TradingGroup: A Multi-Agent Trading System with Self-Reflection and Data-Synthesis

In one sentence

Summary

The paper addresses the limitations of existing LLM-based trading agents, which often lack inter-agent coordination, structured self-reflection, and access to high-quality, domain-specific post-training data derived from actual trading activities. The authors propose TradingGroup, a multi-agent system designed to enhance decision-making quality and adaptability through a self-reflective architecture and an end-to-end data-synthesis pipeline. The system aims to distill past successes and failures to improve reasoning in analogous future scenarios and to generate instruction data for fine-tuning base LLMs. TradingGroup consists of five specialized agents: News-Sentiment, Financial-Report, Stock-Forecasting, Style-Preference, and Trading-Decision. The News-Sentiment Agent uses an MCP client and reranking to filter and score news. The Financial-Report Agent employs hybrid retrieval (dense and sparse) to extract key indicators from filings. The Stock-Forecasting Agent integrates technical indicators with agent outputs and uses a hybrid gate to constrain predictions. The Style-Preference Agent dynamically selects trading styles based on self-reflection of past performance. A dynamic risk-management module adjusts stop-loss and take-profit thresholds based on volatility and style. The system also features an automated data-synthesis pipeline that logs agent inputs, outputs, and Chain-of-Thought trajectories, labeling them with reward signals for supervised fine-tuning. Backtesting experiments on five stocks (AMZN, NFLX, TSLA, MSFT, COIN) from October 2022 to April 2023 demonstrate that TradingGroup outperforms rule-based, machine learning, reinforcement learning, and existing LLM-based baselines like FinMem and FinAgent. Notably, it achieved a 40.46% cumulative return on AMZN with lower drawdown. The paper also shows that fine-tuning Qwen3-8B with the synthesized data (Qwen3-Trader-8B-PEFT) significantly improves performance over the base model, surpassing GPT-4o-mini on some metrics. Limitations include the reliance on specific backtesting frameworks and the potential for overfitting to historical data patterns, though the authors note the importance of flexible risk configuration.

Research problem

Existing LLM-based trading agents suffer from a lack of inter-agent coordination, insufficient structured self-reflection mechanisms, and limited access to high-quality, domain-specific post-training data that captures both market conditions and agent decisions. This limits their ability to adapt to dynamic market conditions and systematically improve through experience. The paper seeks to address how to integrate performance metrics, agent logs, and risk signals for effective self-reflection and how to automatically collect and label trading-process data for fine-tuning.

Core contributions

• Proposes a novel multi-agent chain trading system with role-specific reasoning capabilities for news, reports, forecasting, style, and decision-making.
• Introduces a self-reflection mechanism that enables multi-level self-correction across multiple agents by distilling past successes and failures.
• Designs a dynamic risk management module that flexibly adjusts stop-loss, take-profit, and position sizing based on trading style and volatility.
• Develops an end-to-end data-synthesis pipeline that automatically evaluates, labels, and collects agent working data for LLM fine-tuning.
• Demonstrates superior backtesting performance against rule-based, ML, RL, and other LLM-based baselines across five real-world stock datasets.

Method

• News-Sentiment Agent: Uses MCP client, Qwen3-Reranker, and vector deduplication to score and aggregate news sentiment.
• Financial-Report Agent: Employs hybrid retrieval (dense Qwen3-Embedding + sparse BGE-M3) and LLM re-ranking to extract financial indicators.
• Stock-Forecasting Agent: Integrates technical indicators (RSI, ATR, SMA) with agent outputs and uses a hybrid gate (hard interception + soft pass) for trend prediction.
• Style-Preference Agent: Uses self-reflection on historical PnL to dynamically select trading styles (aggressive, balanced, conservative).
• Trading-Decision Agent: Merges signals from other agents and account status, using self-reflection on past cases to make Buy/Hold/Sell decisions.
• Risk-Management Module: Calculates dynamic stop-loss/take-profit thresholds based on 10-day volatility and style-specific multipliers.
• Data-Synthesis Pipeline: Logs agent inputs/outputs/CoT, labels them with reward signals (directional accuracy, PnL), and filters high-quality samples for SFT.

Results & findings

• TradingGroup (GPT-4o-mini) achieved 40.46% cumulative return on AMZN, significantly outperforming the best baseline (13.27%).
• On TSLA, AMZN, and COIN, TradingGroup achieved globally optimal overall performance among LLM-based baselines.
• Fine-tuned Qwen3-Trader-8B-PEFT outperformed base Qwen3-8B on all five stocks, with CR of 28.67% on TSLA and 29.11% on NFLX.
• Qwen3-Trader-8B-PEFT surpassed GPT-4o-mini in return metrics on TSLA and NFLX datasets.
• Ablation studies show that removing risk management increases return on NFLX but drastically reduces it on other stocks, highlighting the need for risk control.
• The self-reflection and retrieval components significantly improved performance when combined with risk management.
• The data-synthesis pipeline successfully generated high-quality instruction data for parameter-efficient fine-tuning.

Data & experiments

• Training data: Generated via TradingGroup running on DeepSeek-R1 over two historical windows (2020-2021, 2021-2022), resulting in 1,080 labeled trajectories.
• Test data: FINSABER backtesting framework dataset covering AMZN, NFLX, TSLA, MSFT, and COIN from Oct 2022 to Apr 2023.
• Baselines: Buy & Hold, SMA/WMA Cross, ATR Band, Bollinger Bands, Turn-of-the-Month, ARIMA, XGBoost, FinRL algorithms (A2C, PPO, SAC, TD3, DDPG), FinMem, and FinAgent.
• Metrics: Cumulative Return (CR), Sharpe Ratio (SPR), Maximum Drawdown (MDD), and Annualized Volatility (AV).
• Experiments: Framework comparison using GPT-4o-mini and PEFT experiment fine-tuning Qwen3-8B with synthesized data.

Limitations & caveats

• Backtesting results may not fully reflect real-world trading conditions, such as slippage and liquidity constraints.
• The performance of the self-reflection mechanism depends on the quality of the historical data and the base LLM's reasoning capabilities.
• The data-synthesis pipeline requires significant computational resources for running the agents and generating data.
• The system's performance may vary across different market regimes and asset classes.
• The fine-tuning process might lead to overfitting if not carefully regularized, although the authors used a single epoch.

Curator notes

Discovery status: trading_agent_focus_promoted. Source links: Design and Empirical Study of a Large Language Model-Based Multi-Agent Investment System for Chinese Public REITs; Toward Expert Investment Teams:A Multi-Agent LLM System with Fine-Grained Trading Tasks