Skip to content

Systems and records

Financial restatement version datasets: keeping each revised number

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Finance training data should keep every revised number as its own version instead of overwriting it. Each version records the value, the period it describes, the date it became known, the version it replaces, a reason code, links to evidence and the approver's role. With those fields a reviewer can replay what the books said on any date and reconcile every change to documents and the final reported figure.

Record timeline

Financial restatement

This page covers financial restatement version datasets: keeping each revised number.Illustrative

Key takeaways

  • Keep two dates on every figure: the period it describes and the date it became the official number.
  • Classify the change first; error corrections, estimate changes, principle changes and reclassifications are treated differently in accounting and should be in the data.
  • A short controlled list of reason codes makes revisions countable and comparable across periods.
  • Create a version only when a value or classification changes; reviews that confirm a figure add evidence, not versions.
  • Exclude or escalate revisions tied to investigations, litigation or compensation recovery.

How should finance training data preserve the history of revised numbers?

Store every reported number as a chain of versions that is never overwritten. Each version carries the value, the date it became known, the period it describes, the version it replaces, a reason code from a fixed list, links to the evidence and the name of the person who approved it. A model can then learn both the final answer and how the answer moved.

A financial restatement version dataset is that chain, built for a defined set of facts such as revenue by period, a reserve balance or an accrued liability. It differs from an ordinary ledger export in one way that matters for AI work: it lets a reviewer reconstruct what the books said on any past date and explain each change with documents rather than memory.

The rest of this guide gives the annotation scheme as a field table, a reason-code vocabulary, a worked example, the reconciliation checks a reviewer runs, and the cases where versioning adds noise instead of signal.

Why the current ledger balance hides what the team knew at close

Most finance systems answer the question of what the balance is now, not what it was believed to be at close. Posted correcting entries change the current trial balance, and a later export shows only the corrected total, so the first reported figure and the reason it moved disappear from the extract.

Economic statisticians solved the same problem decades ago. The St. Louis Fed's ALFRED archive stores each release of a data series as a vintage with a real-time period, so a researcher can ask what a figure was on a given date; its documentation uses the January 1990 unemployment rate, first reported as 5.3 percent and later revised to 5.4 percent. Company finance data needs the same two time axes: the period a number describes and the date the number was known.

Public filings show why the distinction matters. The SEC's frames API returns, for each company, the one fact last filed that best fits a calendar period, which is the right default for analysis and the wrong one for training a model on how numbers change. A model trained only on last-filed values never sees that an earlier figure existed.

Which kinds of change deserve different labels?

Label the type of change before anything else, because accounting treats the types differently and a model should too. FASB Statement No. 154 separates three accounting changes (a change in accounting principle, a change in accounting estimate and a change in the reporting entity) from the correction of an error, and defines a restatement as revising previously issued financial statements to reflect the correction of an error.

For public companies, the SEC's compensation recovery rules add a second split inside error corrections: a restatement to correct an error that is material to the previously issued statements, often called a Big R restatement, and a correction of an error that would cause a material misstatement if it were corrected in the current period or left uncorrected, often called a little r restatement. When the board or authorized officers conclude that earlier statements should no longer be relied on, Form 8-K Item 4.02 requires a filing that names the affected periods.

Which kinds of change deserve different labels?
Change classWhat it meansTypical evidenceHow the version is treated
Error correction, material to prior statementsPrior issued statements were wrong and are reissuedNon-reliance filing, amended report, audit committee minutesNew version for every affected period, flagged as restatement
Error correction, corrected in current comparativesPrior figures revised in this period's comparative columnsRevision note in the current report, auditor scheduleNew version with the comparative report as the disclosure reference
Error corrected in the current period onlyImmaterial error booked as an out-of-period entryJournal entry with memo, materiality memoNo new version of the old period; the current period carries a tagged entry
Change in accounting estimateNew information changes an estimate going forwardActuarial or reserve study, approval memoNo revision of prior periods; reason code on the current-period value
Change in accounting principleA different method applied as if always usedPolicy memo, transition disclosureNew versions of prior periods, flagged as retrospective application
Change in reporting entityThe set of entities consolidated changesAcquisition or reorganization documentsNew versions with an entity-scope field that changes
ReclassificationAmounts move between lines with no change in totalMapping change log, presentation memoNew version of each line, linked as a pair that nets to zero

The version and reason-for-change annotation scheme

The asset below is the minimum field set for one version of one fact. A fact is a single number defined by entity, line item, period, basis and unit; every row in the dataset is one version of one fact, and a fact with no revisions simply has one version.

Keep the as-of date and the period as separate fields. Mixing them is the most common reason a revision dataset cannot be replayed.

The version and reason-for-change annotation scheme
FieldContentRuleExample (illustrative)
fact_keyEntity, line item or account, period, basis, unitStable across versions; never editedEntity 02, net revenue, 2025-Q2, GAAP, thousands
version_noSequence number within the factStarts at 1, increments by one, no gaps3
valueThe number as reported or posted at that versionSame unit and sign convention as fact_keyIndex value 97.0 against v1 = 100
known_as_ofDate and time the version became the official figureFrom posting, report or filing date, not export date2025-10-29 report issue date
supersedesThe version this one replacesRequired for every version after the firstv2
change_classOne class from the change tableSingle value; disputes go to the reviewerError correction in comparatives
reason_code and reason_textCode from the controlled list plus one plain sentenceText names the cause, not the personCUT-OFF: shipments invoiced before transfer of control
evidence_refsLinks to entries, workpapers, filings or memosAt least one document per version after v1JE 4471, cutoff test schedule, comparative note
approved_by and approved_onRole of the approver and the approval dateRole, not name, in the licensed copyController, 2025-10-24
disclosure_refWhere the change was disclosed, if anywhereFiling or report section, or noneQ3 report, revision of prior period note

A controlled list of reason codes

Reason codes make revisions countable, which is what lets a model or an analyst learn patterns such as which accounts are revised most often and why. Keep the list short enough that two reviewers assign the same code to the same case, and add a code only when a case fits none of the existing ones.

At a glance

A controlled list of reason codes

  • CUT-OFF: a transaction recorded in the wrong period.
  • LATE-DOC: a source document such as a vendor invoice or bank confirmation arrived after close.
  • CALC: a formula, spreadsheet or allocation error.
  • MAPPING: an account mapped to the wrong line, or a chart-of-accounts remap.
  • POLICY-MISAPPLIED: the right policy was applied wrongly; distinct from a change of policy.
  • ESTIMATE-UPDATE: a reserve, accrual or fair value updated for new information.
  • PRINCIPLE-CHANGE: a new accounting method applied retrospectively.
  • ENTITY-SCOPE: an acquisition, disposal or reorganization changed the consolidated group.
  • AUDIT-ADJ: an adjustment proposed by the external auditor and booked by the company.
  • DUPLICATE: the same transaction recorded twice.
  • FX-RATE: the wrong rate or rate date applied in translation or remeasurement.

Worked example (illustrative): a June cutoff error found in the third-quarter review

The figures in this example are invented to show the mechanics; they describe no real company. A distributor closes June with 412 shipments recognized as second-quarter revenue. Net revenue for the quarter is reported internally and indexed here at 100.

During the third-quarter review, the controller's cutoff test finds that 37 of those shipments left the warehouse in July. The team books a correcting entry, assesses the error as immaterial to the second-quarter statements but material if corrected entirely in the third quarter, and revises the comparative column in the third-quarter report.

  • What a model can learn here: the gap between flash and reported figures, the typical size and direction of cutoff corrections, and which documents settled the question.
  • The exception inside the example: the year-end audit touched the figure without changing it, so it adds an evidence link to v3 instead of a new version. Versions record changes in value or classification, not every review.
Worked example (illustrative): a June cutoff error found in the third-quarter review
VersionKnown as ofValue (index)Reason codeEvidence
v1July close, internal flash101.5none (first figure)Flash report, revenue subledger
v2Second-quarter report issued100.0LATE-DOCCredit memos received after flash, approval by controller
v3Third-quarter report issued97.0CUT-OFFCutoff test schedule, bills of lading, correcting entry, comparative note
v4 (not created)Year-end audit97.0noneAuditor agreed v3; no new version because the value did not change

Can a reviewer reconcile every version to documents, authority and the final state?

Yes, if each version passes four checks before it enters the dataset. A version that fails any of them is either fixed from the source records or excluded, and the exclusion is logged with a reason.

At a glance

Can a reviewer reconcile every version to documents, authority and the final state

The audit profession already enforces this discipline on its own files. PCAOB AS 1215 lets auditors add documentation after the documentation completion date but forbids deleting it, and requires each addition to show the date it was added, who prepared it and why. A revision dataset that applies the same three facts to every version is easy to defend.

  • Chain check: version numbers run without gaps, every version after the first names the version it supersedes, and only the last version is open.
  • Authority check: every version after the first has an approver role and an approval date that falls on or before known_as_of.
  • Document check: every evidence reference resolves to a retained record, and the record's date is consistent with the version's timing.
  • Final-state check: the latest version of each fact ties to the most recent issued statement, filed amendment or closed trial balance for that period; for public filers, an amended filing's cover page carries an amendment flag and, as XBRL US guidance notes, a required amendment description.
  • Replay check: rebuilding the dataset as of any past report date reproduces that report's figures.

Exceptions: changes that should not become separate versions, and cases to set aside

Do not create versions for changes that carry no decision. Automated currency revaluation, recurring reversals of accruals and rounding differences between subledger and report are mechanical; tag them on the entry if useful, but a separate version for each one buries the revisions a model should learn from.

Treat a chart-of-accounts remap as one mapping version applied to many facts, with a single reason record, rather than thousands of independent revisions. Otherwise a model learns that reclassification is the most common event in finance, which is an artifact of the migration.

Set aside, or bring in a finance process expert and counsel before using, any revision connected to an investigation, a whistleblower matter, litigation, a regulatory inquiry or an executive compensation recovery. Those records carry obligations and sensitivities that a data project should not decide on its own. Narrow the scope when approvals or evidence are missing for more than a small share of versions in a period; a revision history with unexplained jumps teaches guessing.

Can revision histories be offered as a licensed dataset?

Revision histories with reasons and evidence are rarer than ledgers, because most companies keep only the corrected balance in an exportable form. Finance teams that license such histories through SourceX get rights review, contracting, preparation and delivery handled as one engagement.

Ownership of the records, permission to share them and contractual licensing rights are separate questions, and confidentiality terms with auditors, lenders or customers can limit what leaves the company. The company's signature on scope and terms comes before any delivery, and the first question, whether the company qualifies, is answered by a brief assessment with no data upload.

Frequently asked questions

What is the difference between a restatement and a revision of prior-period figures?

Under FASB Statement No. 154, a restatement revises previously issued financial statements to correct an error. In practice, a correction material to the prior statements leads to reissued statements, often called a Big R restatement, while a smaller one is corrected in the current report's comparative columns, often called a little r revision. Record both as new versions with different change classes.

Should a change in accounting estimate create a new version of the prior period?

No. A change in estimate is applied going forward, so the prior period's figure stays as reported. Record the new estimate on the current period with an ESTIMATE-UPDATE reason code and link the study or memo that drove it.

How do I reconstruct what the books said on a past date?

Filter each fact to the latest version whose known_as_of date is on or before the date you care about. This only works if known_as_of comes from posting, report or filing dates rather than the date the data was exported, which is why that field needs its own source.

Can public filing data stand in for internal revision history?

Partly. Original and amended filings show reported figures before and after a restatement, and the SEC's APIs expose the facts in each filing. They do not show internal flash figures, late documents, approvals or the workpapers behind a change, which are what make a revision dataset useful for training.

Who should review a revision dataset before it is shared?

A finance process expert who knows the company's close, plus a data quality reviewer who runs the chain, authority, document, final-state and replay checks. Counsel should review anything touching investigations, compensation recovery or confidentiality terms with auditors and lenders.

Sources

  • FASB Statement No. 154 separates changes in accounting principle, estimate and reporting entity from error corrections and defines a restatement as revising previously issued financial statements to correct an error. Source
  • The SEC's compliance guide says recovery policies must cover restatements for errors material to prior statements and errors that would cause a material misstatement if corrected or left uncorrected in the current period. Source
  • Form 8-K Item 4.02 requires disclosure when the board, a committee or authorized officers conclude previously issued statements should no longer be relied upon, including the periods affected. Source
  • The SEC's xbrl/frames API aggregates one fact per reporting entity, last filed, that most closely fits the requested calendar period. Source
  • ALFRED archives FRED data with real-time periods so users can retrieve values as they were known on past dates, using the January 1990 unemployment rate revision as its example. Source
  • AS 1215 forbids deleting audit documentation after the documentation completion date and requires added documentation to show the date, the preparer and the reason. Source
  • XBRL US cover page guidance says AmendmentDescription is required when the AmendmentFlag element is set to true. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies