DEV Community

Cover image for I stopped reading the headline jobs number. Table B-1 tells me something different.
Ava Bagherzadeh
Ava Bagherzadeh

Posted on

I stopped reading the headline jobs number. Table B-1 tells me something different.

#career #watercooler #data #news

The US August employment report came out on September 4. It was good. Total nonfarm payroll employment rose by 162,000, against a prior 12-month average monthly gain of 31,000. The unemployment rate held at 4.1 percent. Economists polled by FactSet had forecast 65,000, so the print came in well above what the street expected.

Then I looked at my own industry's line.

Information employment declined by 23,000 in August, following losses that had averaged 8,000 per month over the prior 12 months. So the national total ran at more than five times its own recent trend, and the information sector fell at nearly three times its own recent trend, in the same month, in the same release.

Inside that 23,000, the report breaks it down further. Computing infrastructure providers, data processing and web hosting lost 8,000. Publishing lost 7,000. Broadcasting and content providers lost 5,000.

If you write software, run infrastructure, or produce content for a living, there is a decent chance your employer is counted in that line. The headline number is not describing you. It is describing an average that includes you, weighted so lightly that you disappear into it.

Why the aggregate cannot describe you

The total is a sum, and sums hide their own composition. In August, food services and drinking places added 59,000 against its own prior 12-month average of 12,000, and local government education added 42,000. Between them, those two categories account for most of the beat.

That is not a criticism of the report. The report is doing exactly what it says. It is a criticism of reading only the first sentence of it.

The comparison that matters to a person is not their industry against the national total. It is their industry against its own recent trend. The two questions have different answers and only one of them is about you.

Health care is the cleanest example in this release. Health care employment went up in August, by 13,000. A positive number. It is also well below health care's own prior 12-month average of 32,000. So a health care worker reading "employment continued to trend up" is reading a true sentence about a sector that hired at less than half its recent pace. A positive number can be a slowdown. You only see that if you know what the sector's own normal is.

Do this in five minutes

Nothing here needs an account, a data subscription, or a spreadsheet. It is a browser and about five minutes, once a month.

  1. Open the current Employment Situation summary at bls.gov/news.release/empsit.nr0.htm. This is the primary source. Every number quoted in every article about the report is in here.
  2. Search the page for your industry's name. The summary text names the industries that moved, and it gives each one its own prior 12-month average in the same sentence. That sentence is the whole comparison, already done for you. This is the step most people skip.
  3. If your industry is not named in the summary, open Table B-1, "Employees on nonfarm payrolls by industry sector and selected industry detail", at bls.gov/news.release/empsit.t17.htm. It carries the detailed industry rows.
  4. In the seasonally adjusted block of that table, take your industry's level for the current month and subtract its level for the same month one year earlier. Both columns are right there. Divide by 12. That is your industry's own average monthly change over the last year, computed by you, from the primary table.
  5. Compare this month's change against that average. Same direction and similar size means this month is ordinary. A large gap in either direction is the thing worth reading about.
  6. Read the revisions paragraph before you conclude anything. It is near the top of the summary.

Six steps, one table, one subtraction. That is the entire method.

Telling a trend from one month of noise

Two habits keep me from over-reading a single print.

The first is the revisions. In this release, June was revised up by 11,000, from +20,000 to +31,000. July was revised up by 44,000, from -23,000 to +21,000. July changed sign. A month that was reported as a contraction is now reported as growth. A first print is an estimate that gets two more passes, and the second pass moved that one by more than double the original figure.

The second is the sampling error. The technical note for this release puts the 90 percent confidence interval for the monthly change in total nonfarm employment at plus or minus 122,000. For the monthly change in the unemployment rate it is about plus or minus 0.3 percentage point. So a month-over-month move of 40,000 in the total is inside the noise band, and a change in the unemployment rate from 4.1 to 4.3 is not clearly a change at all.

BLS does not publish a separate interval for every industry in that note, so I do not have a clean noise band for the information line on its own. What I have instead is the run of months. The report says information losses averaged 8,000 per month over the prior 12 months. That is twelve consecutive months of the same direction. One month at 23,000 could be noise. A year of negative months with a bad one on top is a trend, and it is the only part of this release I would actually plan around.

That is the practical rule I use. One month is a data point. Twelve months in the same direction is a signal. The report gives you both, in the same sentence, for free.

Where I actually apply this

I check my sector's line the morning the report drops, and it changes how I spend the following month. When information is shedding and health care is still net positive but decelerating, I read that as a market where the same resume gets a different response depending on which industry code the employer sits under, and I widen where I look before I widen how much I send.

I build AI Applyd, which submits applications on the twelve hiring systems companies actually run, and none of the five minutes above needs it or mentions it again.

The part I do not know

Information employment has fallen for more than a year while total payrolls kept climbing. I do not know what that line is actually measuring.

The CES classifies an establishment by what the employer does, not by what the worker does. The same backend engineer is counted under information at a publisher, and under professional and business services at a consultancy, and under retail trade at a retailer with an in-house platform team. So a year of information losses could be AI substitution, or the unwinding of a hiring overshoot, or a slow reclassification as software work moves out of software companies and into everyone else. Those three stories are indistinguishable in the top-line series, and they imply completely different things about whether the work is disappearing or just moving.

If you know a way to separate them using published CES or QCEW data, I would genuinely like to read it. I have not found one.

Sources, all primary or directly linked:

Top comments (0)