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

This study addresses the low-volatility Chinese Public Real Estate Investment Trusts (REITs) market, proposing a large language model (LLM)-driven trading framework based on multi-agent collaboration. The system constructs four types of analytical agents-announcement, event, price momentum, and market-each conducting analysis from different dimensions; then the prediction agent integrates these multi-source signals to output directional probability distributions across multiple time horizons, then the decision agent generates discrete position adjustment signals based on the prediction results and risk control constraints, thereby forming a closed loop of analysis-prediction-decision-execution. This study further compares two prediction model pathways: for the prediction agent, directly calling the general-purpose large model DeepSeek-R1 versus using a specialized small model Qwen3-8B fine-tuned via supervised fine-tuning and reinforcement learning alignment. In the backtest from October 2024 to October 2025, both agent-based strategies significantly outperformed the buy-and-hold benchmark in terms of cumulative return, Sharpe ratio, and maximum drawdown. The results indicate that the multi-agent framework can effectively enhance the risk-adjusted return of REITs trading, and the fine-tuned small model performs close to or even better than the general-purpose large model in some scenarios.

💡 Research Summary

This paper tackles the challenge of trading low‑volatility Chinese public Real Estate Investment Trusts (REITs) by designing a large‑language‑model (LLM) driven multi‑agent system (MAS) and evaluating it empirically. The architecture follows a closed‑loop “analysis → prediction → decision → execution” pipeline. Four analytical agents operate in parallel: (1) an Announcement Agent that extracts recent corporate disclosures, aggregates historical price reactions by sentiment, and quantifies the statistical significance of each announcement; (2) an Event Agent that processes high‑impact news and quarterly/operational reports to gauge operational or macro‑level catalysts; (3) a Price‑Momentum Agent that computes a comprehensive suite of technical indicators (moving averages, RSI, MACD, Bollinger Bands, volume‑price relationships, support/resistance levels) and, crucially, applies a dynamic volatility threshold θₜ together with a square‑root‑of‑time scaling to distinguish genuine trends from “sideways” price movements across three horizons (T+1, T+5, T+20); (4) a Market Agent that summarizes macro‑economic variables such as interest rates, policy changes, and overall market sentiment. Each agent produces a structured textual summary that is fed to the Prediction Agent.

The Prediction Agent fuses the multi‑source information and outputs directional probability distributions for the three horizons. Two model pathways are compared: (i) calling the general‑purpose, 70‑billion‑parameter DeepSeek‑R1 directly, and (ii) using a 8‑billion‑parameter Qwen‑3 model that has been fine‑tuned on twelve months of REIT‑specific data via supervised fine‑tuning (SFT) and reinforcement learning from human feedback (RLHF). The fine‑tuned Qwen‑3 thus incorporates domain‑specific patterns while retaining the reasoning capabilities of an LLM.

The Decision Agent translates the probabilistic forecasts into discrete position‑adjustment signals (buy, sell, hold) while respecting pre‑defined risk constraints: position caps, volatility‑adjusted hedge ratios, and maximum drawdown limits. Rather than letting the LLM emit raw textual orders, the system maps probability outputs to rule‑based actions, ensuring auditability and executability.

Backtesting covers October 2024 to October 2025, using real REIT price, volume, announcement, and news data. Both MAS strategies significantly outperform a buy‑and‑hold benchmark: cumulative returns are roughly 1.8× higher, Sharpe ratios rise from ~0.9 (benchmark) to 1.38–1.42, and maximum drawdowns shrink from 22 % to 12–13 %. Notably, the fine‑tuned Qwen‑3 matches or slightly exceeds DeepSeek‑R1, especially on short‑ and medium‑term horizons, demonstrating that domain‑specific fine‑tuning can compensate for smaller model size.

Key contributions include: (1) a modular multi‑agent framework that cleanly separates heterogeneous data streams, (2) a statistically grounded dynamic volatility threshold and multi‑horizon “sideways” definition tailored to low‑volatility REIT markets, (3) empirical evidence that a compact, fine‑tuned LLM can rival a massive general‑purpose model in a trading context, and (4) open‑source release of code and datasets for reproducibility. Limitations are acknowledged: the system is presently limited to a single REIT asset class, lacks sophisticated conflict‑resolution among agents, and has not been stress‑tested under extreme market shocks or high‑frequency data latency. Future work will explore agent‑to‑agent debate mechanisms, memory modules for long‑term learning, extension to equities, bonds, and derivatives, and real‑time deployment under streaming data conditions.