Point-in-Time

Definition

A point-in-time view of data contains only what was actually known and available at a chosen past moment. Revisions, restatements and records added later are excluded. Point-in-time correctness means every historical analysis reads data this way, so the past is never judged with hindsight.

In plain terms

Most databases show the latest version of each fact. If a company restates last year’s revenue, today’s table shows the restated number and the original one is gone.

That is fine for reporting on the present. It is wrong for asking “what would we have decided back then?”, because the restated figure did not exist yet. To answer that question you need two things for every value: what it was, and when it became known. Keeping both is what makes a dataset point-in-time capable (see Bitemporal Data for the underlying idea).

Example

The numbers below are synthetic.

A forecast is evaluated “as of March 1”. The usable value is 100, because that is all anyone could have seen on March 1. Using 96 would quietly feed a future correction into the past — the mistake described in Look-ahead Bias.

Common misunderstandings

• “It just means recent data.” No. It means the data as it stood then, however old.
• “A timestamp on each row is enough.” Only if the timestamp says when the value became known, not when the event happened. The two differ more often than people expect.
• “It only matters for numbers that get revised.” It also applies to definitions and membership: which items existed, what a category included, which labels had been assigned.