REFERENCE BUILD 01 / 2026-10-16OFFLINE FIXTURES · EXPLICIT CONTRACTS · REPRODUCIBLE OUTPUT
Investment
Research APIs
Time & revisions

Build an as-of join that explains what it can prove

Join observations to decision cutoffs using availability evidence, revision rules and coverage checks rather than observation dates alone.

Investment Research APIs · 16 October 2026 · Reference edition 1

An as-of join selects records under a declared information policy. Its output is only as strong as the availability evidence, revision coverage and tie rules behind that policy; a generic cutoff does not guarantee point-in-time correctness.

Specify the decision information set

Begin with decision identifiers, cutoff timestamps, required series and the accepted availability bases. State whether date-only records are permitted and how same-day ambiguity is handled. Then define whether the latest eligible observation period or a specific target period is needed for each series.

A merge keyed only by observation_date answers a period-alignment question. It does not answer when the value became usable. A current annual value dated December 31 can be incorrectly joined to a January decision even if that value was first retrieved much later. The query must carry availability evidence into the selection logic.

Trace a synthetic selection

Consider a synthetic series with January value 100 available on February 10 and a revision to 101.2 available on March 10. A February 20 date-level cutoff selects the earlier vintage if both versions are retained and the daily comparison policy is appropriate. A March 20 cutoff may select the revision.

Now add a February observation of 102 available on March 15. A March 20 query asking for the most recent known period selects February 102, while a query asking specifically for January selects January 101.2. Both results can be correct because the target-period rule differs. Record that rule instead of hiding it in a sorting expression.

Enforce a partial order, not a convenient sort

First filter by permitted availability evidence and cutoff. Next resolve versions within the same observation identity using documented revision fields. Finally apply the required observation-period rule. Combining all three operations into a single latest-row sort can select a newer period whose version was not yet eligible.

When candidates tie on date-only availability, an accession or revision key may distinguish identity without establishing time order. Keep multiple candidates or return an ambiguity error unless the mapping has a defensible tie rule. Lexicographically sorting identifiers does not create chronological evidence that the provider never supplied.

In the supplied selector, different values at the latest eligible date cause rejection. Equal-value ties use lexicographic provenance only for deterministic output, without claiming chronological precedence. Selection retains one version for each observation identity.

Carry coverage into the result

Suppose a synthetic decision requires three series but only two have eligible observations. Return coverage two of three with the missing series identified. Do not quietly calculate a three-factor signal from two inputs or substitute a current value to make the row complete.

Staleness also needs a declared rule. An eligible value can be too old for the intended analysis even though it passes the cutoff. Measure age using the relevant observation period and expected publication cadence, not only retrieval age. A freshly downloaded five-year-old observation should not satisfy a requirement for a recent annual measure.

Retain the proof and the remaining uncertainty

Export the selected revision keys, cutoff, availability basis, mapping version and excluded-candidate reasons. A reviewer should be able to replay the selection without access to an undocumented notebook state. For a retrieval-based World Bank snapshot, explain that earlier public availability is not established by the retained record.

FRED real-time dates and SEC filing dates offer useful source-specific evidence, while RFC 3339 helps represent exact system cutoffs. They do not remove historical coverage gaps or guarantee that a whole backtest is causal. The fixture workbench demonstrates deterministic selection and explicit ambiguity using invented records; live research needs a separately evidenced historical information set.

Keep in the decision record

  • Apply availability, revision and target-period rules separately.
  • Return missing and stale inputs with the selected records.
  • A deterministic cutoff result is not a universal point-in-time guarantee.

Source notes

  1. FRED API V1 — Real-Time Periods
  2. RFC 3339 — Date and Time on the Internet: Timestamps
  3. SEC — EDGAR Application Programming Interfaces

Primary specifications support the interface descriptions. All downloadable response fixtures are synthetic. Access and rights checks are dated in the source inventory.

Continue the analysis

Version 1 · Published 16 October 2026. Inspect the method · Open the contract lab

Search the technical reference