We read the sales page of 1,180 MetaTrader EAs. 74.6% describe a strategy with no cap on its losses.

Short version: we classified 1,180 expert advisors currently for sale by the loss mechanism their own sales copy describes. 880 of them — 74.6% — use martingale, grid, averaging, locking or loss-recovery logic without a hard stop loss. Those strategies have no upper bound on a single losing sequence. A sufficiently long one-directional move takes the account to zero. Not "might lose money": every strategy might lose money. Zero.

Why "no cap on losses" is a different question from "is it profitable"

Almost every argument about trading robots is about whether they make money. That question is hard, needs months of data, and the answer keeps changing. There is a much easier question underneath it, and it has a yes-or-no answer you can get from the product description alone:

Is there an upper bound on what one losing sequence can cost?

An EA in the second row can trade profitably for two years. Many do — that is exactly what makes them sell. The equity curve looks like a straight line at 45 degrees, because a strategy that never closes a loser has no losers on the chart. The losses are all sitting open, unrealised, invisible on the profit line right up until the margin call.

This is why "it has never blown up" tells you almost nothing. It has never blown up yet. June to August 2025 removed a batch of gold EAs that had run clean for years. The market was not unusual in some mysterious way. The mechanism simply came due.

How we counted

Every number below comes from text the vendors themselves published. We did not run these EAs, and this study makes no claim about whether any of them is profitable.

1. 1,180 products on sale were collected with their full descriptions and specification tables.
2. Each description was searched for terms describing the mechanisms above, plus their opposites — "hard stop loss", "fixed lot", "no martingale", "drawdown control".
3. A product was placed in unbounded if it describes at least one unbounded mechanism and does not describe a hard stop.
4. A product went to worth testing if it describes risk controls and no unbounded mechanism.
5. Everything too vague to call went to not enough information rather than being guessed at.

The one detail that changed the answer by 57 products

Our first pass matched the word "martingale" anywhere in the text. That flagged this sentence as a martingale EA:

No martingale and no grid are used.

It also flagged every vendor who wrote "we do not average down". Those are precisely the products most worth testing, and the naive count was throwing them out. The fix was to check for negation within the 40 characters before a match, and — this part matters — not to look across a sentence boundary, because this pattern is extremely common:

This is not a grid system. However, a martingale multiplier is applied after two consecutive losses.

Reading only the first sentence clears the product. Reading the whole paragraph as one string flags it for the wrong reason. Sentence-level negation moved the unbounded share from 64.7% down to 59.9% on an earlier sample, and the corrected classifier is what produced every figure in this article.

The counts

Mechanisms found, counted separately because one product often uses several:

The part we did not expect

Separately, we extracted every performance figure the vendors put in writing: win rates, maximum drawdown, monthly and annual return, and account-multiplication claims. 793 products — 67.2% — make at least one.

Then we crossed the two lists. The result is consistent enough that it is worth using as a filter:

The better the numbers on the page, the more likely the product is in the group with no floor.

There is no conspiracy needed to explain this. A 90%+ win rate is what a martingale grid produces by construction: it closes winners and holds losers, so nearly every closed trade is a winner. The claim is often not a lie. It is an accurate description of a statistic that has been chosen because it cannot show the risk.

The same logic applies to a low reported drawdown. If drawdown is measured on closed trades, a strategy that never closes a loser reports a drawdown near zero while carrying an open position that would wipe the account.

The multiplication claims, in the vendors' own words

These are quotations, not accusations. Each figure is on the vendor's own product page, and it may well have happened on some account at some point. What the figure does not tell you is how many accounts running the same settings did not survive to produce it.

What this study does not show

We want to be exact about the limits, because a study that overclaims is no better than a sales page that does.

• It is not a test. Nothing here was run. This is a classification of published descriptions.
• It cannot see undisclosed behaviour. A product whose description mentions no mechanism may still martingale in code. The 97 in "not enough information" are unresolved, not cleared.
• Unbounded does not mean worthless. A grid EA on a mean-reverting pair with a small, capped ladder and a portfolio-level stop is a legitimate design. Our classifier cannot see the cap, so it errs toward the harsher call.
• "Worth testing" means only that. 203 products passed the mechanism check. That says nothing about whether any of them makes money — which is the next thing we are doing.

What we do with this

We are not delisting the 880. People search for those products by name and deserve to find out how they work before they pay for one, and the mechanism is the single most useful thing to know. Each of those product pages states its mechanism plainly.

What they will never get is a place on a recommendation list, a spot in a ranking, or one of our monitored accounts. A test of an unbounded strategy has a known answer and only the date is unresolved.

The 203 in "worth testing" go into a backtest queue: real tick data, real spreads, vendor default settings with no optimisation, and a period that includes June to August 2025. Backtesting is used here to exclude, never to select — a good backtest does not mean an EA will survive, but a bad one settles the question. Results, including for the ones we expected to do well, get published on this site — the first run is here, and most of it is failures.

Every product page on this site carries the result of this classifier at the top: the tier, and the specific mechanisms found. Where we found none, the page says that too rather than implying safety. The classifier runs against the vendor's original description — not our rewrite of it — so that the verdict is always checkable against what they published themselves.