Contents
The honest framingWhere models fit todayWhere they do notReal examples in the wildA checklist for AI-flavoured bot claimsThe honest framing
Language models are text predictors with an enormous memory of what people wrote. Trading is the business of positioning capital against prices. The overlap between those two sentences is small, and most of the interesting work in this space lives exactly on that small overlap — summarizing, translating, drafting, explaining — not on the trading decision itself.
Where models fit today
| Job | Why a model helps | Maturity |
|---|---|---|
| Explaining a strategy file | Converts code into prose for a first read | Works well today |
| Drafting indicator scaffolding | Writes the repetitive boilerplate of a strategy class | Works, with review |
| Summarizing news and filings | Compresses text that a human then reads | Works, with sources kept |
| Writing exchange-API glue | Answers in code for endpoints, signing, pagination | Works well, with tests |
| Making the trading decision | Prices, positions, risk — the hard part | Not where the value is |
Where they do not
- Price prediction. A model trained on text has no privileged view of a candle it has never seen; the market's next move is not written anywhere it could have read it.
- Discretion at the order. "Let the model decide whether to enter" is not a strategy; it is an unresolved responsibility.
- Unattended execution. Any pipeline where a model output becomes an order without a human-shaped gate is a pipeline waiting for the day the model is confidently wrong.
Real examples in the wild
There are genuine, useful integrations in open source: bots that ask a model to annotate a backtest report in plain language; tools that turn a written plan into a first-draft strategy file for a human to fix; assistants that answer questions about a strategy's own code. These work because the model is doing what it is built for — reading, writing, explaining — while the strategy, the risk rules, and the final button stay in deterministic code.
A checklist for AI-flavoured bot claims
- Where exactly does the model sit in the pipeline? If the answer is "it decides", keep walking.
- What happens when the model is unavailable at candle close? Deterministic fallback or no trades?
- Can you reproduce the model's contribution with the model switched off? If not, it was load-bearing in a bad place.
- Who audits the prompts and the model version? A prompt change is a strategy change.