LLM agents inside the trading loop
A multi-agent trading system uses LLMs for market analysis while retaining conventional models and deterministic rules for portfolio construction and execution.
Paper
Inside the Trading Loop: Markets, Models, and Moves with LLM Agents
Authors
Renee Jia, Di Mu, Yao Yu, and Tianrui Liu
Date
8 November 2026
Inside the Trading Loop describes a hybrid trading architecture in which LLM agents analyse market information but do not directly execute trades. Separate agents process company news and macroeconomic information, returning structured reports to a coordinating agent that determines stock weights and cash reserves. Portfolio construction remains momentum-based, while execution follows rule-based procedures with deterministic fallbacks.
The system's broker-recorded paper account returned 34.5% between 11 March and 28 August 2026, compared with 14.4% for the S&P 500 ETF (SPY) over the same period. The higher return came with a larger maximum drawdown.
The authors also examine different LLM-enabled configurations through retrospective replays, including the effects of cash allocation, selective agent invocation and the structure of agent responses.
The results offer an example of how LLMs can be incorporated into an existing systematic trading architecture without handing them direct execution authority.
