TradingAgents A-Share Edition: Localization Practice of Multi-Agent Collaborative Quantitative Trading System

TradingAgents A-Share Edition is an open-source multi-agent collaborative quantitative trading system localized for the characteristics of China's A-share market. Based on the original US stock TradingAgents framework, it adapts to A-share features such as price limit (up/down), T+1 trading system, and retail-dominated market sentiment. Through reconstructing the data collection layer, adding a policy agent, optimizing technical and sentiment agents, and adjusting execution layer rules, it achieves quantitative trading decisions more tailored to the A-share market.

Why is an A-share specialized version needed?

The A-share market has unique characteristics: 1. Price limit system (10% for main board / 20% for STAR Market) affects liquidity and risk control; 2. T+1 system changes the time scale of position management; 3. High retail investor proportion leads to large emotional fluctuations; 4. Data availability and quality have specificities (e.g., stock suspension, financial data standards).

Review of the Original TradingAgents Framework

The original framework simulates the division of labor in an investment team: fundamental analysts (financial reports/macroeconomics), technical analysts (price/volume), sentiment analysts (public opinion), risk managers (stop-loss/drawdown), and a decision committee (comprehensive decision-making). However, it was designed for US stocks and needs to be adapted to A-shares.

Overall Architecture

Data Collection Layer (A-share specialized) → Feature Engineering Layer (localized factors) → Multi-Agent Analysis Layer (including new policy agent) → Decision Fusion Layer → Execution Layer (A-share rule adaptation)

Key Module Transformations

1. Data Collection: Connect to A-share data sources such as East Money/10jqka.com.cn/Tushare, integrate characteristic data like Dragon and Tiger List/Northbound Funds, and automate cleaning (suspension, rights adjustment, etc.);
2. New Policy Agent: Track regulatory/industry/macroeconomic policies and interpret policy impacts;
3. Technical Agent Optimization: Add capital flow analysis (main force/northbound/margin trading and short selling) and A-share characteristic indicators (limit-up factor, chip distribution);
4. Sentiment Agent: Replace with Chinese NLP models (ERNIE/MacBERT) and monitor Chinese communities like Xueqiu/Guba;
5. Execution Layer: Support A-share order types (price cage, call auction) and strengthen T+1/price limit risk control.

Decision-Making Process

1. Information Collection: Each agent independently analyzes data;
2. Opinion Statement: Submit position (long/short/neutral), confidence level, and arguments;
3. Cross-Examination: Agents challenge each other;
4. Consensus Formation: The decision committee integrates opinions to generate trading instructions.

Dynamic Weight Adjustment

Adjust agent weights according to market conditions: volatile market (technical + sentiment), trending market (fundamental), policy-sensitive period (policy), earnings season (fundamental).

Backtesting Specialization

• Slippage Model: Distinguish between large-cap/small-cap liquidity and consider inability to trade due to price limits;
• Cost Model: Include A-share fees such as stamp duty/transfer fee/commission;
• Market Impact: Estimate the impact of large orders and evaluate strategy capacity.

Live Trading Recommendations

• Broker Interfaces: Recommend programmatic APIs like Zhongtai XTP, Hang Seng PTrade;
• Risk Control: Independent risk control module, set stop-loss lines, support manual intervention;
• Operation and Maintenance: Log system records decision-making process, monitors agent status, and conducts regular backtesting verification.

Applicable Scenarios

• Quantitative research teams: Rapidly iterate strategies and test agent configurations;
• Individual traders: Lower the threshold for system construction;
• Teaching and research: Cases of AI in quantitative investment.

Usage Recommendations

• Start with small-scale simulation trading/small capital;
• Continuously optimize agent analysis capabilities;
• Regularly adjust parameters to adapt to market changes;
• Establish a sound risk control system (stop-loss, position upper limit).

Limitations

• High dependence on data quality;
• Models may overfit historical data;
• High demand for computing resources;
• Changes in regulatory policies affect strategy effectiveness.

Future Directions

• Technology: Integrate large language models (LLM), reinforcement learning optimization, federated learning;
• Functionality: Expand to multiple assets (funds/futures/options), global market coverage, personalized agent customization.