Definition
The share of features, weighted by model importance, whose offline (training) and online (serving) computed values agree within tolerance on matched entities and timestamps.
The Measurement Gap
There is no metric that tells you training-serving feature skew continuously, which means offline accuracy transfers to production on faith.
Formula
FPI = sum(imp_f * agree_f) / sum(imp_f), agree_f = matched values within tolerance t_f
Proposed formula, pending validation pilots.
At a Glance
Decision Supported
Whether offline accuracy claims transfer to production, and which feature pipelines need reconciliation before the next promotion.