DIAG-15 · Diagnosis · Invariant — holds by construction, no expiry
Data has a half-life
The law
Under unmodeled stochastic parameter drift with noisy observations, windowed-estimator error is sampling variance (falling in window length) plus staleness error (rising in it); the MSE-optimal lookback is finite and interior, shrinking with drift rate, growing with event sparsity.
Scope
Trailing-window or discounted estimators of drifting parameters (baselines, CVRs, bid landscapes), after deterministic structure (seasonality, known regime changes) is handled separately.
What this licenses you to do
Measure drift as the excess dispersion of rolling estimates over the sampling-noise floor, then shorten the lookback (or reset at known regime changes) until staleness and sampling error balance.