AlphaTransform: A Quantitative Trading Strategy Generation and Backtesting Framework

AlphaTransform is a Reinforcement Learning-based Alpha Generation and Alpha Backtesting system.

This repository contains a formulaic alpha generator using Reinforcement Learning (with a Transformer as the policy network) in alphagen, as well as a backtesting system in backtester that tests alphas generated by the alpha generator that can be used to train and test the alpha generator.

Architecture

Here is a high-level overview of AlphaTransform's architecture:

Usage

Config

The default parameters in config.py works on the input data provided in the data directory, but feel free to change the parameters if necessary (e.g. if using your own dataset or additional operators). The training of the RL agent will depend on these specified parameters.

Training

To train a model on the provided data in the data directory, run:

python3 train_maskable_ppo.py <seed> <loss_metric> <num_steps>

where <seed> is any integer serving as a random seed for reproducible output, and <loss_metric> is a choice of IC, RIC, or Sharpe. <num_steps> refers to the number of steps that the RL training loop will run for, and is an optional parameter (with a default of 100,000).

An example training command is:

python3 train_maskable_ppo.py 42 Sharpe

Testing

To test a previously trained model, run:

python3 test_maskable_ppo.py <checkpoint_path> <num_distinct_alphas_to_generate> 

where <checkpoint_path> is the file path of the saved model checkpoint. <num_distinct_alphas_to_generate> is the number of distinct alphas (with unique loss values) to generate in the batch, and is an optional parameter (with a default of 100).

An example test command is:

python3 test_maskable_ppo.py ./checkpoints/new_Sharpe_42_20231105171143/100352_steps.zip

References

The data provided in the data directory was computed from the Google Finance dataset provided by:

@article{feng2019temporal,
  title={Temporal relational ranking for stock prediction},
  author={Feng, Fuli and He, Xiangnan and Wang, Xiang and Luo, Cheng and Liu, Yiqun and Chua, Tat-Seng},
  journal={ACM Transactions on Information Systems (TOIS)},
  volume={37},
  number={2},
  pages={27},
  year={2019},
  publisher={ACM}
}

The design and representation of the Alpha Generator were inspired and modified from this paper:

@inproceedings{alphagen,
    author = {Yu, Shuo and Xue, Hongyan and Ao, Xiang and Pan, Feiyang and He, Jia and Tu, Dandan and He, Qing},
    title = {Generating Synergistic Formulaic Alpha Collections via Reinforcement Learning},
    year = {2023},
    doi = {10.1145/3580305.3599831},
    booktitle = {Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
}