Replication Package for Cointegration-Based DRL Cryptocurrency Statistical Arbitrage

This replication package supports the study “Asset-Universe Design and Action-Space Granularity in Cointegration-Based Deep Reinforcement Learning for Cryptocurrency Statistical Arbitrage.” The package contains the code, configuration files, saved model-summary outputs, manuscript-ready tables, figures, validation logs, and documentation required to reproduce the reported manuscript tables and figures. The study evaluates a cointegration-based deep reinforcement learning framework for cryptocurrency statistical arbitrage using four predefined asset universes, three DRL agents, two action-space specifications, and a non-DRL COIN benchmark. The empirical design uses 30-minute USDT-denominated Binance spot-market data, with a training period from 2023-01-01 to 2024-12-10, a separation window from 2024-12-11 to 2024-12-17, and an out-of-sample testing period from 2024-12-18 to 2025-05-17. Package contents The archive includes: Python scripts for input validation, table recreation, figure generation, result validation, and Binance kline download scaffolding. Configuration files defining the study design, asset universes, universe-label crosswalk, action spaces, transaction-cost setting, training/testing periods, and robustness seeds. Saved source-output files used to recreate the manuscript tables. Manuscript-ready CSV tables covering asset universes, model results, binary-versus-granular action-space comparisons, robustness statistics, seed completion checks, and universe-label crosswalks. PNG figures summarizing best-strategy performance, return/drawdown profiles, action-space effects, and robustness ranges. Documentation files describing the package structure, data folder, code execution order, raw-data layout, and files not used. Reproduction logs and SHA-256 checksums for verification. Reproduction scope This package reproduces the manuscript tables and figures from saved model-summary outputs. It is designed to support transparent verification of the reported empirical results and figure-generation workflow. Raw Binance 30-minute OHLCV files are not bundled in the package. Trained DRL model checkpoint files are also not included. Instead, the package provides the asset-universe configuration files and a downloader scaffold that can be used to reconstruct raw OHLCV data from Binance Public Data, subject to the source provider’s terms of use. Main reproduction steps From the package root, the manuscript outputs can be recreated by running: python code/00_validate_inputs.pypython code/02_recreate_tables.pypython code/03_make_figures.pypython code/04_validate_results.py On macOS or Linux, the full reproduction workflow can also be run with: bash run_all.sh On Windows, use: run_all.bat Raw data reconstruction To list the required Binance Public Data URLs without downloading files, run: python code/01_download_binance_klines.py --all --dry-run To download monthly 30-minute USDT spot kline files for all configured universes, run: python code/01_download_binance_klines.py --all --start 2023-01 --end 2025-05 Downloaded files should be stored under data/raw_ohlcv/ using the folder structure described in data/raw_ohlcv/README_RAW_DATA.md. Software requirements The package is intended for Python 3.10 or later. The minimal listed dependency is: matplotlib>=3.7 A Conda environment file is also provided as environment.yml. Important notes The package is intended for table and figure replication from saved result summaries, not for full retraining of all DRL models from raw market data. Users who wish to extend the study or rerun the full modeling pipeline should first reconstruct the raw OHLCV data using the provided downloader scaffold and then implement or connect the relevant DRL training pipeline. No raw Binance market data are redistributed in this archive. Users are responsible for ensuring that any downloaded market data are used in accordance with Binance Public Data terms and their institutional requirements. Suggested Zenodo metadata Resource type: Dataset or Software, depending on your journal’s preference. I would choose Software if you want to emphasize the scripts; choose Dataset if the journal mainly wants replication materials. Version: 1.0.0 Keywords: Cryptocurrency statistical arbitrage; deep reinforcement learning; cointegration; asset-universe design; action-space granularity; DQN; PPO; A2C; pairs trading; reproducibility; Binance OHLCV; robustness analysis Related identifier: Add the article DOI once available. For now, use “is supplement to” only after the manuscript has a DOI. Suggested citation text after Zenodo DOI is issued: Author Surname, Initials. (2026). Replication package for cointegration-based DRL cryptocurrency statistical arbitrage (Version 1.0.0) [Software and data]. Zenodo. https://doi.org/10.5281/zenodo.20624264