What AI and Algorithmic Trading Actually Mean
"AI trading" is marketing shorthand for three very different things; learn to tell a rules-based bot from real machine learning from an AI-assisted tool, and most of the hype falls away.
An AI bot places smart buy and sell orders around the price to profit from volatility.
Pick the category the tool actually falls into, then check your answer.
Try the classifier above before reading on: sort each marketing claim into one of three buckets and watch how flimsy the label gets. "AI trading" is one of the most oversold phrases in the market. The same two words get stamped on everything from a simple bot following one fixed rule to a subscription that just resells someone else's buy and sell calls. Here is the good news: behind all the branding, any of these tools can only really be one of three things. Once you can name which one you are looking at, most of the hype falls away and you know exactly what to ask next. That filter is the single most useful skill in this whole course, so we start here.
The question that actually matters
The instinct is to ask "is this automated?". That is the wrong first question, because automation is already normal. Rules-based programs run the majority of United States equity trading volume, on the order of 60 to 75 per cent by common estimates. Automation is not new, and it is not sinister by itself.
The useful question is different: what kind of automation is this, and how much is it really deciding? A tool that follows one rule you can read is worlds apart from a model that invents its own rules, and both are different again from a tool that just helps you decide. The three categories below run from least to most autonomous.
Rules-based (algorithmic) trading
This is a program running fixed if-then rules that a human wrote in advance. A rule might read: "if the 50-day average crosses above the 200-day average, buy; if it crosses back below, sell." The program follows that exactly, tirelessly, without emotion. It never invents a new rule and it never learns.
Two words describe it well: deterministic and auditable. Given the same data it always does the same thing, and you can read the exact rule that produced any trade. Most so-called bots you will meet as a retail trader live right here, no matter what the landing page calls them.
Machine-learning trading
This one is different in kind. Instead of following rules a human wrote, the model studies historical data, works out its own patterns, and can adjust them as new data arrives. Nobody typed the rules; the model found them.
That makes it genuinely more powerful and also far more fragile. Because the model chases whatever fits the data it was shown, it easily learns noise that will not repeat (that trap is called overfitting, and it has its own module later). Real machine-learning trading is hard, hungry for data, and mostly the world of well-funded quantitative firms. Very few consumer products actually do it, whatever they claim.
AI-assisted trading
Here the human stays in charge. The tool augments your decisions instead of trading on its own: it might summarise news, flag a chart setup, suggest a position size, or answer a question, but you still pull the trigger.
This is the lowest-autonomy category, and for a retail trader it is often the most useful in practice, precisely because a person keeps judgement over the moments that matter most. "AI-assisted" and "fully automated" sit at opposite ends of the control spectrum. Treating one as the other is a classic marketing move.
What most "AI trading" products really are
Here is the punchline. The large majority of retail products sold as artificial intelligence are simple rules-based bots (a grid, or a moving-average crossover) or signal subscriptions that resell someone else's calls. Words like "AI-powered", "neural", "deep", "self-learning", and "quantum" are chosen to hide a plain engine and to justify a fee.
So before you judge any tool, force it into a category. Ask: is it running fixed rules, is it genuinely learning from data, or is it just helping a human decide? The category you land on tells you which questions to ask next and how much of the marketing to discount.
::fold{title="Decoding three "AI" products"} Same two words on the box, three completely different things inside. Try classifying each and asking the right follow-up.
- Product A: "AI bot places smart buy and sell orders around the price to profit from volatility." Decoded: a grid bot, pure rules-based automation. Follow-up: what happens when the market trends out of the grid?
- Product B: "Our deep-learning model predicts the next candle with 92% accuracy." Decoded: a machine-learning claim. Follow-up: show me the out-of-sample results and the costs (see the next fold for why the 92% alone means little).
- Product C: "AI copilot flags setups and explains the risk; you place the trade." Decoded: AI-assisted, human in control. Follow-up: does it improve my decisions without taking them away from me?
Learning to run this three-way sort in seconds is most of the battle. ::
Why a 92% win rate can still lose money
A high win rate sounds unbeatable, but it says nothing about payoff or costs. Suppose a tool wins 92% of the time but each win makes 1 while each loss costs 15.
Over 100 trades: you win 92 trades for 92 x 1 = 92, and you lose 8 trades for 8 x 15 = 120. Net result: a loss of 28, before fees and slippage make it worse. The "92%" is also almost always measured in-sample, meaning on the same data the model was tuned on, so even the number itself is usually flattering.
The lesson: win rate on its own is not evidence. Always ask about the size of the wins versus the losses, and about the costs.
Why AI-assisted can be the safest category
It feels backwards that the least autonomous tool is often the safest, but there is a solid reason. Fixed rules and trained models only know the situations they were built on. When something rare and outside that experience shows up (the kind of event covered later in the black-swan module), a human can recognise "this is not normal" and step in, while an automated system will keep confidently executing.
Keeping a person on the trigger is not a sign of a primitive tool. It is a deliberate safeguard over exactly the moments that do the most damage.
Quick knowledge check
In one line each, how do the three categories differ? Rules-based trading runs fixed if-then rules a human wrote; machine-learning trading works out its own patterns from data; AI-assisted trading helps a human who still makes the final call.
Why is "is this automated?" the wrong first question? Because automation is already mainstream (most US equity volume is rules-based), so the answer is almost always yes and tells you nothing. The useful question is what kind of automation it is and how much it actually decides.
A tool advertises a 92% win rate. Why is that not enough to trust it? Win rate ignores the size of wins versus losses and the costs, so a high win rate paired with lopsided payoffs can still lose money, and the figure is usually measured in-sample and unverified.
Sources
- SEC (Division of Trading and Markets), "Staff Report on Algorithmic Trading in U.S. Capital Markets": rules-based, pre-programmed order execution with little or no human interaction.
- arXiv (academic), citing Hendershott et al. (2011) and Brogaard et al. (2014): a majority of US equity volume is algorithmic (commonly cited around 60 to 75%).
- ISO/IEC 22989, "Artificial intelligence — Concepts and terminology": models that learn patterns from data instead of following explicitly written rules.
- SEC and FINRA (Joint Investor Alert), "Automated Investment Tools": a strategy reduced to pre-set entry and exit rules, with human oversight still needed.