Backtests Don’t Fail — Data Does

When strategies underperform in live markets, logic is blamed. Parameters are adjusted. Models are retrained. Rarely is the dataset interrogated with equal scrutiny. Backtests rely entirely on historical inputs. If those inputs contain survivorship bias, timestamp drift, silent revisions, or missing sessions, conclusions are compromised before research begins. Mutable historical datasets are especially dangerous. If yesterday’s file can change without explicit versioning, then historical performance cannot be reproduced precisely. Reproducibility is foundational to capital discipline. Without immutable raw archives, versioned derivations, and hash-verified outputs, a backtest becomes narrative rather than measurement. Audit-grade infrastructure requires canonical source retention, documented transformation logic, deterministic rebuild capability, and cryptographic verification of exported datasets. If a dataset cannot be reconstructed bit-for-bit from preserved inputs, then historical performance cannot be defended under inspection.