Design and Empirical Study of a Large Language Model-Based Multi-Agent Investment System for Chinese Public REITs

Battle Count: 128

Highly relevant to quantitative trading as it presents a complete closed-loop trading system (analysis-prediction-decision-execution) with multi-agent architecture, dynamic volatility thresholds, multi-horizon directional prediction, discrete position management, and risk control constraints. The comparison between general-purpose and fine-tuned specialized LLMs for prediction is directly applicable to trading strategy development. The four-quadrant macro framework and technical indicator integration provide structured decision-making logic. However, it is domain-specific to Chinese REITs and lacks comparison with traditional quant strategies.

The system involves multiple components: four analytical agents with distinct data pipelines (technical indicators, announcement analysis, event detection, macro market analysis), a prediction agent with multi-horizon output, a decision agent with position management logic, dynamic volatility threshold calculation, four-quadrant macro framework, and a two-stage training pipeline (SFT + GSPO). The fine-tuning process requires teacher model distillation, structured data construction, and reward design. However, the architecture is modular and the code is open-sourced, reducing practical implementation barriers.

Project code and data are open-sourced on GitHub (https://github.com/adennng/REITs-MAS-LLM). The methodology is well-documented with specific parameter settings (e.g., Nv=30 days, quantile boundaries at 30%/70%, thresholds tau_high=1.4, tau_low=0.7). Training data period (Oct 2023 - Sep 2024) and backtest period (Oct 2024 - Oct 2025) are clearly defined. However, some implementation details of the agent prompts and the exact GSPO reward weights (alpha, beta) are not fully specified. The 28 REITs fund list is provided.