71 / 97
Point-in-Time Data
Working definition
Data recorded as it was actually known on each historical date — first-print economic releases, as-of universe membership, unrevised financials — rather than the revised series that exists only in hindsight.
Most historical datasets are the present’s version of the past. Economic series are revised: headline employment and GDP figures are restated in subsequent releases, sometimes drastically, so the “historical” value a standard database shows for a given month is a number that nobody possessed in that month. Corporate financials are restated. Index membership lists are published as they stand today, not as they stood on each date. Prices are retroactively adjusted. A backtest run on such data is simulating a market participant who traded on information from the future — politely, a look-ahead bias; precisely, a fiction.
Point-in-time data is the corrective: every value is stored with the timestamp at which it became knowable, and queries return what was knowable as of the simulated date. The first print of a release, not its final revision — because the first print is what moved the market. The universe as constituted on the day, including the members that later delisted — because their absence is exactly the conditioning that survivorship bias smuggles in. The distinction is institutionalised where the stakes are understood: central-bank archival databases preserve data vintages — the full sequence of what each series looked like at each publication date — precisely because the revised series and the tradable series are different objects.
The discipline is expensive and unglamorous. It multiplies storage, complicates every query, and produces backtests with worse numbers, since the flattering revisions and the buried failures are exactly what it removes. That is the point. A simulation on revised data measures a strategy against a memory of history — tidied, corrected, and survivor-only. A simulation on point-in-time data measures it against something resembling the history that occurred. Maintaining the distinction requires knowing what a dataset contains and when each part of it arrived, which is the concern of tick data provenance extended across every input a strategy consumes.
Commonly confused with
Neighbouring concepts that get used interchangeably, and the distinction that actually separates them.
- Historical data
Most historical datasets are the present's version of the past — revised, corrected, and tidied. Point-in-time data is the past's version of itself. The two disagree precisely where a strategy would have been trading on the difference.
- The revised series
The first print is what moved the market; the revision is what the database remembers. For a backtest the first print is the tradable object and the revision is a fact about publishing history, not about the market anyone traded.
- Tick data provenance
Provenance is about what happened to a dataset since capture — cleaning, backfill, merges. Point-in-time is about what was knowable at each historical moment. You need provenance discipline to achieve point-in-time honesty, and the two are not the same claim.
- A data vintage
A vintage is one snapshot: what a series looked like at one publication date. Point-in-time data is the full sequence of vintages, queried as of a simulated date. Central-bank archival databases keep vintages precisely because the revised series and the tradable series are different objects.
How to measure it in your own data
A definition you cannot test is a definition you have to take on trust. This is the shortest honest route from the concept to a number you computed yourself.
- Records you need
For every input, the value together with the timestamp at which it became knowable — not the timestamp of the period it describes. For revised series that means first prints; for universes it means membership as constituted on the day, including the members that later delisted.
- What you compute
Query the dataset as of each simulated date and compare against what your backtest actually consumed. Any value that differs is a value the simulation had and the market did not.
- What the answer tells you
Expect the honest version to produce worse numbers, and treat that as the measurement working. The flattering revisions and the buried failures are exactly what the discipline removes. A simulation on revised data measures a strategy against a memory of history — tidied, corrected and survivor-only. If switching to point-in-time changes nothing, verify that the switch actually happened.
If this has already cost you
Whether a series has been quietly revised since the dates it describes is a property of the file, and it is checkable.
- Dataset Forensics Assay“Can I trust the data file my research stands on?”Will not establish: That the data is correct. Structure can prove a file defective; it cannot prove it faithful to the market — only a second independent source can, and the report names that limit.
Intake is not open yet, so none of these can be commissioned today. They are listed here so you know the measurement exists and what it would and would not settle — the launch list hears first.
Questions and answers
Why does using revised economic data count as look-ahead bias?
Because the revised figure did not exist on the date the simulation trades it. Headline employment and GDP numbers are restated in later releases, sometimes substantially, so the value a standard database shows for a given month is one nobody possessed in that month. The strategy is reacting to information from the future, politely described.
What is a data vintage?
The full sequence of what a series looked like at each publication date, rather than only its current state. Central-bank archival databases preserve vintages deliberately, because the revised series and the series anyone could have traded are different objects and conflating them destroys the second.
Is point-in-time discipline worth the cost?
Point-in-time discipline is expensive and unglamorous — it multiplies storage, complicates every query, and produces worse backtest numbers. That last part is the argument for it rather than against. The numbers get worse because the flattery is being removed, and a result that survives the removal is worth more than the one that preceded it.
Does this only affect economic data?
No. Corporate financials are restated, index membership lists are published as they stand today rather than as they stood on each date, and prices are retroactively adjusted. Every one of those is the same failure: a value stored under the timestamp of the period it describes rather than the moment it became knowable.
Related terms
Derived from the links this entry makes and the entries that link back to it.
In the research
Point-in-Time Data comes up in ten research notes on this site, and this entry lists three of them.