Point-in-Time Financial Data: An AI Research Guide

Point-in-time financial data preserves the version of a fact available at a specified historical moment, including both its reporting period and availability date. In this AllMind research guide, Microsoft's segment recast illustrates why a correct historical number can still be the wrong input for reconstructing an earlier decision. A current filing may restate an older period using revised definitions. Preserve the original and revised records separately; a source citation alone cannot establish what an investor could have known at the time.

This educational teardown uses Microsoft's published segment tables to show the problem. We compare the same fiscal year under two disclosed presentations and calculate the resulting difference in an apparent growth rate. This is an analysis of source documents, not a test of any data vendor or an investment backtest.

One fiscal year, two defensible segment tables

Microsoft's FY2024 release and its FY2025 release both contain revenue for the year ended June 30, 2024. The consolidated total agrees. Several segment values differ because the later table uses a recast presentation.

Sources: Microsoft's FY2024 segment table and FY2025 segment table, accessed September 22, 2026. Differences are our calculations. Download the comparison data.

Microsoft explains that it announced changes to segment composition in August 2024 and recast prior segment information for its FY2025 reporting. Its accounting note says the changes did not affect consolidated financial statements. This is a segment-presentation change; describing it as an accounting-error restatement would misstate the source. Microsoft FY2025 annual report, Note 1

An AI answer that cites the later table can therefore return a different FY2024 segment figure from an answer citing the original release. Neither source is necessarily wrong. The unresolved question is which presentation the analyst intended to use.

The comparison also has a historical limit. Opening an old release page today is not the same as holding a timestamped capture made on its publication date. These two pages demonstrate the version difference. They do not, by themselves, prove every intermediate date on which Microsoft or a data provider first made the recast series available.

A mixed-basis growth rate changes the story

The later table reports FY2025 Productivity and Business Processes revenue of $120,810 million. Comparing that amount with the recast FY2024 value produces growth of approximately 13.10%. Comparing it with the older presentation of FY2024 produces an apparent 55.43% increase.

Both calculations divide the FY2025 value by a FY2024 value and subtract one. Only the first uses the matched presentation disclosed in the FY2025 table. The second combines different segment definitions. The arithmetic runs correctly; the comparison does not describe like-for-like growth.

This is a useful failure case for a research workflow because a model can generate the second answer with two genuine citations. A reviewer who checks only whether the cited numbers appear on the linked pages may miss the inconsistency. The necessary check is semantic: do the figures measure the same business perimeter?

A current operating analysis will often use recast history to improve comparability. A reconstruction of what an analyst knew earlier needs the information set available at that earlier cutoff. Those are different research questions, and the data request should name which one it answers.

Preserve the dates that answer different questions

An observation date describes the economic period. A publication date identifies when a source disclosed the fact. A provider's availability timestamp records when that service delivered or indexed it. A revision timestamp marks a later change. Depending on the job, an analyst may need all four.

This is a proposed data contract for research, not a claim that every provider exposes each field. A daily dataset may be sufficient for a monthly review while being inadequate for reconstructing an intraday decision. Match the temporal precision to the question before evaluating the service.

When a filing appears after the market closes, a date-only cutoff can also be ambiguous. Preserve a timezone and release time when the decision depends on information available during the trading session. Avoid inventing that precision from a dataset that provides only calendar dates.

Why a latest-fact API does not solve historical reconstruction

The SEC's XBRL frames endpoint selects a last-filed fact that most closely fits a requested calendar period. That is useful for collecting comparable periods, but it is not a promise to reconstruct what was known at a past decision time. The SEC separately documents company facts and filing submissions. SEC API documentation

For a historical exercise, retain the filing identity and the permitted source version. Do not assume that requesting an older fiscal year automatically excludes later amendments or recast disclosures. The year being measured and the date of knowledge remain separate even when an API returns both in one object.

Raw facts also need selection rules. The same concept can appear in multiple filings, reporting intervals, units, or dimensional contexts. A quarter-derived value may use two cumulative observations. Every input to that subtraction must obey the intended basis and historical cutoff.

The general procurement checks belong in our AI-ready financial data guide. The narrower point here is that a current feed and a historical-vintage store answer different questions. Verify the storage and selection behavior on a source with a known revision before relying on an availability claim.

The same issue appears in macro series and estimates

Economic data offers a clear reference model. The Federal Reserve Bank of St. Louis documents real-time periods for retrieving information as it existed historically. ALFRED supports vintage views of economic series. Those capabilities help distinguish a present estimate of an earlier period from the value available when that period was first reported. FRED real-time periods and ALFRED help

Consensus estimates need an analogous distinction between the period being forecast and the time the estimate was observed. A current estimate for next fiscal year cannot automatically stand in for the estimate held before an earnings announcement. The contributor population, adjustment basis, and fiscal-period mapping may also matter.

Keep these dimensions in the request. State whether the exercise needs the latest comparable history, the earliest published figure, or the exact vintage available at a declared cutoff. Ask for revisions separately when the change itself is the research signal.

For an estimates-specific specification, use the consensus record guide. A single phrase such as “historical consensus” is too broad to establish the snapshot rules needed for an event study.

Give the research system a question it can answer honestly

A useful historical request names the issuer, metric, fiscal period, units, definition, and cutoff. It also states the source hierarchy and how the system should handle a missing vintage. If the required historical version is unavailable, the output should identify that gap rather than substitute today's value.

The answer should return its calculation inputs alongside its conclusion. For a growth comparison, ask the reviewer to verify that both periods use the same segment perimeter. For a historical decision reconstruction, verify that every cited source was available by the cutoff. These checks can reach different conclusions about the suitability of the same fact.

There is an additional limitation with general language models: restricting retrieved documents does not prove the model's prior knowledge excludes later events. A carefully bounded document exercise can test source discipline. It cannot, by itself, establish a clean historical trading experiment free of future information.

The same caution applies to universe selection. Reconstructing the data for today's surviving companies does not recreate the universe an investor could have selected years ago. Preserve historical constituents and selection rules when the analysis moves from individual facts to a strategy backtest.

A practical acceptance exercise

Use a source pair with a disclosed version change, such as the Microsoft tables above. Request the original presentation and the later comparable presentation separately. Then request a growth comparison and inspect whether the system combines the intended versions.

Keep a short review record with four outcomes: version identified, source located, definition reconciled, and cutoff supported. Mark a missing availability timestamp as unresolved. Do not turn a correct number into proof of historical reproducibility when its availability remains unknown.

If the output is going into Excel, carry the version labels into the workbook and apply the AI model audit checks. A correct citation in the research chat is insufficient if the exported model loses the basis that made the comparison valid.

Sources and methodology

We inspected the linked Microsoft, SEC, and St. Louis Fed pages on September 22, 2026. The tables contain issuer-reported historical values; differences and growth rates are our calculations using unrounded inputs. We did not audit a vendor's point-in-time database, establish the first availability of every recast observation, or run an investment strategy. The example establishes a version-selection problem and a method for investigating it.