A plain-English guide to the NFP report.

The Bureau of Labor Statistics releases two surveys on the first Friday of every month, and the market collapses them into one number: "the jobs report". They are not one thing. They are the Current Employment Statistics survey (CES), which produces the nonfarm payrolls headline, and the Current Population Survey (CPS), which produces the unemployment rate. The two surveys measure adjacent things, by different methods, with different sample sizes, and they can disagree for extended periods without either being wrong. This is the explainer on what each one is, the gotchas in the construction, and the pieces of the report that move FX disproportionately to the headline.

Two surveys, one release

• CES, the establishment survey. Each month the BLS samples roughly 119,000 businesses and government agencies, covering about 622,000 individual worksites. The survey asks how many people were on the payroll for the pay period that included the 12th of the reference month. It produces the nonfarm payroll change (the headline), the average hourly earnings figure, the average weekly hours, and a detailed industry breakdown.
• CPS, the household survey. Each month the US Census Bureau interviews roughly 60,000 households for the BLS. It asks the household whether each working-age member was employed, unemployed, or out of the labour force during the reference week. It produces the unemployment rate, the participation rate, the employment-to-population ratio, and demographic breakdowns (age, race, education, gender).

Two structural differences fall out of the definitions. First, CES counts jobs; a person with two jobs is counted twice. CPS counts people; a person with two jobs is counted once. In any given month with rising multiple-jobholding, CES rises faster than CPS. Second, CES excludes the self-employed, unpaid family workers, household workers, and farm workers; CPS captures all of them. In months when self-employment is rising fast, CPS shows employment gains that CES doesn't.

The two estimates side by side

The birth-death model

The CES sample is a frame of existing businesses. By construction, it cannot capture a business that opens for the first time during the reference month, because that business is not in the sample frame yet. To address this, the BLS uses a statistical model to estimate the number of jobs created by new business formation and to net out the jobs lost when businesses close. The model is called the net birth-death model, and it is the single largest source of methodological controversy in NFP.

In a typical month the net birth-death adjustment adds roughly 50,000 to 150,000 to the seasonally-adjusted headline figure. The contribution is largest in the spring and smallest in the winter, with a seasonal pattern that follows the formation cycle. The model is benchmarked once a year against actual administrative-records data from the unemployment-insurance system, and the benchmark revision is almost always smaller than the headline error bars implied by the model. In years when small-business formation deviates sharply from history, the benchmark revision can be very large; the March 2024 benchmark revised the prior year down by roughly 800,000 jobs.

Seasonal adjustment

The headline NFP figure is seasonally adjusted using the X-13ARIMA-SEATS method, which decomposes each series into trend, seasonal, and irregular components and reports the trend plus irregular. Two things to know about how this affects the read.

• January and July are the noisiest months. Year-end retail and summer hiring patterns are large relative to the trend, so the seasonal adjustment is doing more work in those months. A 50k surprise in January is half as informative as a 50k surprise in May or November.
• The seasonal factors themselves get revised. Every February's release includes a seasonal-factor revision that can re-shape the prior year's trajectory. Reading old prints uncritically is hazardous.

Revisions cadence

Three layers of revision sit underneath each NFP release.

1. Monthly revisions. The release covers the most recent month plus revised data for the previous two months. The two-month revision can run ±100k routinely; ±200k is unusual but not rare. Reading the headline without the revisions is reading half the report.
2. Annual benchmark. Every February, the BLS replaces the model-based estimates with the administrative-records universe count for the prior March. The benchmark can rewrite a year of trajectory.
3. Quinquennial Census revisions. Every five years the underlying population controls in CPS are rebuilt from the Decennial Census, which can produce level shifts in the unemployment-rate series.

What FX traders actually read

The headline drives the immediate price reaction; the revisions and the composition drive the durable one. A 200k headline with negative 80k of revisions is a 120k print, and that is what an experienced desk responds to within ninety seconds.

Three sub-components of the release matter more for the dollar than the headline:

• Unemployment rate. Comes from CPS. A tenth-of-a-percentage-point change at the headline NFP level shifts the OIS-implied policy path more than a 50k surprise on the establishment side. The OIS curve reaction is the cleanest read.
• Average hourly earnings. Reads through to the PCE forecast for the next two to three months. A hot AHE is the strongest single signal that the Fed will not be able to cut on schedule.
• Hours worked. Aggregate income equals employment times wages times hours. A flat headline with rising hours is rising aggregate income. The hours component is where the cleanest read of consumer-spending capacity sits.

What the report does not tell you

1. The state of the labour market in real time. The reference week ends roughly three weeks before the release; by the time the print lands, the data is twenty days old. The Beveridge curve and the weekly jobless-claims series fill in between releases.
2. Sectoral churn. The headline aggregates everything; the same number can come from healthcare + hospitality or from manufacturing + tech. Read the industry table.
3. Wage formation by skill. AHE is an average across the workforce; it can be moved by composition effects rather than wage growth. The Atlanta Fed Wage Growth Tracker, which follows individual workers' wages, is the cleaner read.
4. Hidden slack. The U-6 unemployment rate (includes marginally attached and part-time-for-economic- reasons workers) is roughly twice U-3 and is the better read of true slack in soft labour-market regimes.

Related reading

• Today's end-of-week analysis applies this directly: End of week: the thesis missed. Here is the post-mortem.
• The companion piece on ADP: A plain-English guide to ADP vs nonfarm payrolls
• The Beveridge-curve angle on the same labour data: A plain-English guide to the Beveridge curve
• The aggregating framework: A plain-English guide to economic surprise indices