AlphaEval

The implementation of AlphaEval: A Comprehensive and Efficient Evaluation Framework for Formula Alpha Mining.

Overview

This repository contains implementations of various factor mining models and the AlphaEval evaluation framework. The codebase is organized into two main components:

1. Factor Mining Models: Algorithms for discovering trading factors.
2. AlphaEval Evaluation Model: A backtesting and evaluation framework to assess the performance of generated factors.

Acknowledgements

This project reuses ideas and code from the following open-source projects, to whose authors we extend our sincere thanks:

• gplearn
• AlphaGen
• AlphaForge
• AlphaQCM

Data Preparation

In a manner similar to AlphaGen, we leverage Qlib for data storage. and pull our data from the free, open-source BaoStock service. After installing Qlib and baostock, run the script data_collection/fetch_baostock_data.py to download the data. If it is invalid, there is also other data preparation method on the website Qlib

The next, Modify the correspoding path/to/your/qlib_data in all python files to the data you downloaded.

Factor Mining Models

The following factor mining models have been implemented or reproduced by the authors of this project:

• gplearn (including Random Baseline)
• AutoAlpha
• AlphaEvolve
• Fama
• AlphaAgent

Running Instructions for the above models:

python gplearn.py --start_time 2010-01-01 --end_time 2019-12-31 --population_size 1000 --hall_of_fame 50 --n_components 10 --generations 5
python autoalpha.py --start_time 2010-01-01 --end_time 2019-12-31 --population_size 1000 --hall_of_fame 50 --n_components 10 --generations 5
python alphaevolve.py --start_time 2010-01-01 --end_time 2019-12-31 --population_size 1000 --hall_of_fame 50 --n_components 10 --generations 5
python fama.py
python alphaagent.py

The code for the following open-source projects is used directly from their original repositories. For setup and usage instructions, please refer to the README files in their respective folders. Copyright remains with the original authors:

• AlphaGen
• AlphaForge
• AlphaQCM

AlphaEval Evaluation Model

Once you have generated a set of candidate factors, you can evaluate their performance using the AlphaEval framework located in backtest/modeltester. A simplified working example is provided in the Jupyter notebook:

backtest/test.ipynb

Special Note: For the AlphaEvolve project, we have created a custom my_qlib to support new operators such as “RelationRank” incorporating the additional operators introduced in the AlphaEvolve paper. During testing, please use my_modeltester alongside it.