"AI-powered trading" shows up in the marketing of an enormous number of retail EAs, and in a meaningful share of those cases, there's remarkably little machine learning actually happening underneath the label. It's worth separating the real thing from the marketing.

What machine learning is not doing

It's not predicting the future. Despite how the phrase "AI trading" is often used, no legitimate machine learning system in a retail EA is reliably forecasting where price will be in an hour, a day, or a week. Markets are influenced by too many unpredictable, non-stationary factors for that to be a realistic claim — and any seller implying otherwise is selling a story, not a system.

It's not a black box that "just knows." A model that can't explain, at least at a mechanism level, what it's actually scoring and why isn't more trustworthy for being mysterious — it's less trustworthy, because there's no way to sanity-check whether it's doing something reasonable or something coincidental.

What machine learning can genuinely do

Score the quality of a setup, not predict a price. A well-built ML filter isn't trying to guess future price movement — it's evaluating whether a current setup resembles conditions that have historically led to good outcomes versus poor ones, based on real, accumulated trade data. That's a fundamentally different (and far more achievable) task than prediction.

Learn from live outcomes, not just a static backtest. The meaningful distinction is between a model trained once on historical data and left alone, versus one that continues learning from real, live trading outcomes as they accumulate — filtering out the kinds of setups that have actually underperformed in real conditions, not just in a backtest.

Filter, not decide alone. In a well-built system, the ML component is one layer among several — structure, momentum, risk-reward, and news filters typically all have to agree before machine learning is even consulted, and the ML layer's job is narrower: veto the weaker candidates among setups that already passed everything else.

How to tell the difference from the outside

Ask what the model is actually trained on. "Real, live trading outcomes, retrained regularly" is a very different (and verifiable) answer than "advanced AI algorithms" with no further explanation.

Ask what happens when there isn't enough data yet. A genuinely data-driven system should refuse to trade a new instrument or condition until it has enough real samples to trust — not fall back on a guessed default and call it AI anyway.

Ask whether the ML component is disclosed as one layer among several, or presented as the entire strategy. A single model making every decision alone is a much bigger claim (and a much bigger risk) than a model narrowly filtering candidates that already passed independent structural and risk checks.

What this looks like in practice

QMS Trading's EAs use machine learning as a genuine filtering layer — trained on real, accumulated live trading outcomes rather than a static backtest, applied on top of independent structure, momentum, and risk-reward checks rather than as the sole decision-maker. A symbol simply won't trade until there's enough real data to trust. See exactly how the filtering works, layer by layer.

Machine learning in trading is real, useful, and genuinely different from a rules-only system — but it's a filter, not a fortune teller. Anyone selling you the second version is selling you marketing.