Statistical agencies publish early and correct later. A first estimate is produced from partial information under a deadline, and subsequent releases incorporate late responses, corrections and methodological updates. The number is a draft, and the market almost always trades the draft.
This creates a specific and underappreciated analytical hazard. A backtest run on the current version of a series uses numbers that did not exist at the time the trade would have been placed. The model is being shown information from the future, and the effect is not small: revisions can change the sign of a weekly change, which means a strategy conditioned on the sign is being tested on data that would not have been available.
Revision behaviour is itself informative and worth measuring. A series whose revisions are large relative to its typical change is one whose weekly print carries little information; a series with small revisions can be taken more nearly at face value. That ratio is a property of the series and it is checkable.
Revisions also come in two kinds that behave differently. Routine revisions incorporate late data and tend to be small and unbiased. Benchmark or methodological revisions change the definition or the estimation approach and can shift a whole history at once, which breaks any model calibrated on the old basis. The first is noise to be accounted for; the second is a discontinuity to be detected.