AI trading bots and ML EAs explained simply

On this page

In short

An AI trading bot is an EA whose rules were not written by hand but learned from data. A model looks at years of prices and learns which patterns came before a move. Live it applies these patterns.

The schematic shows this as simply as possible: from the training data the model has learned an expectation (the line) and buys when price sits far below it. As long as the market behaves as in training, that works. After a regime change the model keeps buying, stop after stop, because it has learned nothing else.

Then there is the marketing: many products call themselves AI without containing a learning model. Our AI trading bot review shows what such EAs deliver live. Why good backtests often do not hold up live is explained in Why EAs fail live.

The strategy as a schematic

A drawn price path, no market data. First the case where the rule works, then the same routine when it fails.

Closed: ±0 pips
Open: –
Open volume: –

Schematic, no market data: a drawn price path, results in pips weighted to 0.10 lots. Not a backtest and no statement about any real EA.

How the strategy trades

A model is trained on historical price data, for example a neural network or a decision tree. It learns which patterns came before a move in that period.
After training the model is frozen. Live it usually no longer changes; it only applies what it learned.
On every new bar the model calculates an expectation or a signal.
If price deviates strongly from that expectation, the EA trades, usually with a stop and a target like any other EA.
If market behaviour changes, the learned patterns no longer fit. Without retraining the model keeps trading by the old rule.

Where the risk sits

A model can only know what was in its training data. If the market switches to another regime (trend turns sideways, calm turns volatile, a central bank changes course), the learned patterns no longer fit. The model does not notice; it simply keeps trading.

The more parameters a model has, the more easily it fits random patterns of the past (overfitting). The backtest then looks excellent, and the live curve breaks down. An AI EA therefore needs a long live record, not a nicer backtest.

In the EA market AI is above all a sales argument. Many products with AI in their name use classic indicators, a grid or martingale, and no verifiable model. What counts is not the label but how the EA trades: stop or no stop, how volume grows, how it loses.

When it works, when it fails

Tends to work when …

The market behaves live much as it did in the training data.
The model is regularly checked against new data and retrained.
Fixed stops and risk limits that apply independently of the model.

Tends to fail when …

A regime change: the model keeps trading patterns that no longer hold.
Overfitting: the model learned noise, not structure.
The AI label hides a grid or martingale without a stop.

How to recognise it in an EA

Traces in the trade log of a live account, whatever the vendor calls the EA.

A product description with terms like neural network, deep learning or GPT, but no verifiable details on model, data and training.
A trade log that looks like a classic EA: trend following, grid or mean reversion.
Very good backtests over exactly the period the model was trained on.
A live curve that looks clearly different after a market change than before.

Typical settings

Setting	Typical	Effect
Training period	e.g. the last 2–10 years	What did not occur there, the model does not know.
Retraining	never, monthly or continuous	Without retraining the model goes stale.
Signal threshold	minimum confidence or deviation	Higher = fewer trades.
Stop / risk limit	fixed per trade or per day	Should apply independently of the model.

Common ranges from EA manuals and set files, not a recommendation.

EAs in our tests built this way

8 graded EAs from our tests carry this design. Grade on a 0 to 10 scale, one formula for all; no live figures.

All EA reviews →How we grade EAs →

All strategies →All designs compared: EA strategies explained →

Frequently asked questions

Do AI trading bots work?

Some models work in certain market phases, none in all of them. Whether a specific AI EA is any good is only shown by a long live record across different market phases. The AI label alone says nothing about it.

What is the difference between AI and machine learning in EAs?

In the EA market both terms are used almost interchangeably. Machine learning means models that learn rules from data, such as decision trees or neural networks. AI is the broader umbrella term and is often used for products that contain no learning model at all.

What is a regime change?

A lasting change in how the market behaves, for instance from a calm trend to a volatile sideways phase. Rules that worked in the old regime then no longer fit. For learning models this is the central risk.

How do I spot exaggerated AI marketing?

By promises of high profits, missing details on model and training data, backtests only, and a short or missing live account. A serious vendor explains what the model does and where it fails.

Video: the strategy in under a minute

The same animation as a video with subtitles, for those who prefer to watch.

AI trading bots explained simply