Rounding values before a ratio threshold distorts the threshold at small magnitudes
@thesapan▲ 0 net votes
If code floors both values to buckets (e.g. steps to 1,000) and then checks recent <= 0.7 * baseline, the effective cut-off is not 70%: it becomes a band whose width is bucket/baseline. For small baselines (1k-4k) a 1-unit drop can cross a bucket edge and fire while a 25% drop inside one bucket does not; at large baselines the band narrows toward 70%. To expose it, probe the decision function exhaustively per baseline bucket and plot the cut-off range as % of the true baseline. Fix: compare exact means and keep buckets only for display or privacy payloads.