. We know of no regulator or study that follows a representative sample of forex EAs or their users, so any "most robots fail" or "X% of EAs are profitable" figure is a guess. What does exist is loss data on retail trading populations:

Source, checked 2026-10-11PopulationFinding
ESMA, 27 March 2018Retail CFD accounts, from EU national regulators' analyses of CFD trading74–89% of retail accounts typically lose money
SEBI, 23 September 2024Individual traders in Indian equity futures and options, FY22 to FY2493% incurred losses; aggregate losses above ₹1.8 lakh crore over three years
Chague, De-Losso and Giovannetti, FGV working paperIndividuals who began day trading Brazilian equity futures in 2013–2015 and persisted for at least 300 days97% lost money

None of these separates EA users from discretionary traders, and none measures forex robots. They set a base rate for caution, not an EA failure rate.

One practical fact from the CFTC's August 2022 advisory on OTC forex (checked 2026-10-11): in the US, registered OTC forex dealers must give any customer or prospective customer the percentage of profitable and unprofitable accounts on request. Ask before you deposit.

The sharper question: what evidence supports this EA after realistic costs, and what would break it?

Why do forex EAs fail?

These are failure mechanisms to inspect, not ranked causes from a survey.

The backtest was fiction. Try enough settings and something will always look incredible; the winner gets published and the failed versions disappear. A backtest can also read data it wouldn't have had live (look-ahead bias), and even an honest one is one path at modelled costs. If the seller can't say how many versions failed, the winner means less than it looks.

Costs ate the edge. EAs that trade often for small gains are the most exposed to spread, commission, slippage and swap, and those costs are not fixed: they tend to widen exactly when markets move fast. An edge smaller than real friction isn't an edge.

The regime left. Every EA is built for a condition: trend, range, a volatility band, a session. When a trend ends or a range breaks, the EA hasn't changed; the market has. Owners often misread it as the EA breaking, and the post-regime audit for gold EAs works through it with gold's January 2026 high.

The sizing was a time bomb. Grid, martingale and recovery logic turn many small losses into one delayed large one. That gets its own section below.

The owner overrode it. Stops widened in a drawdown, risk doubled after a loss, a pause after a bad week and a restart at the worst moment. Automation doesn't remove the owner, so stop rules written before the drawdown are part of the system.

Can an EA be profitable?

It can, if it automates a real edge and executes it consistently. A serious EA needs:

  • a reason the edge should exist
  • a backtest that didn't just fit noise, and survival out of sample
  • realistic costs
  • position sizing that survives bad sequences
  • drawdown behaviour written down before the drawdown
  • a way to stand aside when the regime changes
  • rules or source the operator can inspect

Notice what isn't on the list: a secret indicator, an "AI brain", a near-perfect win rate, a vendor screenshot, a monthly return.

That last one is worth a minute of arithmetic. A steady 10% a month compounds to 1.10¹² ≈ 3.14 times the starting balance, about +214% a year. A steady 15% a month compounds to about +435% a year. The CFTC's advisory on AI trading bots (checked 2026-10-11) describes one such pitch: in its Mirror Trading International case study, a pool "guaranteed at least a 10 percent monthly return (or more than 200 percent per year) trading foreign currencies." A system can make 10% in one month by taking enough risk. Selling it as a steady rate is the red flag.

If a robot reliably compounded 10% a month at survivable risk, its owner wouldn't need your purchase.

Are grid and martingale EAs safe?

Not by default. A grid EA adds positions as price moves against its open basket. A martingale increases size after a loss. Recovery logic does some of both, aiming to close a losing cluster at a net profit. All three turn many small losses into fewer, later, larger ones. So is a grid trading EA safe? Only if its total exposure is capped and it accepts losses before they threaten the account, and you can confirm that in the rules, not on the curve.

In calm, mean-reverting markets price comes back, baskets close green and the curve climbs in neat steps. In a sustained one-way move it doesn't come back soon enough: each add improves the average entry and raises the floating loss until margin runs out. The account is finite; the adverse move doesn't have to be. In martingale form the size doubles at each step (1, 2, 4, 8, 16) until the next step is larger than the account can carry. That isn't bad luck. It's arithmetic.

What the curve showsWhat the structure may hide
A smooth staircase upRisk stored in open trades, not closed
Months of small winsA payoff of many small wins and one rare large loss
"No losing trades"Losses deferred into an open basket
A tiny drawdown on the balance lineFloating loss that hasn't been realised yet
A very high win rateNegative skew, not safety

The balance curve and the floating loss

A balance curve moves only when trades close. A grid's losing positions stay open, so their loss sits in equity, off the balance line, sometimes for weeks. A grid can show a smooth balance staircase while its equity is deep underwater; the cliff reaches the balance line only when the basket is finally closed or liquidated. Ask for the equity curve and the maximum floating loss, not just the balance curve.

A one-way downward price trend with grid entries stacking against it; the floating-loss area grows taller at each level until it reaches a red margin-call wall.
The grid answers a trend by adding exposure. The account has a wall; price does not.

Why tuning the backtest makes it worse

Widen the grid, add levels or delay the cutoff, and the backtest improves because the worst period in the sample is now survived. That usually pushes the failure point just past the worst move in your data, so the live failure needs a bigger move and costs more. Test the move that wasn't in the backtest: a longer one-way trend than the instrument has shown, a gap through several levels, a spread spike at full exposure.

"But mine has a stop"

A real control defines the loss before the account is threatened: a hard stop per position or basket, capped total exposure, a maximum basket size, no unbounded lot increase after a loss. A cosmetic control, like an account-level stop at a catastrophic drawdown, fires after most of the damage; it turns a wipe-out into a large realised loss, which isn't bounded risk. "Recovery mode" is often martingale with better branding: a risk increase, not risk removal.

Four questions to read off the rules

  1. Does it add to losing positions? If yes, treat it as negative-skew exposure until it's shown to be bounded.
  2. Does position size grow after a loss? If yes, martingale or recovery sizing is in there, and the burden of proof rises.
  3. Where is the hard stop? A fixed stop on every position or basket, set before the trade, not an emergency exit after the damage.
  4. What caps total exposure? Maximum open positions, maximum lots, maximum basket risk.

A seller who won't let you check has answered. The average down calculator shows how the loss at a stop grows with each add, and the forex margin calculator shows the margin each stacked position ties up.

Is an AI trading bot a scam? Five checks.

Not always, but "AI" proves nothing. Real tools use machine learning; the word still tells you nothing about whether there's an edge after costs, whether the track record is real, or whether you can get your money back.

Regulators have said so directly. The CFTC's advisory, AI Won't Turn Trading Bots into Money Machines (checked 2026-10-11), states that "AI technology can't predict the future or sudden market changes," and warns about bots, "trade signal strategies, and crypto-asset trading schemes that promise unreasonably high or guaranteed returns." The FBI's IC3 2025 annual report (checked 2026-10-11) puts losses in investment complaints with a reported AI nexus above $632 million, against more than $8 billion lost to investment scams overall, and describes investment clubs using AI-generated videos and voices of celebrities, CEOs and trusted figures.

The fakes usually wear one of four faces:

  • The deepfake endorsement: a familiar face, an "AI trading platform", a link to a deposit page.
  • The black-box bot with a leaderboard: a curve that only climbs, no logic you can read, no risk you can set.
  • The manufactured track record: withdrawal screenshots, five-star walls and "verified" badges, all cheap to generate.
  • The "AI signal" group: a channel that celebrates the wins and deletes the losses.

Five checks before any money moves:

  1. Strip the face off. Ignore the celebrity, the video and the testimonials, and assume each can be generated. What checkable claim is left? If it's "a big return and a deposit button", stop.
  2. Demand the mechanism. Ask for the entry and exit logic, the risk controls you can set, the cost assumptions and the failure modes. "Proprietary AI" isn't an answer.
  3. Reproduce, don't admire. Can you run it on your own demo account, at your broker's costs, and see the losing periods? A record you can only look at is decoration.
  4. Find the people. Look for a named team you can find outside the funnel, and a registration you can check if they handle client money. The CFTC also suggests a reverse image search on key personnel and checking how old the website's domain is.
  5. Test the exit before you enter. Does your money stay in your own broker account, or move to their platform? If it moves, find out exactly how withdrawals work first. In the crypto investment scams the IC3 report describes, victims who try to withdraw "will be charged taxes and fees" before the scammers disappear with the funds.
The anatomy of an AI investment scam as a one-way funnel: a deepfake celebrity ad, then an urgent direct message, then a fake app showing a climbing dashboard, then a 'pay a fee to withdraw' wall, then the money is gone.
Every stage is built to get money in. None is built to let it out.

"Not a scam" and "has an edge" are two different bars: what a genuine AI tool builds still has to pass the same cost and sizing tests as any EA. For the useful side, see how to use AI for trading, how to build a trading bot with AI and the risks of AI trading agents with broker access.

Are marketplace ratings and verified track records proof?

No. Buyers review in the first weeks, inside whatever regime the EA happens to suit, so a grid in a friendly market can collect five-star reviews for months before its first real trend. Survivorship does the rest: dead systems get delisted or relaunched as a new version, so the list you browse is the list that hasn't failed publicly yet.

A verified track record shows trades in one account, over one window, chosen by the person showing it. Check the dates, deposits and withdrawals, whether other accounts were left out, and whether it shows equity or only balance. A rating says people liked it. It doesn't say the math survives.

How do you test an EA before risking real money?

Four steps, on demo or in the strategy tester, before any live money.

1. Read it, or demand the answers. With source, find the lines for direction, position sizing, order sending, stops, the drawdown rule and the kill switch. Without source, ask what the EA does when its favourite market disappears; for a gold EA, what it did after gold's January 2026 high. Evidence that stops before the hard window is an answer.

2. Re-test it where it wasn't born. Run different years, the recent hostile months and the high-spread periods, not only the seller's window (how to backtest an EA in MT5, and why every tick and real ticks differ). Then vary the order of outcomes: the MT5 report analyzer reads your own MT5 report in the browser and can resample its recorded results, while the Monte Carlo simulator draws synthetic paths from a win rate and payoff you enter, not from your trades.

3. Cost it like the worst day. Re-run with spread doubled, slippage on every entry, commission and swap. If doubling the spread kills it, the edge was thinner than the friction.

4. Size it for the sequence you haven't seen. Take the worst drawdown in testing and double it. At your intended lot size, does the account survive? If not, the EA isn't "risky"; it's too large for the account. If it adds to losers, repeat the test with a trend that doesn't stop. The risk of ruin calculator estimates the chance of reaching a loss floor from a win rate, payoff and risk per trade.

Passing all four guarantees nothing. It earns a small, pre-sized forward test with review points written down in advance; there is no universal answer to how many trades that takes.

Should you buy an EA, rent signals or build your own?

Renting signals gets you decisions without the reasons. When results turn, you can't tell a normal losing streak from a broken system, and when the service stops, nothing stays with you.

Buying a closed EA gets you automation without inspection. You can't know whether it's a working system or a screenshot with code attached until the market tests it.

Building with AI is now realistic, but the code was never the hard part; the edge is. With a validation process AI is an accelerator; on vibes it overfits faster than anything before it.

Buying source is the one version of buying where all four test steps are possible. That's inspectability, not profitability: bad code can be open too.

Five questions show what you're actually getting:

  1. Can I read the logic, not just run it?
  2. Can I change what matters: sizing, risk caps, sessions, spread guards?
  3. Can I audit the assumptions behind the backtest and the costs?
  4. When it fails, can I diagnose why?
  5. What is the seller paid to preserve: my subscription, or the product's reputation under inspection?

So, are forex EAs worth it?

Are forex EAs worth it? Not on the strength of a screenshot, a rating or a win rate. One can be worth considering when you can inspect its rules, re-test it at your broker's costs, size it for a drawdown worse than its backtest, and watch it while it runs. That's a due-diligence standard, not a promise. An EA you can't verify is still an unverified bet.

Disclosure

We build and sell MTR, a hedge-grid + recovery EA for gold, and other source-code trading tools. Nothing here is investment advice, and no EA on this page, ours included, has a guaranteed result. The loss figures above describe the populations their sources measured; none of them predicts yours.

Frequently asked questions

Are forex robots a scam?

Not all of them, but many are sold on claims no robot can back up: a steady monthly return, a near-perfect win rate, a famous face. A legitimate EA lets you see what it does, run it on your own demo account at your broker's costs, and keep your money in your own broker account. Treat a product that refuses those checks, or asks for a fee to release a withdrawal, as a scam until it proves otherwise.

Can an EA really make 10% a month?

A system can make 10% in one month by taking enough risk. As a steady rate, 10% a month compounds to about 214% a year and 15% a month to about 435%. The CFTC cites a forex scheme that guaranteed at least 10% a month. Treat any steady monthly return as a red flag unless the seller shows full risk, drawdown and failure history.

What is the difference between a grid EA and a martingale EA?

A grid EA opens extra positions at set intervals as price moves against its open trades, which improves the average entry. A martingale EA increases position size after a loss, so a smaller reversal recovers the sequence. Many recovery EAs combine the two. Both turn many small losses into fewer, later, larger ones, so check whether total exposure is capped and where the hard stop sits.

Are free forex EAs worth trying?

As something to learn from, sometimes. Free doesn't exempt an EA from spread, slippage, regime change or bad sizing, and a free EA can be a marketing funnel, an abandoned experiment or a recovery system with no hard stop. Run the same four-step test on a demo account before it touches real money.

Do prop firms allow EAs?

It depends on the firm and its current rules, so don't assume a universal ban or approval. FTMO, for example, allows EAs, but warns that a third-party EA running the same strategy for other traders risks being denied a funded account under its per-strategy capital cap. A grid or recovery EA can also breach a daily loss limit that counts floating losses while its basket is still open. Read the firm's rules before paying the fee.

How much money do you need to run an EA?

Enough that the worst tested drawdown, doubled, doesn't force you to stop at the bottom, at the smallest lot size your broker allows. For a grid or recovery EA that also means margin for every add it can make. If surviving that math needs money you can't afford to lose, the account is too small for the EA.

Why did my EA work for months and then stop?

Usually because the market left the conditions the EA was built for: a trend EA struggles in chop, a range EA in a trend, and a grid when price stops coming back. Before deleting it or adding size, check how it behaved in the hostile stretch: did it pause, cut risk or keep doing the same thing?

Is a verified Myfxbook record proof that an EA works?

It shows that trades happened in one account over one window, chosen by the person showing it. It doesn't show the strategy surviving conditions it hasn't met, and a balance curve can hide open losses. Check the dates, deposits and withdrawals, any accounts left out and the equity curve, then run your own tests.

Continue reading

AI Found a Winner.

Test ten thousand strategies and one will look brilliant by pure luck. AI tests ten thousand before lunch. Here is the test quant funds run — and retail skips — that tells a real edge from a coincidence.

Jun 3 · ProEA Lab

A small child reaches up toward a single dazzling green crystal of light, surrounded by towering green-eyed machines and their reaching hands — the one alluring winner that lures you out of thousands of tries.

What Is Look-Ahead Bias? How to Find It in a Backtest

Look-ahead bias is a backtest using information that didn't exist yet when each trade was decided. The five places it hides, how to trace each one, and why a one-bar delay test can point at a leak but can't prove one.

Jun 3 · ProEA Lab

A child floats at the center of a vast web of green light and broken clock faces, ringed by towering green-eyed machines, a cracked hourglass and candlestick charts: the trader caught inside a backtest that can see the time the live market hides.

AI Backtest Overfitting: How to Audit a Perfect Result Before You Trust It

Describe a strategy to AI, or let an EA/indicator generator optimize it, and you'll get a gorgeous equity curve in minutes. You didn't find an edge — you ran a search. And a big enough search always finds a winner, even when there's nothing there.

Jun 4 · ProEA Lab

A child meditating cross-legged, calm at the center of a vast rainbow cosmic vortex spiraling into a black hole, ringed by drifting debris — stillness at the middle of the chaos of an endless search.