ChatGPT Trading Bot

What people mean by it

There are two common set-ups. In the first, the language model is a coding helper. You describe a rule in words, it writes the code, and the finished bot is ordinary code that does the same thing every time. In the second, the bot sends market data or news text to a language model on every decision and trades on the answer. The second is harder to test, because the same input can produce a different answer on a different day, the model can be changed or retired by its provider, and you cannot write down the rule it is following.

This page is mostly about the first set-up, because it is the one you can test. For the structure of a bot see Python trading bot.

What a language model can safely do

• Write a first draft of code from a clear written rule, which you then read line by line and run on test data.
• Explain code you do not understand, including code from a seller. Ask it to explain each step, then check its explanation against what the code does.
• Turn a rule into a checklist: entry, exit, size, stop, what happens on a lost connection.
• Suggest tests and edge cases you had not thought of, such as a gap, a rejected order or a missing bar.
• Explain terms such as drawdown or slippage, which you can then confirm in an independent source.

What it cannot do

• Know the future. It writes fluent text from patterns in its training. A confident forecast from it is not evidence.
• Test the bot for you. It has not run your code on your data unless a tool did so, and even then the result is only as good as the data and the costs in it. See backtesting.
• Be relied on for facts. It can state an order type, a fee or a platform rule with confidence and be wrong. Rules vary by broker and exchange, so check yours.
• Take responsibility. You run the code, so you own what it does with your account.

The risks of running generated code

• Look-ahead mistakes. Code that reads a bar and trades on the same bar looks brilliant in a test. See look-ahead bias.
• Made-up functions. A model can call a library function or an order parameter that does not exist, or one that exists and does something else.
• Missing risk handling. Drafts often lack a position limit, a daily loss limit, a stop that does not depend on the bot, and handling for a failed order.
• Units and signs. Lots versus units, long versus short, percent versus fraction. A bug here can multiply a position.
• Secrets. Keys pasted into a chat or into code can leak. Keep them out of both, give them the smallest permission, and never allow withdrawals.
• Unreviewed changes. Each time you ask the model to "improve" the code, the whole bot needs testing again.

A worked example: a one-line mistake

Everything here is made up. We generated 500 random daily price moves with no pattern in them. The idea is simple: be long for a day if the market has been going up. Two versions of the code differ by one index:

• Peeks: be long on day i if day i's move is up. This is what a draft can do when it reads "today's close above the last close" and applies it to the same day.
• Fixed: be long on day i if day i-1's move was up, which is all that was known at the time.

Calculated by us on 29 Oct 2026, exact for these seeds, with no costs. The prices have no pattern, so no honest rule can earn from them. The peeking version still averaged +59.60 points because it only collects moves it already knows are up. The fixed version averaged -0.28 and finished above zero in 101 of 200 series, which is luck. Both versions run without an error message, and a backtest alone would have rewarded the broken one. The only way to catch it is to read the code and test it on data it has not seen.

A checklist of tests before any money

• Read every line. Be able to say in plain words what each decision uses and when it was known.
• Look-ahead check. Shift the signal by one bar. If the result changes a great deal, find out why.
• Random-data test. Run the bot on random prices. It should not make money. If it does, there is a bug.
• Backtest with costs. Include spread, fees and slippage.
• Untouched data. Test on a period the rule was never tuned on. See overfitting.
• Edge cases. Missing bars, a rejected order, a restart mid-trade, a dropped connection.
• Risk limits. Size per trade, a daily loss limit and a stop that works if the bot does not. See position sizing.
• Forward test and paper trade for weeks and compare with the backtest. See forward testing and paper trading.
• Only then, if at all, a very small size, with a key that cannot withdraw.

What this page does not tell you

It does not say a language model is useless or that a bot written with one will lose. It says the model's output is a draft, and a draft is not a tested bot. The numbers are made up and nothing here is a method to trade.