backtest-overfittinglisted

# Backtest overfitting **Seven guards in this repo ask whether a backtest cheated. This one asks the question that is still open after they all pass: the backtest is clean, and it is *the maximum of a search*.** Two numbers answer it. **PBO** asks how often the in-sample winner lands in the bottom half out of sample — 0.5 means the winner is a coin flip. **MinBTL** asks how many years of data a claimed Sharpe needs before the honest trial count stops explaining it on its own. Every number below is printed by `scripts/overfitting.py` (numpy + scipy, seeds fixed and printed, **24 s**). Nothing here is quoted from a paper without being reproduced. ## 1. CSCV, exactly as defined ✅ Source: Bailey, Borwein, López de Prado & Zhu, *The Probability of Backtest Overfitting*, Journal of Computational Finance 20(4), 2016 — read from the authors' hosted preprint (`davidhbailey.com/dhbpapers/backtest-prob.pdf`; the published JCF text is paywalled and was not reached). Algorithm 2.3: 1. Build a **T x N** performance matrix — one column per configuration you tried, rows aligned on the same calendar. 2. Split the **rows** into an even number **S** of disjoint contiguous blocks. 3. Form all **C(S, S/2)** ways of choosing S/2 blocks as in-sample; the complement is out-of-sample. Both halves are the same size — that is what "symmetric" means, and it is why IS and OS performance are directly comparable. 4. `n*` = the configuration with the best IS performance. Find its **rank** amo