What is ChatGPT genuinely good at in trading?
Language models are excellent at specification: taking a fuzzy description of what you do and turning it into an unambiguous set of rules with every edge case named.
This matters more than it sounds. 'I buy pullbacks in an uptrend' is not a strategy, it is a sentiment. What counts as an uptrend, over what lookback, on which timeframe? How deep is a pullback and how is depth measured? What invalidates the idea? A model that asks those questions one at a time will get you to a testable rule set faster than staring at a chart will.
They are also good at translation — turning rules into code for a specific platform, converting a strategy between languages, or explaining what an unfamiliar indicator actually computes. These are language tasks and they play to the technology's strengths.
And they are good at critique when asked properly. Given a strategy description, a model will readily list the ways it might be fitted, the costs it ignores, and the market conditions it has not met — provided you ask for the objections rather than for encouragement.
Why can't a language model tell you if a strategy works?
A language model cannot tell you whether a strategy works because determining that requires running the rules over price history and counting the outcomes, which is arithmetic on data rather than a question about language.
The model has no access to your instrument's tick history, cannot compute how many trades the rule would have taken, and has no way to check what a comparable random strategy would have scored. What it can do is produce a fluent and plausible assessment, which is worse than useless because fluency reads as confidence.
This is the specific failure mode to guard against. Ask a model whether your strategy is good and it will answer. The answer will be reasonable-sounding, will cite real concepts, and will be untethered from your data. Nothing in the response distinguishes it from an answer about a strategy that does not work.
The fix is procedural rather than technical: use the model to produce the rules and the code, and use a backtester to produce the verdict. Keep those two jobs in different tools so you never mistake one output for the other.
How do you prompt a model for a strategy that can be tested?
Prompt for precision rather than for performance: ask it to make your idea unambiguous and to name every parameter, rather than asking it for a strategy that works.
A productive sequence starts with your own idea, however rough, and asks the model to restate it as numbered rules with every threshold explicit and marked as a parameter. Then ask what a tester would have to decide that the rules do not specify — the ambiguous cases are exactly where a backtest quietly resolves things in your favour.
Next ask for the ways the rule set could produce a flattering backtest: which parameters would be tempting to tune, where look-ahead could creep in, what costs are being ignored. Models are notably good at this when asked directly, and notably bad at volunteering it.
Then take the rules to a backtester and grade them out-of-sample. If you use the model to iterate on the rules after seeing results, count each iteration — a conversation is a search, and thirty rounds of adjustment against the same history is thirty trials whether or not it felt like tuning.
What about strategies a model suggests on its own?
Strategies produced from a bare prompt tend to be composites of widely published material, which means they are the strategies most likely to have been arbitraged and least likely to contain anything distinctive.
That is not a criticism of the model; it is a description of what it was trained on. A moving average crossover with an RSI filter appears in a great deal of published material, so it is a natural completion to a request for a strategy. Popularity in training data is not evidence about markets.
The more useful direction is to bring your own hypothesis — something from your own trading, your own market observation, or your own instrument — and use the model to sharpen it. Your idea is at least not already in a thousand blog posts.
And whichever way the idea arrived, the verdict comes from the same place: a fixed rule set, graded on data that was not used to build it, with the number of variants tried written down.