Seasonal adjustment and data revisions: why the numbers change

Economic numbers are not fixed when they are published. Almost every headline figure markets trade on, from the monthly jobs count to quarterly GDP, is seasonally adjusted to strip out the regular swings of the calendar, and most are revised as better information arrives. The first number you read is an early estimate built on incomplete data; later versions can differ by more than the surprise that moved markets on release day.

None of this is a flaw or a cover-up. It is how statistical agencies balance speed and accuracy. But it changes how the numbers should be read. This guide explains what seasonal adjustment does and how it is done, what "annualized" means, and how the Bureau of Economic Analysis (BEA) and the Bureau of Labor Statistics (BLS) revise GDP and payrolls, using their own published revision records.

Why seasonal adjustment exists

Many economic series swing with the calendar. Retailers hire in November and December and let workers go in January. Construction slows in winter. Schools close in summer, and education jobs drop with them. Months have different numbers of working days and weekends, and holidays such as Easter move between March and April.

These swings can be larger than the changes economists care about. In unadjusted data, the US loses large numbers of jobs every January as holiday jobs end. That is not news, because it happens every year. Seasonal adjustment estimates the typical pattern for each month from many years of history and removes it, leaving a series in which a rise or fall means something beyond the usual calendar.

How the agencies do it

The Census Bureau produces and maintains X-13ARIMA-SEATS, the standard seasonal adjustment program in US official statistics. It fits a regression model with ARIMA errors (regARIMA) to a series, including model selection, and then estimates the seasonal component using either the nonparametric X-11 method or the ARIMA model-based SEATS method (originally developed at the Bank of Spain). It also produces diagnostics on how stable the adjustment is.

Each agency applies its own choices:

• CPI. BLS computes seasonally adjusted CPI with factors from X-13ARIMA-SEATS. It recalculates the factors once a year, with the January CPI released in February, and the new factors revise the seasonally adjusted data for the previous five years. On February 13, 2026, BLS revised seasonally adjusted indexes for January 2021 through December 2025.
• Payrolls. The BLS's Current Employment Statistics (CES) program uses concurrent seasonal adjustment, re-estimating the factors each month with the newest data. It re-specifies its models each year with the benchmark and updates the latest five years of seasonally adjusted data.
• GDP. The BEA publishes its quarterly and monthly national-accounts estimates seasonally adjusted, so a quarter's growth can be compared directly with the one before.

What "annualized" means

GDP growth is reported as a seasonally adjusted annual rate (SAAR): the growth over one quarter compounded as if it continued for four quarters.

Monthly series such as housing starts are also quoted at annual rates: a figure of 1.22 million starts in September 2026 means September's pace, scaled to a full year.

How GDP gets revised

The BEA estimates each quarter's GDP in stages, as source data arrive:

1. Advance estimate, near the end of the first month after the quarter, based on source data that are incomplete or subject to further revision. Q3 2026's advance estimate came out October 29, 2026.
2. Second estimate, about a month later (November 25, 2026, for Q3).
3. Third estimate, a month after that (December 23, 2026, for Q3).
4. Annual update, released in late September, which generally covers at least the five most recent calendar years and brings in newly available major annual source data.
5. Comprehensive update about every five years, which adds periodic source data like the economic census and can change concepts and methods.

The BEA publishes a record of how large these revisions have been. Over 1999–2025, in its October 2026 comparison:

Two lessons follow. First, the average with sign is close to zero: early estimates are not consistently too high or too low. Second, the typical size of later revisions is large compared with the differences markets argue about. A quarter first reported at 2.0% and later revised by that typical 1.3 points would end up at 0.7% or 3.3%. The BEA also reports that the advance and latest estimates agree on the direction of real GDP growth 97% of the time, and on whether growth is above, near or below trend 92% of the time. Early GDP is a good guide to the broad picture, not to the decimal.

How payrolls get revised

The monthly payroll count comes from a survey of about 119,000 businesses and government agencies covering about 622,000 worksites, roughly a quarter of all nonfarm payroll employees. It is revised in two ways.

Monthly revisions. The first estimate is based on the responses received by the release date. As more responses arrive, BLS revises that month's figure in each of the next two monthly reports. The September 2026 report, for example, revised the July and August figures.

The annual benchmark. Once a year, BLS re-anchors the survey to a near-complete count: the Quarterly Census of Employment and Wages (QCEW), based on the unemployment insurance tax records employers file. The March level of each year is replaced with that count, and the months around it are adjusted to match. The process has three dates:

• A preliminary estimate of the coming revision is published in late summer. For March 2026 it came out on August 28, 2026: −79,000 jobs, or −0.1%.
• The final benchmark is built into the data with the January employment report, published in February. For March 2026, that is due in February 2027.
• Seasonally adjusted data for recent years are revised at the same time.

Part of the gap between the survey and the tax records comes from business openings and closings. A new firm cannot be in the sample until it appears on the survey's list of employers, and a firm that closes simply stops responding. BLS uses a birth-death model to estimate the net effect; BLS notes the estimates would be considerably less accurate without it. Because it is a model, it can miss when business formation changes quickly, which is one reason the survey and the tax records can drift apart.

Reading data that will be revised

• Read the revised prior. When a new jobs report comes out, add the revisions to the previous two months to the headline change.
• Prefer trends to single prints. A three-month average of payroll changes, or several quarters of GDP, is far less sensitive to revisions than one data point.
• Know which vintage you are looking at. Data sites such as FRED show the latest revised figures, not what was known on the release day. The St. Louis Fed's ALFRED archive keeps earlier vintages for anyone studying how markets reacted at the time.
• Cross-check with other sources. Payrolls can be compared with the household survey and with weekly jobless claims (Jobless claims: reading the weekly unemployment insurance report); GDP with gross domestic income and with monthly spending and production data.

Pitfalls and limitations

• Residual seasonality. Adjustment is an estimate. If the seasonal pattern shifts, as it can after a large shock, some seasonality may remain, or normal months may be over-corrected for a few years.
• Shocks distort the factors. Extreme months, such as the 2020 pandemic collapse, can pull seasonal factors off unless the agency treats them as outliers.
• Revisions are not errors. Agencies publish early estimates on purpose, knowing they will improve them; the schedule and size of revisions are published in advance or after the fact.
• The first print still moves markets. Prices respond to the information available at the moment. A later revision changes history, not the trades already made.
• Year-over-year versus annualized. A quarter's annualized rate and its year-over-year rate measure different things and can differ widely.

Data concepts compared

How Finabase does it

Finabase does not publish GDP, payroll, CPI or other macroeconomic data, and has no economic calendar; the chart and tables on this page are ours, built from BEA and BLS publications. The app shows markets, not statistics: Treasury yields and stock-index futures are where you can see first-print reactions on release mornings, and the Economy news topic collects reporting on both the releases and later revisions.

Frequently asked questions

What does seasonally adjusted mean?

It means the agency has removed the typical pattern for that time of year, such as holiday hiring or winter construction slowdowns, so that one month can be compared directly with the previous one. US agencies use the Census Bureau's X-13ARIMA-SEATS program or methods built on it.

Why do jobs numbers get revised?

Because the first estimate uses only the survey responses received by the release date. BLS revises each month twice as more responses arrive, then once a year benchmarks the level to unemployment-insurance tax records, which cover almost all employers.

How much does GDP get revised?

From 1999 to 2025, the BEA's advance estimate of quarterly real GDP growth differed from the latest estimate by 1.3 percentage points on average, ignoring direction. The two agreed on whether the economy grew or shrank 97% of the time.

What is the difference between annualized and year-over-year growth?

Annualized growth compounds one quarter's change as if it continued for a year; Q3 2026's 0.83% rise in real GDP becomes 3.4% annualized. Year-over-year growth compares with the same quarter a year earlier.

Summary

Seasonal adjustment removes the calendar's regular swings so monthly and quarterly changes mean something, and GDP growth is quoted as a seasonally adjusted annual rate, so a 0.83% quarterly rise is reported as 3.4%. First estimates are deliberately early: GDP is revised in second and third estimates, in late-September annual updates and in comprehensive updates about every five years, and the BEA's own record shows an average 1.3-point gap between the advance and latest growth rates. Payrolls are revised twice in the following months and benchmarked once a year to tax records; the March 2025 level came down by 898,000 jobs. Read the revised prior, prefer trends, and remember that markets trade the number available on the day.