Can ChatGPT build you a trading bot?

Can ChatGPT actually trade stocks?

Not on its own. A chat session has no brokerage account, no order routing and no custody of anything. What it can do is write the program that connects to a broker you have an account with, using keys you supply — and products exist that package exactly that, so the model sits behind a button rather than behind a code editor.

So the mechanical answer is yes, by way of code and an API. The useful answer is that access was never the constraint. Placing an order has been programmatically available to retail traders for years, and it has never been the reason a strategy did or did not work.

The constraint is knowing whether the rule you are about to automate describes something about the market or something about your search for a rule. Nothing about connecting a model to a broker addresses that, and connecting it faster addresses it less.

Why is the bot the easy part?

A bot is an execution wrapper. It takes a decision rule you already have and converts it into orders on a schedule, without getting bored, distracted or hopeful. That is genuinely valuable — discretion is where a lot of otherwise sound plans go to die — but notice what it is not. It adds no edge. It cannot add one. It multiplies whatever you point it at.

Which means the sign of the thing matters more than the size. Automating a rule that has real structure behind it compounds a small advantage reliably. Automating a coincidence converts a coincidence into a schedule, executed faster and with less friction than you could manage by hand.

Every part of the bot except one is a solved engineering problem with mature libraries behind it. The unsolved part is not in the code.

What breaks when the strategy came out of a chat window?

The specific failure is not that the model produces bad strategies. It is that the conversation is itself a search, and nothing records how large the search was.

You ask for a strategy. You backtest it. It is fine, not exciting. You ask for a variation, then a different asset, then a shorter lookback. Over a week you have perhaps twenty candidates and you keep the one that tested best. You have run a search over twenty variants — which is exactly the process that manufactures results that do not repeat — and the number twenty appears nowhere, in any artifact, ever. The model cannot tell you it either: earlier turns fall out of context and separate sessions never shared one.

The count of variants tried is the single largest input into whether a backtested result means anything. Chat is a medium that structurally destroys it, and tool-using agents make it worse by moving the number from twenty into the hundreds inside an afternoon.

Backtested and hypothetical results are not indicative of future performance. Nothing on this page is investment advice.

What checks would make a chat-derived strategy believable?

The question becomes answerable the moment you stop looking at the equity curve and start asking how the candidate was chosen. Four measurements do most of the work, and all four are standard quantitative finance rather than anything proprietary.

• A genuinely held-out period. Data that played no part in any decision — not in choosing the rule, not in tuning it, not in deciding to keep asking for more ideas. Once you have adjusted anything in response to it, it is in-sample.
• Costs applied before you look. Spread, commission, slippage and, at size, market impact. A frequently trading rule can be profitable before costs and reliably unprofitable after them.
• Performance discounted for the number of variants tried. The best of two hundred candidates has to clear a much higher bar than the first thing you tested, and the deflated Sharpe ratio is the standard way to say how much higher.
• An estimate of how often the in-sample winner would land below median out of sample. That is what probability of backtest overfitting measures, and it is the number that speaks directly to the failure above.

None of these tells you a strategy will make money. Each tells you how much weight the backtest can carry. That is a smaller claim and it is the only honest one on offer.

Does a better model fix this?

No, and this is the part worth sitting with, because the intuitive answer is yes.

The missing input is not a fact about markets that a larger model would know. It is a fact about your process: how many things you tried before this one, and how much data stayed untouched while you tried them. No amount of capability inside the model gives it access to that, because the information does not exist anywhere in the conversation.

Worse, capability pushes in the wrong direction. A stronger model with tool access generates and tests more candidates per hour, which raises the bar the eventual winner must clear rather than lowering it. Better models make the search cheaper, and cheap search is the mechanism that produces false results in the first place.

What actually fixes it is structural: something other than the proposer has to keep the count and hold the data.

Common questions