1. Archive the decision record, not just its colour
A historical label without its inputs and rule cannot explain what happened. Preserve a decision record containing the cutoff, selected observations, vintage identities, transformed features, rule version, output and any abstention reason. Distinguish the time the record was created from the historical time it describes. A reconstruction created today is not a contemporaneous record merely because it carries an old date. The archive should let a reader determine whether an entry was originally observed, later reconstructed or corrected after a processing error. Those statuses answer different evidential questions.
2. Separate data identity from method identity
Give the input snapshot and classification method separate identifiers. A method specification should include feature formulas, units, thresholds, equality rules, missing-input policy and any scaling window. A snapshot should identify observation periods and selected versions. FRED's vintage-date endpoint lists dates when a series gained new values or revisions, excluding releases that left its data unchanged. That distinction matters: a vintage list is not a complete event log. Keep calendar and receipt evidence separately when the research question requires proof that an expected release was checked.
3. Preserve a chain of derivation
W3C's PROV overview frames provenance around the entities, activities and people involved in producing information. Applied to a research archive, the useful chain is source snapshot, eligibility selection, feature calculation, classification and recorded output. Preserve the relationships between these items, not just a folder full of exports. This is a documentation recommendation rather than a claim that any particular archive implements the W3C model. A reviewer should be able to locate the exact input used by a feature and the exact feature used by a label without guessing from filenames.
4. Correct by addition rather than silent replacement
When a value is revised or a defect is found, retain the earlier record and append a successor with a reason and link to the affected entry. Separate source revision, method change, processing correction and presentation-only change. A corrected result may be the appropriate current view while the earlier record remains necessary for decision review. Avoid describing every successor as a better forecast. A correction can fix a coding error, and a revision can incorporate later information; neither establishes that the corrected result was available when the original decision was made.
5. Use a fictional two-by-two reconstruction
The table crosses two invented data snapshots with two invented rules. Snapshot V1 contains activity 1.9; V2 revises it to 2.1. Rule R1 assigns higher at 2.0 or above; R2 uses 1.7. The second feature is held fixed. V1 with R1 is lower, while each other combination is higher. Only the V1/R1 cell represents the assumed original record. The other cells are retrospective comparisons. Changing data and rules simultaneously would conceal that either change is sufficient to cross the boundary in this fictional case.
6. Test replay and integrity as separate properties
An integrity check asks whether archived content has changed relative to a stored reference. A replay check asks whether the preserved inputs and method reproduce the output. A content digest helps with the first question but does not, by itself, prove when a record existed or whether its inputs were historically available. A successful replay answers the second question but cannot validate an economic interpretation. Preserve independent timing evidence where needed. If either check fails, record the failure and its scope rather than replacing the reference until the result appears consistent.
7. Explain what changed in a revision report
A useful archive comparison reports the affected decisions, changed input versions, changed method fields and resulting label differences. Include unchanged outputs when their underlying inputs changed, because label stability can hide substantial measurement movement. Identify unresolved historical gaps and avoid claiming that an incomplete archive proves continuous availability. Readers should see which view is decision-time history and which is current reconstruction. These practices make disagreements inspectable without converting documentation into evidence of forecasting success. They are research recordkeeping principles, not investment advice or an instruction to act on any classification.
Inspect the comparison
| Fictional snapshot | Activity units | R1: higher at 2.0 | R2: higher at 1.7 |
|---|---|---|---|
| V1 | 1.9 | Lower: assumed original record | Higher: method comparison |
| V2 | 2.1 | Higher: revised-data comparison | Higher: both changes |
Primary references
References support the stated methods and data conventions. Original fictional examples are not observations from the cited sources.