Replication package: Does machine learning add value to trend-following? Evidence from a purged walk-forward evaluation of a Donchian channel breakout system

Replication package for the article "Does machine learning add value to trend-following? Evidence from a purged walk-forward evaluation of a Donchian channel breakout system" (Ettouhami T., El Kabbouri M.). The study applies a stacked machine learning ensemble (XGBoost, LightGBM, logistic regression) as a meta-labeling filter on a weekly Donchian channel breakout rule over a point-in-time universe of liquid US equities, evaluated out-of-sample from 2012 to 2026 under an embargoed purged walk-forward scheme. Contents: - code/ : the full Python analysis code and pinned package versions (requirements.txt, Python 3.14.6) - Weekly data with Universe/ : raw input data — weekly OHLCV bars for US equities (2000–2026), weekly SPY and VIX series, and point-in-time weekly universe membership (source: QuantConnect) - data/ : derived results — the 7,056 scored out-of-sample candidate signals, the trade log, the 15 walk-forward folds, the two 100-draw permutation controls, and all robustness exercises - README.md : description of every file and instructions to reproduce all results (run time about 17 minutes)