Guides
What is point-in-time data? (bitemporal, explained)
Point-in-time data means every observation carries two timestamps: the period it describes, and the moment it became public. With both, a query can ask what was believed about this period on this date — and get an answer that reflects what anyone could actually have known.
The technical term is bitemporal. One axis is the observation or event time, the other is the release or knowledge time. Most databases collapse the second axis by keeping only the newest value for each period, which is why they can tell you what is now believed about 2019 but not what was believed in 2019.
The distinction is not academic. Statistical agencies revise their estimates on published schedules, companies amend filings, index providers reconstitute membership, and regulators publish holdings weeks after the date they describe. In every one of those cases the first published value and the current value differ, and only one of them existed on the date a historical decision was made.
A useful test for any data source: ask it a question about a past date and see whether the answer depends on when you ask. If today's answer about 2019 differs from the answer you would have gotten in 2019, and the source cannot give you the second one, it is not point-in-time — regardless of how much history it contains.
What's covered
- Two timestamps per observation: what it describes, when it published
- Bitemporal storage keeps every vintage rather than the latest value
- As-of queries filter on the publication timestamp
- Revisions become new rows, never edits to existing ones
- Applies to macro series, fundamentals, holdings and index membership
- Latest-value databases cannot reconstruct vintages after the fact
Questions
Is point-in-time the same as having a long history?
No, and conflating them is the most common misunderstanding. A source can hold decades of data and still store only the current value for each date, which tells you what is believed now about the past rather than what was believed then.
Where does the term hindcast come from?
From meteorology, where a hindcast validates a model by replaying history using only the information available at each moment. It is the same discipline applied to markets, which is where this product takes its name.
Does this matter for AI agents specifically?
More than for humans, because an agent has no intuition that a number looks too good for its date. It will report whatever the data layer returns, with fluent confidence, so the correctness has to be enforced by the storage model rather than by the reader's judgment.
Sources
Related
Join the waitlist for early access.
Join the waitlist. Vote the roadmap. First in, first served.