Why economic data gets revised

New reports, improved source data, benchmarks and method changes can revise earlier economic estimates.

An estimate can improve after publication

A producer often publishes an initial estimate before all source reports are available. Later responses, corrections and more complete administrative data can change the estimate for the same reference period. The economy did not go back in time; the statistical estimate changed.

This revision is different from a new observation. Comparing August with July describes a period-to-period change. Comparing two published estimates of July describes a revision. Both can occur in one release, and the currently revised July value can differ from the one a reader saw a month earlier.

Different programs revise differently

Payroll employment can receive monthly and benchmark revisions. GDP has advance, second and third estimates plus subsequent account updates. Seasonal-factor re-estimation can revise adjusted history. A program’s unadjusted index may follow a different policy. Do not use one generic final/preliminary label for all these processes.

Method changes require another category. A new classification, population control or geographic definition can change what the series represents. Such breaks may complicate a historical comparison even when individual published observations are correctly copied. Read producer notes and retain versioned definitions.

What a frozen snapshot proves

The September 30, 2026 site snapshot preserves the observations and metadata acquired for this publication. Its raw files and checksums make that acquired dataset reproducible. It does not establish that every figure was available on an earlier date, or supply original agency release times that are absent from source metadata.

An as-known-on analysis needs authentic vintage observations and an explicit information cutoff. Date-only release records cannot prove intraday knowledge. A backtest that silently includes today’s revisions can use information unavailable at the historical decision date. This site does not offer arbitrary pre-acquisition as-of exports.

How updates are handled

Acquisition produces a candidate snapshot. Differences in matching series and periods are reviewed as revisions, while new periods are classified separately. Approved versions replace the visible dataset only after validation. A correction to a site processing error is logged separately from an agency revision. Previously approved snapshots remain frozen rather than being overwritten.

Typical revision schedules

BEA publishes an advance estimate of quarterly GDP about one month after the quarter ends, followed by second and third estimates at roughly monthly intervals. Annual updates, usually released in September, revise several years of history with more complete source data. Comprehensive updates, conducted roughly every five years, can introduce new definitions and methods and revise the entire history, sometimes reaching back many decades.

For payroll employment, BLS revises each monthly estimate in the two following monthly releases as more employer reports arrive. Once a year, payroll levels are benchmarked to counts derived largely from unemployment insurance tax records, with the March level as the reference point. The benchmark revision is published with January data in February and also revises adjusted history through seasonal re-estimation.

How professionals work with revised data

Researchers who study forecasting or policy decisions use real-time datasets, which record what each statistic looked like on each publication date. The Federal Reserve Bank of St. Louis maintains ALFRED, an archive of past vintages of many series, and the Federal Reserve Bank of Philadelphia maintains a real-time dataset for macroeconomists. These sources make it possible to reconstruct what was known at a decision date.

Analysts who track monthly releases often watch the revisions as closely as the new figure. A strong first estimate accompanied by large downward revisions to the prior two months can describe a weaker picture than the headline alone suggests. Reporting the net revision alongside the new estimate gives readers a more complete account of the release. Producers often state that combined net revision in the release text itself, which makes it easy to cite.

Common errors and correct citations

One error is to compare a newly published figure with the originally reported figure for the prior period, instead of with the revised prior figure in the same release. Another is to describe a revision as the economy changing, as in “the economy lost jobs last spring,” when the release revised an estimate of a past period. A revision changes the estimate, not the events.

A third error is assuming that a widely quoted number is permanent. Text written months after a release should check whether the figure has since been revised. A durable citation names the series, reference period, estimate stage where relevant and access date: for example, “Real GDP, second quarter, third estimate, BEA, accessed on a stated date,” so a reader can locate the same vintage later.

Worked example: separating a new change from a revision

Suppose a hypothetical payroll series, in thousands of jobs, first shows 100,000 for May and 100,180 for June. A month later, the next release revises May to 100,020 and June to 100,150.

  1. First reported June change: 100,180 − 100,000 = 180 thousand jobs.
  2. Revised June change: 100,150 − 100,020 = 130 thousand jobs.
  3. Revision to the June monthly change: 130 − 180 = −50 thousand jobs.
  4. Revision to the June level: 100,150 − 100,180 = −30 thousand jobs, smaller than the change revision because May was revised up by 100,020 − 100,000 = 20 thousand.
  5. If a later benchmark lowers the March level by 0.2%, a level of 100,000 would be revised by 100,000 × 0.002 = 200 thousand jobs.

Takeaway: A revision to a level and a revision to a monthly change are different numbers. Report which one you mean and compare figures from the same release.

Frequently asked questions

Why does the government revise jobs numbers?

The first payroll estimate is based on employer reports received by the publication deadline. Later responses are added in the next two monthly releases, and once a year estimates are benchmarked to more complete counts based largely on unemployment insurance tax records. Revisions reflect more complete information, not a change in what happened.

Is the third GDP estimate final?

No. The third estimate is the last of the initial quarterly estimates, but BEA revises GDP again in annual updates, usually released in September, and in comprehensive updates roughly every five years. A quarter’s GDP growth can change years after it was first reported.

Where can I find the originally published numbers?

The Federal Reserve Bank of St. Louis ALFRED archive stores past vintages for many U.S. series, and the Federal Reserve Bank of Philadelphia maintains a real-time dataset for macroeconomists. Agencies also keep archived news releases, which show the figures as published on each release date.

Are revisions a sign of error or manipulation?

Revisions are a planned part of statistical production. Producers publish early estimates for timeliness and update them as fuller source data arrive, under documented schedules and methods. Large revisions can reveal how uncertain early estimates are, which is why producers publish revision analyses and why careful users cite the estimate stage.

What can we help you find?

Type at least two letters to start searching.

Browse economic guides Esc to close