Effectiveness of Fine-Tuning LLMs for Trading Agents

Determine whether fine-tuning large language models used within financial trading agents improves trading performance compared with using only in-context learning without fine-tuning.

Background

The survey notes that most LLM-based trading systems rely on closed-source models and primarily employ in-context learning rather than parameter fine-tuning. Among the reviewed works, only SEP (Koa et al., 2024) involves tuning during training, highlighting a gap in systematic evaluation of fine-tuning strategies for this application domain.

Understanding whether fine-tuning confers measurable advantages over in-context learning for trading—such as improved returns, risk-adjusted performance, or robustness—remains unresolved and is explicitly identified as an open question by the authors.