AI trading bots: separating the technique from the label

What AI actually does in trading

What it is genuinely good at: classification and filtering. Given a large set of historical setups with known outcomes, a model can learn which combinations of conditions tended to resolve badly, and reject those. That is a real, useful capability and it maps well onto what machine learning does well.

What it is not good at: predicting price. Markets are adversarial and largely efficient at the timescales retail traders operate on. Any reliably predictive signal attracts capital until it stops being predictive. A product claiming its model predicts direction is claiming to have solved a problem that very well-funded institutions have not.

What is usually happening instead: a conventional strategy with an "AI" label, or a large language model used to summarise news. The second is genuinely useful for research and is nothing like a trading model.

Why AI overfits worse

Overfitting is the failure mode of every data-driven strategy: the model learns the noise in its training data instead of a durable relationship. Machine learning makes it worse for a simple structural reason, capacity.

A moving-average crossover has two or three parameters. A neural network can have millions. With that much capacity, a model can memorise its training set almost perfectly, producing a backtest that looks extraordinary and generalises to nothing.

The defences are well established, and their absence is diagnostic:

• Walk-forward validation. Train on a window, test on the next unseen window, roll forward. Anything else leaks future information into training.
• Out-of-sample holdout that is genuinely untouched. If the holdout was used to pick the model, it is not a holdout.
• Live forward performance. The only test that cannot be gamed.

A vendor who cannot describe their validation approach has almost certainly not done any.

Five questions that expose a label

1. What kind of model is it? "Neural network", "gradient boosting", "reinforcement learning" are answers. "Proprietary AI" is not.
2. What are the inputs? A real system has a feature set someone can describe.
3. What is the output? A probability? A classification? A direction? Vagueness here usually means there is no model.
4. How was it validated? Walk-forward and out-of-sample, or nothing.
5. How often is it retrained? A model trained once in 2023 is applying 2023's relationships to today.

Any product that cannot answer four of the five is using "AI" as a marketing adjective.

Where MarketPro uses AI, precisely

MarketPro uses an AI layer at exactly one point in the pipeline: reviewing candidate signals before they are published.

The strategy engine finds setups. The AI layer then sees those candidates alongside their diagnostics and the instrument's own published history, and decides whether each one should reach users. It can reject a technically valid setup for reasons the scan has no representation for. Event risk inside the expected holding period, a spread that consumes the edge, a target needing more room than the session typically offers.

Three things it deliberately does not do:

• It does not generate signals. Its job is rejection, not creation.
• It does not predict price. It assesses whether a candidate meets the publishing bar.
• It does not execute anything. MarketPro has no access to any trading account.

We describe it this narrowly on purpose. "AI-reviewed" is a claim about a review step, and it would be trivially easy to imply something larger. Our editorial policy sets out what we will and will not claim.

What an automated system cannot do

Four limits are structural rather than a matter of build quality, and understanding them is most of what separates people who use automation well from people who lose money to it.

It cannot know what it has not been shown. A rule set encodes past relationships. When the regime changes (a trending market turns to chop, a correlation that held for two years breaks) the system keeps applying yesterday's logic with full confidence and no awareness that anything has changed.

Backtests overstate almost everything. A backtest with clean fills, fixed spread and no slippage is not a simulation of trading, it is a simulation of arithmetic. Add realistic spread, variable execution and the requeues that happen in fast markets, and a strategy that looked excellent frequently becomes marginal.

Optimisation finds coincidences. Tune enough parameters against enough history and you will find a setting that fits it perfectly. That setting describes the noise in that particular sample, not a property of the market, and it stops working the moment it meets data it was not fitted to. This is curve-fitting, and it is the single most common reason a purchased system fails.

Recovery logic hides risk instead of removing it. Martingale and grid systems produce beautiful equity curves for months, because doubling into a losing position converts many small losses into rare enormous ones. The curve is not evidence of an edge; it is evidence that the loss has not arrived yet.

The MarketPro Expert Advisor

MarketPro is building an Expert Advisor for MetaTrader 4 and MetaTrader 5. It is not available to download yet, and this page will say so until it is.

What is open today is the waiting list. Install the app, open the EA tab, and join it; everyone on the list is emailed when the builds go live. There is no charge and no card involved in joining.

When it does ship, three things will be true about it by design:

• You install and run it. It executes on your terminal, on your account, under your configuration. MarketPro never touches your account and exercises no discretion over your funds.
• It ships with a kill switch and an explicit as-is licence with no performance guarantee.
• Availability is limited by region. Distribution is restricted, and eligibility is determined by residence rather than by a checkbox on a page.

Meanwhile the thing that is live is the signal feed: vetted trade ideas with entry, stop and three targets that you place yourself. That is the manual equivalent of what the EA will automate, and it is available today with one free signal a day.