What Is AI Trading?
Search "what is ai trading" and you get five confident answers describing five different products. Here is what the term actually covers, what AI is genuinely good and bad at in markets, and the risks most explainers leave out.
Type "what is ai trading" into a search bar and the answers all sound authoritative, and none of them describe the same thing. One page means a chatbot that flags an oversold stock. Another means an execution algorithm running unattended inside a fund. A third means a bot with standing access to someone's brokerage account. These are not the same product, they do not carry the same risk, and treating them as one thing is most of why the term causes confusion.
AI trading is a spectrum, not one thing
At the broadest level, AI trading means using machine learning or other AI techniques somewhere in the process of deciding what to buy or sell. That definition is wide enough to cover four meaningfully different setups, each with its own risk profile.
- Signal generation. A model flags a possible opportunity — this instrument looks unusual today, sentiment on this pair just shifted — and a person decides everything that happens after that. The AI's job ends at the alert.
- Decision support. The model scores or ranks candidate setups, prioritizing what a trader looks at first, without placing an order itself. The person still pulls the trigger.
- Semi-autonomous execution. The model prepares or proposes a trade — direction, size, stop — but a human, or an explicit rule the human set in advance, has to approve it before it reaches a broker.
- Fully autonomous execution. The model observes the market, decides, and sends the order, with no human in the loop for that specific decision.
What AI is actually good at here
None of this is about hype. It is about matching the tool to the job. AI genuinely does a few things better than a person watching charts by hand:
- Finding patterns across more data than a person can hold in working memory: years of price history, several correlated instruments, all at once, without fatigue.
- Reading text at volume. A news or sentiment model can score thousands of headlines and posts before a person finishes reading the first ten.
- Executing a fixed rule exactly as written, every time. It will not second-guess a stop-loss at 2 a.m. because the trade "feels" like it will turn around.
What AI is actually bad at here
The failure modes are not about the model being unintelligent. They come from what the tool actually is: a pattern-matcher trained on the past, asked to operate in a market that keeps changing.
- Explaining itself. Most models are opaque by default: you get an output, not the reasoning behind it. That is tolerable for a recommendation engine. It is a real problem when the output moves money out of an account unattended.
- Knowing when history stops applying. A rate cycle turning, a liquidity shock, a new dominant participant in a market: any of these can invalidate a pattern that looked reliable across years of backtested data, and the model has no built-in way to notice the ground shifted.
- Handling situations outside its training data. Ask it to react to something it has not effectively seen before and it still produces an answer, with the same apparent confidence as always. It does not know what it does not know.
The gap between a backtested edge and a live one
This is the risk most explainers skip, and it is the one that actually costs people money. A strategy can produce a clean, rising equity curve on years of historical data and still lose money from the day it goes live — and that has nothing to do with bad luck.
Three causes show up again and again. Overfitting happens when a model or rule set is tuned until it fits the noise in one specific dataset rather than a pattern that will repeat. Lookahead bias happens when the backtest quietly used information that would not actually have been available at the moment of the decision. And execution reality — spread, slippage, and latency — is what a spreadsheet ignores and a live account pays on every trade.
Two more risks worth naming
- Automation is not the same as safety. Removing human emotion from execution removes one source of error. It does not remove data errors, connectivity failures, or a strategy that was wrong to begin with. It just executes the mistake faster and more consistently.
- Vendor and data dependency. An AI trading system is only as good as the data feed and the model behind it. If either goes stale, wrong, or unavailable, the system does not know to stop unless something outside the model tells it to.
Where Cortiq sits on this spectrum
Cortiq is a full execution agent, not a signal service, closer to the far end of the list above. Two design choices bring the risks in this piece back under control instead of ignoring them. Every session is locked to a single instrument, on purpose: depth on one market compounds into a real, inspectable track record, rather than spreading evidence thin across a basket of symbols. And the AI is bring-your-own: Cortiq connects to an AI provider you choose and pay for separately, with every decision logged in plain text, before, during, and after, so the opacity problem above has an actual audit trail instead of a black box.
That does not remove the backtest-to-live gap or the need for supervision — nothing does. It keeps the failure mode visible and bounded instead of hidden. The full mechanics — the trade cycle, the guardrails, what the system does not do — are in [How Cortiq Actually Works](/how-it-works).
Key takeaways
- AI trading is not one product. It is a spectrum from a plain alert (signal generation) to a system that decides and executes with no human in the loop.
- AI is genuinely good at scanning large datasets and text at volume, and executing a fixed rule without fatigue. It is bad at explaining itself, and at knowing when a pattern has stopped working.
- A clean backtest is a hypothesis, not a proof — overfitting, lookahead bias, and real-world execution costs are why a backtested edge and a live one are not the same thing.
- Automation removes emotional error. It does not remove data risk, model risk, or a bad strategy — it executes whatever it is given, correct or not, faster and more consistently.
- Where a system sits on the spectrum, and what sits outside the model to bound its mistakes, matters more than any performance claim attached to it.
The next post in this series goes one level deeper: the four different things people actually mean when they say "AI in trading," and what has to sit outside the model for any of them to be safe to run unattended — see [AI in Trading: How It Actually Works](/blog/ai-in-trading-how-it-works).