AI stock trading risks

By the TradeAgentic engineering team·Published 30 August 2026·Updated 30 August 2026

The risks are not the ones people expect

Asked what could go wrong, most people describe the system being wrong about a stock. That is not a risk, it is the normal case, and it happens constantly to every system trading equities. The risks worth naming are the ones that turn an ordinary bad week into something you do not recover from.

Every one of them is a design question rather than a market question, which means every one of them is checkable before you commit capital.

What compounds rather than bounds?

Protection that lives in the software

The single most consequential shortcut a vendor can take, and it is rarely volunteered. If a stop exists only in a running process, a crash, a reboot, a network drop or a closed laptop removes it. The position stays open and unprotected and nothing tells you.

Silent degradation on bad data

Feeds go stale and quotes go wide. A system that keeps trading on numbers it cannot verify is guessing with confidence, which is worse than halting. The correct behaviour is standing aside and saying so on screen.

Concentration nobody measured

Approaches that look independent can take correlated positions and concentrate risk without either appearing to misbehave. This is what turns a bad week into a redemption conversation.

Size that increases after a loss

Any system doing this has a martingale in it, whatever the marketing calls it. It is reliably catastrophic and its presence is a reason to walk away.

What is specific to the AI part?

Confident reasoning from a wrong input

A model given bad data produces a well-argued conclusion from it. Garbage in does not produce obvious garbage out here, it produces persuasive garbage, which is harder to catch.

Look-ahead through training data

A model may know what happened after the period a backtest covers, so it appears to predict what it is remembering. The research has to be built so it cannot, and getting this wrong is invisible from the results.

Retraining on somebody else's schedule

If the model runs on a vendor's infrastructure, your system's behaviour can change without your approval. Running locally removes the exposure rather than managing it.

Control logic riding on model output

Ask a vendor to draw the line between what the model decides and what deterministic code decides. If they cannot, the line probably is not there.

What makes the rest survivable?

Bounds outside the software

Loss limits, a control that flattens everything, protective orders resting at the brokerage, and a refusal to trade when startup checks fail. If the system's own optimisation can widen any of them, they are suggestions.

A record that survives the decision

Without it you cannot tell a bad process from bad luck, which means you cannot improve anything or explain anything to anyone.

Your own sizing

The largest single determinant of your outcome and entirely yours. Size the first live allocation so total loss would be annoying rather than serious.

What no design protects against

Gaps

A stop cannot execute at your price if the market opens past it. Avoiding known events matters more than stop placement, and no vendor can engineer this away.

Broker outages

Your protection rests at the brokerage, which is the right place for it, and the brokerage is still a single point of failure you do not control.

Trading itself

It involves substantial risk of loss. Controls bound the damage rather than eliminating it, and any vendor implying otherwise is telling you something untrue.