Marketing calls all of it ‘AI’. Under the bonnet there are four very different technologies, and knowing which one you’re looking at tells you most of what you need to know. A guide for the technically curious.
AI trading tools are having their moment. App stores are full of them, ads for them follow you around the internet, and every one of them describes itself in roughly the same words: intelligent, automated, data-driven. What the marketing never tells you is what the software actually does – and for anyone with a technical bent, that is the single most useful question, because the category label ‘AI trading’ covers everything from a moving-average calculator to a genuinely sophisticated machine-learning pipeline.
This guide unpacks the layers, in ascending order of sophistication.
Layer one: rule-based automation wearing an AI badge
A large share of ‘AI’ trading tools are classical rule engines: if the 50-day average crosses the 200-day, buy; if the RSI passes 70, sell. These are if-statements, some of them decades old, executed quickly and dressed in contemporary branding. There is nothing wrong with rule-based trading – it is transparent, testable and predictable, which are virtues – but it is automation, not intelligence, and it has been available in retail platforms since the 2000s. If a tool’s documentation talks about ‘signals’ and ‘indicators’ without ever mentioning a model or training data, this is usually the layer you are in.
Layer two: statistical models and genuine machine learning
The next tier up uses actual learned models – regressions, gradient-boosted trees, sometimes neural networks – trained on historical market data to score opportunities or forecast short-horizon moves. This is real machine learning, and it comes with machine learning’s classic failure mode: overfitting. A model tuned until it explains the past perfectly has usually learned the noise rather than the signal, and financial data is mostly noise. The tell-tale symptom is the immaculate backtest – performance curves that glide upward through a decade of history – followed by unremarkable live results. Any developer who has watched a model ace validation and faceplant in production will recognise the pattern; in trading, the faceplant costs money.
Layer three: LLMs reading the news
The newest layer uses large language models for sentiment and news analysis – parsing headlines, filings and social feeds to gauge mood around an asset. It is a genuinely new capability and an active research area, but be clear about what it does: it summarises what is being said, quickly. It does not know the future, and sentiment signals decay within minutes to hours because everyone else’s LLM read the same headline. Tools leaning heavily on ‘ChatGPT-powered analysis’ in their marketing are usually thin wrappers around an API call – technically real, economically marginal.
Layer four: execution – the part nobody markets
Whatever the intelligence layer, every tool ends at the same place: an order sent to a brokerage, usually via API keys you grant it. This unglamorous layer is where users actually get hurt. Slippage between signal and fill, spreads that widen at exactly the moments strategies trade most, rate limits and failed orders on volatile days, and – the security-critical detail – what permissions those API keys carry. A trading key should never have withdrawal rights, and any tool that asks for them should be closed immediately. If you evaluate nothing else technically, evaluate this layer.
How to test a tool’s claims without funding it blind
Three practical checks. First, demand live, dated results and ignore backtests entirely – simulations are controlled by the vendor. Second, compound any claimed monthly return over a year and see whether the number survives contact with reality; a ‘reliable 10% a month’ is a claimed 214% a year, which would make its author the greatest investor in history rather than a $49 subscription. Third, find out how the tool makes money – subscriptions are clean, while per-trade fees or broker kickbacks mean the vendor profits from your activity whether or not you do.
Better still, look for testing done with real money by someone independent. The Investors Centre’s reviews of AI trading software available to UK users are built exactly that way – the firm deposits its own funds with each tool and measures delivered performance against marketed claims, including the execution-layer costs that never appear in a demo. Their headline finding is one every engineer will find familiar: the gap between the spec sheet and production behaviour is the rule, not the exception.
The honest summary for a technical reader
Most retail ‘AI trading’ is automation plus marketing. The genuine machine-learning tier exists but battles overfitting in one of the noisiest datasets there is. The LLM tier is real technology with rapidly decaying signal value. And the execution layer – keys, fills, costs – is where the practical risk lives, whatever the intelligence above it. Treat the category the way you would treat any vendor’s claims about production software: assume nothing, verify with small live tests, and grant the minimum permissions that let the thing run.
