Test enough strategy variants on a short sample and one of them will look brilliant by pure chance. MinBTL (Bailey, Borwein, López de Prado & Zhu, 2014) answers: given N trials, how many years of data does a claimed Sharpe need before it deserves any trust?
The expected maximum Sharpe among N skill-less strategies grows with N: E[max] ≈ (1−γ)·z₁₋₁/ₙ + γ·z₁₋₁/₍ₙₑ₎ (γ ≈ 0.5772). For the claimed Sharpe to exceed what noise alone produces, the sample must satisfy T ≥ (E[max] / SR)² years (with an upper bound 2·ln N / SR²). Double N and the required history grows — the more you search, the more data you owe.
Almost never. Every parameter tweak, universe change, or discarded idea you peeked at counts. Research platforms that let you iterate fast silently inflate N into the hundreds.
Either reduce N (pre-register fewer, better-motivated trials), demand a higher Sharpe bar, or add breadth (more independent assets = more effective data). Do not shorten the question to fit the answer.
Same mathematics, opposite direction: MinBTL fixes the trust level and solves for required data; the DSR fixes your data and solves for the trust level.