Skip to the evidence
AI Trading SignalsDefinition / Population / Proof

AI claims / Consumer evidence

AI Features: Vintage, Publication and Receipt Clocks

Separate the date a fact describes from the date and version a model could actually use.

An observation date is not availability

A monthly economic value describes a period, while its release and later revision occur on other dates. A market feature can also be computed only after its source window finishes. Preserve event period, publication time, receipt time, revision identity and decision time separately. ALFRED's vintage-data help supports investigation of revised economic records. It does not supply an arbitrary model's intraday reception log or establish that the value reached that model before a decision.

A file downloaded today can contain corrected historical values. Their old observation labels do not prove that those final values were available to a historical experiment. Ask which vintage and transformation each original prediction used. If the original snapshot was not retained, the review should narrow its reproducibility claim instead of labelling a current reconstruction the original real-time history.

Join data as available, not as later finalised

A feature join needs a declared availability rule and source identity. Do not connect a decision to the later revised value merely because it describes the same economic month. Keep the original join version so another reader can reproduce which rows were eligible at each decision. A documented rule is still different from proof that the vendor executed it. The actual run manifest and receipts would need to establish that stronger implementation claim.

Derived features add their own completion time

A bar close, full-session range or aggregate can be unavailable until its observation window is complete. Retain the window and evaluation clock. A model may use a partial-window feature legitimately if that definition is frozen, but later replacing it with the completed-window value changes the experiment. Missing observations and timezones should remain explicit. The consumer enquiry is whether the inputs were available under the stated method, not whether the final chart looks persuasive.

A source request that can be answered

Ask for the model's feature specification, source versions, availability mapping, retained input row and decision record. If commercial confidentiality prevents a full dataset release, a narrower redacted timing example may answer a limited question. It does not complete a full audit. Credentials, hashes and product manuals cannot substitute for the missing input chronology. A trustworthy review keeps the record's strongest supported level and the next unresolved request together.

Completed fictional evidence case

A fictional value describes month M but is first published at day 12 as 2.0. It is revised on day 20 to 2.4. A decision on day 15 can use the 2.0 release if its receipt is established; it cannot acquire the later 2.4 merely because both values describe M. A modern reconstruction using 2.4 therefore needs to be labelled revised-history research. The example is not a real macro release or model receipt, and the publication makes no claim that a selected provider committed this error.

Evidence objectRecorded valueClaim boundary
Event periodMonth MNot a receipt time
First releaseDay 12 / 2.0Potentially available on day 15
RevisionDay 20 / 2.4Later information
An original claim, its available references and the bounded review
Illustrative record flow; no vendor experiment or customer account is pictured.

Direct sources and reviewed scope

Public documents, not private vendor experiments. The worked cases are invented; source dates do not establish a model-validation engagement.

Related evidence questions