backtest-validationlisted

# Backtest validation A backtest reports the maximum of a search. The question is never "what was the Sharpe" but **"what was the Sharpe, given how many things were tried, and how much of the sample was independent."** This skill answers both. ## 1. The 2026 stack ```bash pip install purgedcv # purged/combinatorial CV + PSR/DSR. MIT. py>=3.10 pip install skfolio # CPCV composed with sklearn GridSearchCV. BSD-3 pip install arch # SPA / Reality Check / StepM / MCS. THE significance toolkit pip install RiskLabAI # broadest AFML reimplementation. BSD-3. py>=3.12 pip install ml4t-diagnostic # DSR, PBO, CPCV, feature importance. MIT. py>=3.12 pip install jsharpe # PSR, MinTRL + FWER/FDR corrections. MIT. py>=3.11 ``` 🚨 **Do not plan around `mlfinlab`.** Not installable from PyPI; the GitHub source is stubbed (**every function body is `pass`**); licence is proprietary all-rights-reserved; silent since 2023. 🚨 **`fracdiff` is archived** (2023-12) and `requires_python <3.10` — unusable on modern Python. ⚠️ **`pypbo`** is the reference PBO/CSCV implementation but is **not on PyPI** and is **AGPL-3.0**. ⚠️ **`sharpebench`** (v0.17.1, 2026-09-02) covers this whole surface but declares **no repository URL on PyPI** — provenance unverifiable. New and unvetted; do not default to it. ## 2. Cross-validation that does not leak **Any CV that shuffles is catastrophic here.** `train_test_split`, `KFold(shuffle=True)`, default `cross_val_sc