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

In one sentence

Summary

The study addresses the challenge of automating investment decisions in the low-volatility Chinese Public Real Estate Investment Trusts (REITs) market, where traditional single-model approaches often struggle with multi-source information integration and risk control. The authors propose a closed-loop multi-agent system that decomposes the trading process into analysis, prediction, decision, and execution layers. The system employs four distinct analytical agents—announcement, event, price momentum, and market macro—to generate structured insights from diverse data sources. These insights are fused by a prediction agent to output directional probability distributions across multiple time horizons (T+1, T+5, T+20), which are then translated into discrete position adjustment signals by a decision agent constrained by risk limits. This architecture aims to enhance reasoning quality, reduce hallucination risks, and provide auditable, executable decision logic amidst market noise. The experimental design involves a 12-month backtest from October 2024 to October 2025 on 28 Chinese Public REITs listed for over one year. The core comparison evaluates two pathways for the prediction agent: using the general-purpose reasoning model DeepSeek-R1 versus a specialized small model, Qwen3-8B, fine-tuned via supervised fine-tuning (SFT) and Group Sequence Policy Optimization (GSPO) reinforcement learning. The fine-tuning process utilizes teacher-model distillation to create structured reasoning data and aligns the model with real price movements using a dynamic volatility threshold to define 'up', 'down', and 'sideways' labels. The system incorporates technical indicators, historical announcement impacts, event-driven news, and a four-quadrant macro allocation framework based on interest rates and equity market states. Performance is evaluated against a buy-and-hold benchmark using cumulative return, Sharpe ratio, and maximum drawdown metrics. Results indicate that both agent-based strategies significantly outperform the buy-and-hold benchmark in cumulative return, Sharpe ratio, and maximum drawdown control. The DeepSeek-R1 pathway achieved a mean cumulative return of 15.50% with a Sharpe ratio of 1.71, while the fine-tuned Qwen3-8B pathway achieved 13.75% return with a slightly higher Sharpe ratio of 1.77. Both strategies maintained maximum drawdowns under -5%, compared to over -11% for the benchmark. The fine-tuned small model demonstrated competitive performance, suggesting that domain-specific alignment can match or exceed general-purpose models in stability. However, the system showed limitations in capturing rapid trend accelerations and experienced synchronous drawdowns during market corrections, indicating room for improvement in high-volatility adaptability and decision agent proactiveness.

Research problem

The paper addresses the difficulty of automating investment decisions in the low-volatility Chinese Public REITs market, where single LLM models often fail to integrate diverse information sources and maintain robust risk control. It seeks to determine if a multi-agent framework can improve risk-adjusted returns and whether a fine-tuned small model can compete with general-purpose large models in this specific financial domain.

Core contributions

• Proposes a multi-agent LLM trading framework decomposed into announcement, event, price momentum, and market macro agents.
• Introduces a dynamic volatility threshold mechanism to define 'sideways' markets appropriate for low-volatility REITs.
• Develops a four-quadrant macro allocation framework integrating interest rates and equity market states.
• Compares a general-purpose LLM (DeepSeek-R1) against a fine-tuned small model (Qwen3-8B) using SFT and GSPO.
• Demonstrates significant outperformance over buy-and-hold in both return and risk metrics for Chinese Public REITs.

Method

• Four analytical agents: Announcement, Event, Price Momentum, Market Macro.
• Price Momentum Agent: Technical indicators (MA, RSI, MACD, Bollinger Bands) and dynamic volatility thresholds.
• Market Agent: Four-quadrant framework based on 10-year bond yield trends and equity market states.
• Prediction Agent: Fuses multi-source signals into directional probabilities (up/down/side) for T+1, T+5, T+20.
• Decision Agent: Maps predictions to discrete position adjustments under risk constraints.
• Fine-tuning: SFT for structured output, GSPO RL for alignment with price movements.
• Teacher model distillation to generate high-quality reasoning data for SFT.
• Dynamic threshold annotation using square root of time rule for multi-period labels.

Results & findings

• Agent-DeepSeek-R1 mean CR: 15.50%, Sharpe: 1.71, MDD: -4.09%.
• Agent-Qwen3-8B-FT mean CR: 13.75%, Sharpe: 1.77, MDD: -3.46%.
• Buy-and-Hold mean CR: 10.69%, Sharpe: 0.75, MDD: -11.12%.
• Both agent strategies reduced maximum drawdown by more than half compared to benchmark.
• Fine-tuned small model achieved higher Sharpe ratio than general-purpose large model.
• Agent strategies outperformed benchmark on majority of individual funds.
• System struggled to capture rapid trend accelerations in March-April 2025.
• Synchronous drawdowns occurred during market corrections in April, May, and August 2025.

Data & experiments

• Backtest period: October 2024 to October 2025 (12 months).
• Dataset: 28 Chinese Public REITs listed for over one year as of start date.
• Data sources: Daily prices, volume, announcements, news, CSI REITs Index, 10-year bond yields, stock indices.
• Baselines: Buy-and-hold strategy.
• Metrics: Cumulative Return, Sharpe Ratio, Maximum Drawdown.
• Training data: 12 months of historical REIT data (Oct 2023-Sep 2024) for fine-tuning.

Limitations & caveats

• Backtest limited to 12 months, potentially insufficient for long-term robustness.
• System performance degraded during rapid market corrections and trend accelerations.
• Decision agent uses general-purpose LLM, which may be overly conservative.
• Focus on low-volatility REITs may not generalize to high-volatility equities.
• Transaction costs set at 0.03%, which may not reflect all market frictions.
• Fine-tuning data limited to 12 months, potentially missing long-term regime changes.

Curator notes

Discovery status: trading_agent_focus_promoted. Source links: FinMem: A Performance-Enhanced Large Language Model Trading Agent with Layered Memory and Character Design