“AI EA” is one of the most used labels in retail forex. Sometimes it describes a genuine machine-learning model; often it describes a conventional rule-based robot with a modern name. Knowing the difference helps you read any product page more critically.

What counts as AI here

  • Supervised learning — a model (for example a gradient-boosted tree or a neural network) is trained on historical features such as returns, volatility or indicator values to predict a label such as “next bar up or down”.
  • Reinforcement learning — an agent learns a policy by being rewarded for outcomes in a simulated market. Powerful in theory, very prone to learning quirks of the simulator.
  • Optimisation / genetic algorithms — searching thousands of parameter combinations for the best historical result. Frequently called “AI” in marketing; it is really parameter search.
  • Language models — used for summarising news or sentiment. They do not, by themselves, know where price is going.

Why markets are hard for machine learning

  1. Low signal-to-noise. Short-term price changes are dominated by noise. A model can easily “learn” patterns that were pure chance.
  2. Non-stationarity. Relationships change as central-bank policy, volatility regimes and market participants change. A pattern from 2015 may not exist in 2025.
  3. Limited data. Even decades of hourly data are small next to the number of parameters in a neural network.
  4. Costs. A tiny predictive edge can be erased by spread, commission and slippage.

Overfitting in one picture

fit on past data result on new data model complexity / number of tuned parameters →

The more freely a model is tuned to the past, the better it looks in a backtest and the worse it tends to do on data it has not seen. Good research spends most of its effort fighting this effect.

Questions to ask about any “AI” trading system

  • What exactly is the model, and what does it predict?
  • What data was it trained on, and which period was kept completely unseen for testing?
  • How many versions or parameter sets were tried before this one was published?
  • How does it perform after realistic spreads, commissions and slippage?
  • Is there an independent, long, live record — not a backtest or a demo account?

Related: backtesting & forward testing and evaluating EA claims.

Risk reminder. Trading foreign exchange and CFDs on margin carries a high level of risk and may not be suitable for every investor. You can lose more than your initial deposit. Past or simulated performance of any trading system, including automated Expert Advisors, does not guarantee future results.