1. Give every observation a version history
A vintage is a dated view of data, not a different economic period. One quarter can have several estimates released at different times. A useful research record therefore identifies both the observation period and the version available at a chosen cutoff. FRED distinguishes its usual latest-information view from historical real-time queries. Neither view is inherently wrong: today's estimate can suit retrospective description, while an earlier decision requires the information then available. The mistake is substituting one view for the other without changing the research question or the label.
2. Understand why estimates change
BEA describes advance, second and third quarterly GDP estimates and later updates that incorporate additional information. Third does not mean permanently unrevisable. Revisions can also arise from seasonal adjustment: BLS explains that recalculated CPI seasonal factors can revise the previous five years of seasonally adjusted indexes. These are distinct mechanisms, so retain revision notes instead of assigning every change the generic description 'correction.' A changed number can reflect broader measurement information rather than a processing mistake. None of these mechanisms establishes how often a particular regime classifier will change.
3. Choose the historical question explicitly
An as-known study reconstructs the eligible vintage at each decision. A fixed-vintage study applies one specified snapshot across history. A latest-vintage study uses the newest downloaded estimates. Name the choice in the study description and retain the snapshot date. An as-known study should also identify whether it uses first releases only or the latest revision available at each cutoff; those policies differ. Selecting only initial releases can intentionally discard revisions that were already public. Historical realism requires the declared information policy, not an automatic preference for the oldest number.
4. Recompute transformations within a snapshot
Revision effects need not be limited to the newest observation. A denominator, rolling average or standardisation window can change too. Build the eligible level series first, then calculate the feature under a frozen formula. Keep units, seasonal status and comparison periods unchanged when isolating revision effects. For a difference feature, record both operands; for a growth rate, record numerator and denominator. A small displayed revision can produce a classification change near a threshold, while a larger revision far from it may leave the label unchanged. Neither observation alone measures model quality.
5. Work through a fictional boundary crossing
Consider a fictional activity feature for an invented quarter. The estimate published on fictional 2030-05-10 is 1.8 units; its revision on fictional 2030-06-10 is 2.1 units. A predeclared threshold assigns 'higher' at values greater than or equal to 2.0. The revision is 0.3 units and changes the retrospective label from lower to higher. A second feature is held fixed. The May decision remains lower in the as-known archive. Rewriting it as higher would answer what the revised data imply, not what the May information supported.
6. Summarise revisions without selecting a winner
Report how many comparable observations changed value, how many labels changed and which boundaries were crossed. Keep the denominator explicit and separate missing-vintage cases from unchanged cases. Inspect signed revisions as well as their magnitudes, because opposite signs can cancel in an average. Compare like-for-like observation periods and revision horizons. A latest estimate remains an estimate, not an independently observed true regime. Revision stability can describe a classifier's sensitivity to changing measurements, but it cannot establish predictive accuracy, economic causality or the value of acting on a label.
7. Preserve both histories and their limitations
Store the original classification and a separately identified revised-data reconstruction. Link their differences to changed inputs, transformations or rules rather than overwriting the earlier record. If a required vintage is absent, mark the gap and narrow the claim; do not quietly fill it with today's value. Record retrieval dates and source definitions so subsequent researchers can distinguish a fresh revision from a changed series concept. This approach permits both historical learning and faithful decision review. It is a research documentation practice, not investment advice or evidence of historical prediction.
Inspect the comparison
| Fictional publication date | Feature units | Label at fixed 2.0 threshold | Role |
|---|---|---|---|
| 2030-05-10 | 1.8 | Lower | Original decision-time estimate |
| 2030-06-10 | 2.1 | Higher | Later revision of the same period |
Primary references
References support the stated methods and data conventions. Original fictional examples are not observations from the cited sources.