Replication Package for "Observation Frequency as a Trading-Environment Design Variable in Cointegration-Based Deep Reinforcement Learning for Cryptocurrency Statistical Arbitrage"

This repository contains the replication package for the paper “Observation Frequency as a Trading-Environment Design Variable in Cointegration-Based Deep Reinforcement Learning for Cryptocurrency Statistical Arbitrage.” The study examines whether intraday observation frequency should be treated as a methodological design variable in cointegration-based deep reinforcement learning cryptocurrency statistical arbitrage. The empirical framework reconstructs the same economic trading scenario across five observation frequencies: 15 minutes, 30 minutes, 1 hour, 2 hours, and 4 hours. Each scenario preserves a six-day formation window and a one-day trading horizon, while the number of observations, reward intervals, and trading opportunities varies with the sampling interval. The package supports a matched frequency-sensitivity experiment using DQN, PPO, and A2C agents across four fixed cryptocurrency asset panels, two action-space specifications, and three random seeds. The full experimental design produces 360 DRL runs. The package includes scripts and archived outputs required to reproduce the scenario construction, experiment grid, performance tables, hypothesis tests, and figures reported in the manuscript. The replication package is organized to support two levels of reproducibility. First, a fast reproduction workflow recreates the manuscript tables and figures from archived output files. Second, a full replication workflow downloads raw Binance OHLCV data and reruns the complete frequency-sensitivity experiment from raw data. Raw Binance OHLCV files are not included in the archive because of file size considerations, but download scripts are provided. The package includes: Windows CMD scripts for fast reproduction and full raw-data replication; Python scripts for scenario construction, experiment-grid generation, DRL experiment execution, and hypothesis testing; archived source-output files used to recreate the manuscript tables and figures; frequency-sensitivity results for the 360-run matched experiment; Friedman repeated-measures tests and Wilcoxon-Holm pairwise comparison outputs; documentation files, including README, package map, codebook, data-availability statement, requirements file, and environment file; metadata files and checksums for package validation. This record is intended to improve transparency, reproducibility, and reuse of the empirical workflow. The materials are provided for academic research and replication purposes. Veliota DrakopoulouHigher Colleges of Technology, United Arab EmiratesEmbry-Riddle Aeronautical University, United StatesORCID: 0000-0002-1670-8033Email: [e-mail] Suggested Keywords for Zenodo Deep reinforcement learning; cryptocurrency statistical arbitrage; observation frequency; temporal discretization; cointegration; trading-environment design; DQN; PPO; A2C; intraday trading; turnover; frequency sensitivity; Binance; replication package; reproducible research