LLM trading agents
What are LLM trading agents actually for?
A language model is very good at some things a trading system needs and actively unsuited to others. Most disappointment in this area comes from asking one to do the second kind of work. It is worth being specific about which is which.
What are language models genuinely good at?
Reading things that were never structured
Filings, disclosures, transcripts and announcements are written for humans and arrive in no fixed shape. Extracting meaning from them at volume is the single clearest advantage a language model brings to this problem.
Holding several considerations at once
Weighing a disclosure against a calendar against a position, and producing an argument rather than a score, is work that suits a model built on language. It is also what makes the reasoning readable afterwards.
Arguing a case it does not hold
Constructing the strongest case against a trade is a language task, and it is more useful than constructing the case for one. Most candidates should not be taken, and articulating why is the valuable output.
Explaining itself in a form a person can audit
A decision that comes with its reasoning in plain language can be reviewed by whoever runs the desk. A decision that comes as a number cannot.
What are language models bad at?
Arithmetic and position sizing
Language models are unreliable at exact calculation. Anything involving how much to buy, what a stop implies, or what a spread costs belongs in ordinary deterministic code, checked before an order exists.
Consistency across identical inputs
The same question can produce a differently worded answer. That is tolerable in argument and intolerable in a control. Nothing that bounds loss should depend on a model producing the same output twice.
Knowing what it does not know
A model asked for a view will produce one, whether or not the evidence supports it. Confidence has to be established by grading past decisions against outcomes, not by asking the model how sure it is.
Anything time-critical or high-frequency
Model calls are slow relative to markets. A design that needs one in the execution path is the wrong design; the language work belongs in forming views, not in placing orders.
Which failure catches people out?
Look-ahead through the training data
A model may know what happened after the date you are testing, which makes a backtest meaningless in a way that is hard to see. Any research using a language model has to be constructed so the model cannot draw on knowledge from after the point being tested.
Plausible reasoning about a false premise
A model given a wrong input produces a confident, well-argued conclusion from it. Garbage in does not produce obvious garbage out; it produces persuasive garbage, which is worse.
Retraining changing behaviour underneath you
If the model runs on a vendor's infrastructure, it can be updated on their schedule and your system's behaviour changes without your approval. Running locally removes that.
How does this desk use language models?
Language work forms views, code enforces bounds
Reading and arguing are language tasks. Sizing, protection and the limits that cap loss are deterministic code, and no model output may weaken them.
Every argument is recorded and later graded
The reasoning is written down before the trade and scored afterwards against what the market did. That is how confidence is established, rather than by asking the model.
It runs on your machine
The model runs on your own hardware. Nobody retrains it on a schedule you do not control, and your positions are not telemetry for someone else's product.
What should you ask before relying on one?
Which decisions does the model actually make
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, and control logic is riding on model output.
What happens when the model is unavailable
Model calls fail. Ask what the system does then: whether it holds, degrades to something simpler, or continues on stale reasoning. Continuing quietly is the wrong answer.
How much does each decision cost to make
Model inference has a real per-decision cost that scales with how often the system thinks. That cost is part of the economics and it is rarely disclosed.
Common questions
- Can a language model trade on its own?
- It can form and argue views, which is the hard part of the work, but it should not be doing arithmetic, position sizing or anything that bounds loss. Those need deterministic code, because language models are unreliable at exact calculation and inconsistent across identical inputs. A workable design uses the model for judgment and ordinary code for control.
- What are language models bad at in trading?
- Exact arithmetic, consistency across repeated identical inputs, knowing the limits of their own knowledge, and anything time-critical, since a model call is slow relative to markets. They will also reason confidently from a false input, producing persuasive rather than obviously wrong conclusions.
- What is look-ahead bias with language models?
- A model may already know what happened after the date a backtest is testing, because that period was in its training data. It can then appear to predict an outcome it simply remembers. Research using language models has to be built so the model cannot draw on knowledge from after the point being tested, and this is easy to get wrong invisibly.
- Should a language model decide position size?
- No. Sizing is arithmetic, and language models are unreliable at exact calculation and inconsistent across identical inputs. Anything that determines how much capital moves, what a protective stop implies, or what a spread costs belongs in deterministic code that is checked before an order exists.
TradeAgentic is an autonomous trading desk for macOS and Windows, licensed to operators and firms who intend to run it themselves.
Request licensing terms