Adaptive LLM-based multi-agent systems to enhance quantitative trading performance

Traditional quantitative trading approaches have limited capacity for synthesizing heterogeneous data sources and adapting to market regime shifts. We propose a three-layer framework that systematically integrates large language models for semantic processing, multi-agent systems for collaborative intelligence, and Deep Reinforcement Learning for adaptive decision-making. The architecture employs five specialized agents coordinated through a Model Context Provider mechanism to generate daily factor scores that inform the Proximal Policy Optimization algorithm. We conduct rigorous empirical validation across five US equities using 25 years of historical data with strict temporal partitioning to prevent look-ahead bias. During the volatile out-of-sample test period from July 2024 to June 2025, the framework achieves an average annualized return of 53.87% with a Sharpe ratio of 1.702, substantially exceeding the buy-and-hold benchmark’s 26.08% return and Sharpe ratio of 0.765. The maximum drawdown averages 12.54%, compared to 30.24% for the passive strategies. Comprehensive comparisons against 15 baseline models confirm statistically significant superiority across all metrics, with Diebold–Mariano tests yielding p -values < 0.0001 after Bonferroni correction. Systematic ablation studies demonstrate a genuine architectural synergy, with complete three-layer integration outperforming the best two-component combination by 15.35 percentage points. The framework maintains robust performance across volatility regimes, demonstrating its particular strength during periods of high volatility when traditional approaches experience negative returns.