The two-clock problem.
A company announces on June 20 that it migrated platforms on June 1. A researcher discovers the announcement on June 24. The reported event date is June 1; the researcher’s observation date is June 24.
For an analysis of what was known on June 10, this record was not yet available to that researcher. Placing it into a June 10 snapshot would use future knowledge.
Historical views need an availability rule.
A point-in-time view should make its cutoff explicit. Include only observations available by that time, while retaining the date a source says the underlying event happened.
This distinction matters even when the event date is accurate. Accuracy about an event does not make information available earlier than it was collected.
Unknown dates should stay unknown.
A vendor disappears between two collection runs. Without a dated statement, the removal may only be bounded to an interval. The second collection date is not necessarily the cancellation date.
Store the last supporting observation and the first changed observation. Leave the event date unspecified or represent an interval instead of inventing precision.
Corrections are part of the history.
A company match can be corrected, a source withdrawn, or a signal reclassified. Replacing the old record in place makes it difficult to reproduce earlier research.
A better record preserves the original observation, the revision time, and the reason for the change. Researchers can then choose between current best knowledge and what was known at a historical cutoff.
Questions to ask of any historical dataset.
- Which timestamp controls inclusion in a historical snapshot?
- Are reported event dates distinguished from collection dates?
- Are missing dates and observation gaps visible?
- Can revised classifications be reconstructed?
- Did source coverage change during the analysis window?