LLM Agents Improve Trading Risk-Adjusted Returns

Takanobu Kawahara (arXiv preprint submitted Feb 26, 2026) proposes a multi-agent LLM trading framework that decomposes investment analysis into fine-grained tasks rather than coarse instructions. The system is evaluated on Japanese stock prices, financial statements, news, and macro data using leakage-controlled backtesting, where fine-grained decomposition improves risk-adjusted returns and alignment of intermediate outputs drives performance. Portfolio optimization leveraging low index correlation and output variance further boosts returns.