01 ROLE

AI TRADING AGENT

The model researches the strategy — and remains part of the trading decision.

LLM Trading is designed for cases where you want the model to interpret market context instead of reducing the final method to fixed deterministic code. Start with instruments, clusters, chart examples and a research objective. The project keeps the research conversation, governing prompt, backtest evidence and later trading decisions connected to one strategy context.

EXAMPLE RESEARCH CASE Find a repeatable setup across a group of FX pairs and improve the equity curve without removing activity.

The agent can inspect charts, compare winning and failing examples, propose a hypothesis, test it, explain where the hypothesis breaks and revise the strategy prompt for another historical replay.

02 PIPELINE

HOW THE PROJECT WORKS

From market evidence to an LLM strategy that can be replayed and traded.

INPUTS Charts, instruments, clusters and research objective

Define what the model should investigate and which instruments provide tradable or contextual evidence. Hard Settings keep risk permissions and runtime behavior inside a fixed contract.

RESEARCH Chat searches for a repeatable market explanation

Ask the model to find common setups, compare regimes, inspect correlations, explain losing cases, flatten the equity curve or abandon a weak hypothesis and search for another method.

STRATEGY PROMPT Research is compiled into the governing decision logic

The active prompt captures the strategy assumptions and tells the selected GPT model how to interpret charts, open positions, manage risk and decide between trading actions.

LLM BACKTEST Replay the same class of decisions on historical candles

During historical replay the model receives the allowed context and makes LLM trading decisions. Trades, PnL, reasoning and chart evidence become the next research input rather than only a score.

OPTIMIZATION Analyze failure patterns and improve the hypothesis

Backtest conclusions are fed into the next reasoning cycle. The model can identify weak assumptions, revise the prompt and run another replay while project history remains attached to the same workspace.

RUNTIME AI remains in the loop for paper or live decisions

This is the defining feature of LLM Trading: the model is not only the strategy designer. It also interprets the current trading context when the agent is asked to make the next decision.

03 USE CASES

EXAMPLE WORKFLOWS

Use Chat as an active trading researcher, not a one-shot strategy generator.

01

PATTERN DISCOVERY

Find a common setup in several charts

Give the project a group of instruments and ask it to identify a repeatable market structure rather than optimize one fixed rule.

OUTPUT: hypothesis → prompt → replay → evidence.

02

EQUITY CURVE

Work on drawdown or curve smoothness

Ask why losses cluster in one regime, which assumptions create the drawdown and whether the method can be improved without destroying trade activity.

OUTPUT: revised hypothesis tested against the previous backtest.

03

CLUSTERS

Reason across correlated instruments

Use multi-instrument context for relative-value, correlation or hedging reasoning while controlling which cluster members can trade and which remain context-only.

OUTPUT: synchronized decisions informed by cross-market context.

04

RESEARCH RESET

Abandon a weak method and search again

If optimization exposes a bad thesis, return to research instead of forcing the same strategy family. The conversation can move back from optimization to discovery.

OUTPUT: a new research direction without losing project history.

04 STRUCTURE

PROJECT STRUCTURE

One workspace connects contract, research, strategy artifact and replay.

01 Hard Settings Instruments, clusters, sources, risk permissions, triggers and runtime constraints remain explicit.

02 Chat Research conversation where the user can direct the investigation and ask the model to analyze backtest evidence.

03 Strategy Prompt The governing strategy artifact used by the model to interpret context and make decisions.

04 Backtest + Optimization Historical LLM replay, trade evidence, reasoning and iterative prompt improvement remain attached to the project.

MARKET CONTEXT Charts · positions · clusters · permitted sources

GPT DECISION LAYER Research · replay · optimization · runtime reasoning

TRADING ACTIONS Open · close · hold · reverse · risk updates

AVAILABLE GPT TIERS Choose the intelligence level per project.

STANDARD

GPT-5.4 mini

Lower-cost model for routine Builder work and economical LLM trading iterations.

Input: $0.98 / 1M
Cached: $0.10 / 1M
Output: $5.85 / 1M

RECOMMENDED

GPT-5.6 Terra

Stronger reasoning with a practical intelligence-to-cost balance for strategy research and replay.

Input: $2.60 / 1M
Cached: $0.26 / 1M
Output: $15.60 / 1M

FRONTIER

GPT-5.6 Sol

Highest reasoning tier for difficult research, complex market interpretation and strategy debugging.

Input: $6.50 / 1M
Cached: $0.65 / 1M
Output: $39.00 / 1M

ILLUSTRATIVE COST LLM Trading uses model calls during research, historical replay and runtime decisions.

A representative workflow with 10 Chat turns, prompt compilation, one 90-decision replay and five optimization stages is approximately $4.59 on GPT-5.4 mini, $12.20 on GPT-5.6 Terra or $30.51 on GPT-5.6 Sol. This is an example, not a fixed quote; actual usage depends on context size, output and the number of LLM decisions.

01Top up Art Argentum balance

02Select GPT tier

03Usage is metered

04New paid AI work pauses when balance is insufficient

05 START

BUILD AN LLM TRADING PROJECT

Start with the market question. Let the project develop the method around the evidence.

Choose instruments or clusters, define the hard contract and use Chat to direct the research. The project keeps the strategy prompt, replay evidence and later LLM decisions in one continuous workflow.

Build LLM Trading Agent