Official source code of KDD 2025 paper: AlphaAgent: LLM-Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay

AlphaAgent is an autonomous framework that effectively integrates LLM agents for mining interpretable and decay-resistant alpha factors through three specialized agents.

• Idea Agent: Proposes market hypotheses to guide factor creation based on financial theories or emerging trends.
• Factor Agent: Constructs factors based on hypotheses while incorporating regularization mechanisms to avoid duplication and overfitting.
• Eval Agent: Validates practicality, performs backtesting, and iteratively refines factors via feedback loops.

This repository follows the implementation of RD-Agent. You can find its repository at: https://github.com/microsoft/RD-Agent. We would like to extend our sincere gratitude to the RD-Agent team for their pioneering work and contributions to the community.

• Create a new conda environment with Python (3.10 and 3.11 are well-tested in our CI):

  conda create -n alphaagent python=3.10

• Activate the environment:

  conda activate alphaagent

• # Install AlphaAgent
  pip install -e .

• First, clone Qlib source code for runing backtest locally.

  # Clone Qlib source code
  git clone https://github.com/microsoft/qlib.git
  cd qlib
  pip install .
  cd ..
  

• Then, mannully download Chinese stock data via baostock and dump into the Qlib format.

  # Download or update stock data from 2015-01-01 until NOW from baostock
  python prepare_cn_data.py
  
  cd qlib
  
  # Convert csv to Qlib format. Check correct paths before runing. 
  python scripts/dump_bin.py dump_all ... \
  --include_fields open,high,low,close,preclose,volume,amount,turn,factor \
  --csv_path  ~/.qlib/qlib_data/cn_data/raw_data_now \
  --qlib_dir ~/.qlib/qlib_data/cn_data \
  --date_field_name date \
  --symbol_field_name code
  
  # Collect calendar data
  python scripts/data_collector/future_calendar_collector.py --qlib_dir ~/.qlib/qlib_data/cn_data/ --region cn
  
  
  # Download the CSI500/CSI300/CSI100 stock universe
  python scripts/data_collector/cn_index/collector.py --index_name CSI500 --qlib_dir ~/.qlib/qlib_data/cn_data/ --method parse_instruments

• Alternatively, stock data (out-dated) will be automatically downloaded to ~/.qlib/qlib_data/cn_data.

• You can modify backtest configuration files which are located at:

  • Baseline: alphaagent/scenarios/qlib/experiment/factor_template/conf.yaml
  • For Newly proposed factors: alphaagent/scenarios/qlib/experiment/factor_template/conf_cn_combined.yaml
  • For changing train/val/test periods, first remove all cache files in ./git_ignore_folder and ./pickle_cache.
  • For changing the market, remove cache files in ./git_ignore_folder, ./pickle_cache. Then, delete daily_pv_all.h5 and daily_pv_debug.h5 in directory alphaagent/scenarios/qlib/experiment/factor_data_template/.

• For OpenAI compatible API, ensure both OPENAI_BASE_URL and OPENAI_API_KEY are configured in the .env file.
• REASONING_MODEL is used in the idea agent and factor agent, while CHAT_MODEL is for debugging factors and generating feedbacks.
• Slow-thinking models, such as o3-mini are preferred for the REASONING_MODEL.
• To run the project in a local environment (instead of Docker), add USE_LOCAL=True to the .env file.

• Run AlphaAgent based on Qlib Backtesting Framework.

  alphaagent mine --potential_direction "<YOUR_MARKET_HYPOTHESIS>"

• Alternatively, run the following command

  dotenv run -- python alphaagent/app/qlib_rd_loop/factor_alphaagent.py --direction "<YOUR_MARKET_HYPOTHESIS>"

  After running the command, log out and log back in for the changes to take effect.

• Multi-factor backtesting

  alphaagent backtest --factor_path "<PATH_TO_YOUR_CSV_FILE>"

  Your factors need to be stored in a .csv file. Here is an example:

  factor_name,factor_expression
  MACD_Factor,"MACD($close)"
  RSI_Factor,"RSI($close)"
  

• If you need to rerun the baseline results or update backtest configs, remove the cache folders:

  rm -r ./pickle_cache/*
  rm -r ./git_ignore_folder/*

• You can run the following command for our demo program to see the run logs. Note than the entrance is deprecated.

  alphaagent ui --port 19899 --log_dir log/

If you find this work helpful, please cite our paper:

@misc{tang2025alphaagentllmdrivenalphamining,
      title={AlphaAgent: LLM-Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay}, 
      author={Ziyi Tang and Zechuan Chen and Jiarui Yang and Jiayao Mai and Yongsen Zheng and Keze Wang and Jinrui Chen and Liang Lin},
      year={2025},
      eprint={2502.16789},
      archivePrefix={arXiv},
      primaryClass={cs.CE},
      url={https://arxiv.org/abs/2502.16789}, 
}