LLM Agents in Financial Trading: A Systematic Review

A new paper on arXiv surveys the integration of Large Language Models (LLMs) into trading systems, treating them as expert-system decision pipelines. The study screens 77 included studies up to March 9, 2026, with a primary empirical subset of 19 papers meeting minimum criteria for action output and closed-loop evaluation. Key findings reveal protocol incomparability: only 2 of 19 studies report time-consistent split protocols, 1 includes an explicit transaction-cost model, 1 documents survivorship bias handling, 11 specify execution timing, and 15 are codified. The remaining 58 studies provide background context. The research highlights significant methodological inconsistencies in the field, suggesting that current LLM-based trading agents lack standardized evaluation frameworks.