$\text{Alpha}^2$: Discovering Logical Formulaic Alphas using Deep Reinforcement Learning

Link to paper

This repository contains pseudocode and algorithms for the paper "$\text{Alpha}^2$: Discovering Logical Formulaic Alphas using Deep Reinforcement Learning". It does not contain a runnnable version of $\text{Alpha}^2$, but provides the design principals and code structures.

Code Structure

• utils: utility functions for logging and loading configs
• computation_data.py: Generates a data file for the experiment ro run
• run.py: main file for running the experiment
• run.sh: script to start an experiment: first generate computation data, then start the runner
• configs configuration files
• trainer.py: definition of MCTS and network trainer actors for ray
• expression contains definition of the environment, including:
  • evaluate.py defines teh evaluation function
  • legal_actions.py calculates the legal actions when expanding an MCTS node
  • meta_data.py meta data of stock/futures market
  • operands.py definition of operands
  • operators.py definition of operators
  • tokens.py tokens wrap the implementation of operators, and implements a "validity_check" function for legal action check
  • port.py avoid ray recursive import
  • structure.py defines the structure of tokens, tree nodes, dimensions and values
  • tree.py defines the structure and computation of expression trees
• mcts contains MCTS and network related code, which is an modificated version of alphadev

Cite this work

@article{xu2024textalpha2,
    title={$\text{Alpha}^2$: Discovering Logical Formulaic Alphas using Deep Reinforcement Learning},
    author={Feng Xu and Yan Yin and Xinyu Zhang and Tianyuan Liu and Shengyi Jiang and Zongzhang Zhang},
    journal={arXiv preprint arXiv:2406.16505},
    year={2024}
}