The robot is not sold for what it earns, but for what it shows

The product page always carries the same things: an equity curve climbing with barely a dip, a high win rate, and a number that looks like an exam grade — modelling quality 90%.

That report is not proof the robot makes money. It is the output of a simulator, and it is worth knowing what the simulator does inside before giving it the weight of an audit.

What follows is not about whether a trading bot can work — it can — but about how the document they sell it with is built.

The report does not measure the robot, it measures the simulation

MetaTrader 4 offers three modes for testing a strategy, and what separates them is which data the movements inside each candle are invented from. The platform's official help describes them like this:

The first is good for almost nothing. Of the second, the documentation itself warns that its results "can also be considered as estimated ones, not as final ones".

And the third — the most precise of the three, the one that produces the 90% on the product page — is described with this sentence:

"To generate price movements between control points, interpolation based on predefined templates is also used."

Read it again. In the most faithful mode available, the ticks in the backtest are generated by interpolation from templates. They are not the market that happened: they are a plausible reconstruction of what might have happened inside each candle.

For a strategy that enters and exits over hours, that matters little. For one that trades inside the candle — scalping, grids, anything with a stop a few pips away — the gap between the real tick and the interpolated tick is exactly where its result lives.

What does the 90% actually mean?

This is the most expensive misunderstanding of the lot, because the number sounds like a quality grade for the strategy and it is not.

Modelling quality measures the density of historical data available to reconstruct the period, not the soundness of the system. A 90% says the simulator had good one-minute data to work with. It says nothing about whether the logic tested on top of it has an edge.

An overfitted robot with 90% modelling quality is still an overfitted robot. The report will be beautifully built and will still describe an edge that does not exist.

The two costs the simulator never saw

A backtest runs on fixed assumptions. Reality does not.

The spread. The test normally runs at a constant spread, and the spread is not constant: it widens overnight, on thin pairs, and in the seconds around a data release. We measured it in why the spread widens overnight. A strategy whose average profit per trade is a few pips can go from profitable to losing on that line alone.

Execution. In the simulator the order fills at the requested price, every time. In the real account there is slippage, there are rejections and there is latency, and the sign is not neutral: it tends to go against you precisely when the market moves fast, which tends to be when the strategy trades.

Neither shows up in the equity curve you are shown. Both show up in your account.

The seller's question

If the system has a demonstrable edge, why is it for sale?

The honest answer is that there can be legitimate reasons: a developer without the capital to exploit it, or a strategy whose capacity is larger than their account. Those exist, and it would be unfair to deny it.

But look at the arithmetic on the other side. A real edge applied to your own capital grows with that capital and is not consumed by use. A licence sold a thousand times is consumed: if the strategy depends on getting filled at specific levels, a thousand accounts doing the same thing at once degrade the result for all of them. The seller is paid once per copy and does not share that decay.

Which is why the incentives line up closer than anyone would like to the ones described in copy trading and in fake signal services: the seller is paid for the sale, not for your result.

What a serious report does carry

Not all of it is worthless. There is a short list of things an honest report carries and a marketing one does not:

• Out-of-sample results. The period the strategy was tuned on and the period it was validated on, separate and both published. Without that, the curve only demonstrates overfitting.
• Forward performance, dated after the system was published. That is what forward testing measures and the only thing that cannot be manufactured backwards.
• Maximum drawdown, not just profit. A curve without its worst streak is incomplete.
• Metrics beyond win rate — expectancy, and the rest of the ones in expectancy, Sharpe and Sortino.
• Declared assumptions: spread used, modelling mode, and the broker the data came from.

A seller who has all five shows them, because they are the best argument available. One who only has the curve shows the curve.

What this article does not tell you

It does not say robots do not work. Automated trading is a real tool, and the difference between an algorithm and something billed as AI is covered separately.

It does not say every seller lies, or that a backtest is worthless. A well-built backtest discards bad strategies quickly, which is a lot: what it cannot do is prove one is good. We developed that in what backtesting is.

It does not rate specific products or recommend any, and it does not claim the 90% is a fake number. It is a real number that measures something other than what people think it measures.

What it gives you is what to ask for before paying: the out-of-sample period, the forward record, and the spread assumptions. If those three answers do not arrive, the curve you were shown describes an interpolated past, and you are buying it as if it were the future.