Payrolls Fell 23,000 and the Revisions Took Another 103,000

The Headline and the Bigger Number Underneath It

American employers cut 23,000 jobs in July. Economists surveyed beforehand had expected somewhere between 83,000 and 95,000 additions, so the miss against the consensus was well over a hundred thousand.

That is the number that led the coverage. It is not the important one.

The same release revised May down by 66,000, from 129,000 to 63,000, and June down by 37,000, from 57,000 to 20,000. Combined, 103,000 jobs that were reported as having existed did not exist.

Put the three months together and the economy added 60,000 jobs across May, June and July. That is an average of 20,000 a month, against an average of 34,000 a month over the preceding twelve.

Why the Revisions Are the Story

A single weak month is noise. This series is volatile, the sampling error around it is large, and one negative print inside a run of positive ones tells you very little.

A revision is different in kind. It does not say this month was weak. It says the picture you had of the last several months was wrong, in a consistent direction, and every judgement made on the basis of that picture was made with bad information.

The direction of the revision matters as much as the size. Revisions that scatter randomly are measurement noise. Revisions that run the same way month after month indicate something systematic in how the initial estimate is produced, and a systematic error persists into the estimates that have not been revised yet.

Where the Initial Number Comes From

Understanding why first prints move requires knowing what the first print actually is.

The payroll figure comes from a survey of business establishments, and it is published on a schedule that does not wait for every response. The initial estimate is built from the responses received by the deadline. More arrive afterwards, and each subsequent month's release incorporates them, which is why the same month is published three separate times before it is considered final.

There is a second source of movement. A survey can only count firms it knows about, so it cannot directly observe jobs created by businesses too new to be in the sample or lost by businesses that quietly closed. That gap is filled by a statistical estimate rather than a count.

That estimate is built on historical relationships, and it is at its least reliable exactly when the economy is turning, because a turning point is precisely when the past relationship between business formation and business failure stops holding. So the initial estimate tends to be most wrong at the moment its accuracy matters most, and the error runs in a predictable direction: too optimistic as an expansion ends.

The Unemployment Rate Went the Wrong Way for the Right Reason

The unemployment rate fell in July, from 4.2 percent to 4.1 percent, in a month when the economy shed jobs.

That is not a contradiction and it is not an error. The unemployment rate is a fraction. The numerator is people without work who are actively looking, and the denominator is everyone working or looking. Someone who stops looking leaves both.

The labour force participation rate fell to 61.4 percent, a level not seen in more than five years. People left the labour force, and their departure lowered the measured unemployment rate by shrinking the pool it is calculated against.

This is the standing reason to read the participation rate alongside the unemployment rate rather than on its own. A falling unemployment rate driven by rising employment and one driven by people giving up look identical in the headline and describe opposite conditions.

Where the Losses Landed

The composition is more informative than the total, because the total is a net figure that nets out a great deal.

Local government education shed 50,000 positions. Retail trade lost 19,000.

The education figure is worth pausing on because a July number in that category is unusually hard to read. Public education employment is intensely seasonal, with staff leaving at the end of a school year and returning at the start of the next, and the reported figure is an attempt to strip that pattern out using historical norms. When the timing of hiring shifts relative to those norms, the adjustment produces a large number that is partly an artefact of the calendar rather than a change in employment.

Retail is easier to interpret and harder to dismiss. It is a discretionary spending business, it employs a great many people, and it responds to household budgets fairly quickly. A decline there is closer to a genuine demand signal.

What is at least as informative is what did not appear. A net figure of negative 23,000 that contains a 50,000 decline in one category and a 19,000 decline in another means the rest of the economy added jobs on balance. The losses were concentrated rather than broad.

That distinction changes the diagnosis. A recession looks like weakness spreading across unrelated industries at the same time, because a general decline in demand touches everything. Two categories falling while the remainder holds up looks like two specific problems, one of which may be a calendar artefact.

The way to tell them apart is to watch whether the losses stay concentrated. If next month's decline is in the same two places, this is a story about public sector budgets and consumer spending. If it has spread into business services, transport, and construction, it is a story about the economy.

What Three Months at 20,000 Actually Means

The three month average is the figure to hold on to, and it is worth converting into something meaningful.

An economy with a growing working age population needs a certain number of jobs each month simply to keep the employment rate steady. Below that level, employment is not keeping pace with the population even when the reported number is positive.

Twenty thousand a month is far beneath any reasonable estimate of that threshold. So a three month stretch at this rate is not slow growth in the labour market. It is a labour market that has stopped expanding, and the only reason the unemployment rate has not risen to reflect it is that the labour force shrank at the same time.

The comparison against the prior twelve months makes the deceleration concrete. Going from 34,000 a month to 20,000 is a decline of roughly forty percent in the pace of hiring, and it happened over a single quarter. That is the kind of change that shows up in the aggregate data long after employers have already felt it, because hiring decisions are made months before they appear in a payroll count.

It also means the recent past looked considerably better than it was while it was happening. Anyone who read the initial May and June prints in real time saw a labour market adding a reasonable number of jobs. The revised versions describe an economy that had already slowed materially by then, and nobody could have known it from the data available at the time.

Two Surveys, Two Answers

Part of the confusion in any payroll release comes from the fact that the two headline numbers are not produced by the same exercise.

The job count comes from surveying employers. The unemployment rate and the participation rate come from surveying households. They are different samples, asking different questions, and they routinely disagree.

The differences are not errors. A person holding two jobs appears twice in the employer survey and once in the household survey. Someone self employed or working informally appears in the household survey and not on any establishment payroll. Someone who runs their own business full time is employed by one measure and invisible to the other.

That divergence is worth watching in its own right. When household employment holds up while payrolls weaken, it often reflects a shift toward self employment and informal work, which is a different economy from one where payroll jobs are simply being added.

The practical instruction is to stop treating the release as one number with supporting detail. It is two independent measurements of a labour market that neither can see completely, and the interesting information is usually in the gap between them rather than in either one.

Why This Does Not Settle the Inflation Question

A weakening labour market is the mechanism by which tighter policy is supposed to work, so it would be reasonable to read this as evidence that the job is done.

The complication is that the transmission runs through wages, and a labour force shrinking at the same time as employment does not obviously loosen the market. Fewer jobs against fewer available workers can leave the balance between them roughly unchanged, and it is that balance, rather than the level of employment, that determines whether wage growth pushes prices along.

The other complication is that a participation decline can reverse. People who stop looking during a soft patch return when conditions improve, and their return raises measured unemployment without anything getting worse. A policymaker looking at 4.1 percent has to decide how much of it is durable and how much is a temporary withdrawal that will show up again later.

How I Read a Payroll Release Now

I read the revisions before the headline. They are printed in the same release and they are frequently the larger number, and reading them first prevents anchoring on a figure that is about to be restated.

I check the participation rate before drawing any conclusion from the unemployment rate, for the reason above.

I look at the three month average rather than the month, because the sampling error on a single print is wide enough that individual months are close to uninformative on their own.

And I treat the first estimate of any month as provisional in a specific way: not merely uncertain, but likely to be revised in whichever direction the economy is currently heading, because that is the direction the statistical estimate of unobserved firms will be getting wrong.

The Case That This Is Less Bad Than It Reads

The gloomy reading has real weaknesses and they should be stated.

The single largest loss in the report is in local government education, where the seasonal adjustment does strange things in July, and stripping that line out changes the character of the month considerably. A report that loses 23,000 jobs including a 50,000 decline in one seasonally awkward category is not obviously a report about broad economic weakness.

Falling participation is also more than a story about discouragement. An ageing population produces a declining participation rate through retirement alone, and that component is demographic rather than cyclical. Reading every decline as people giving up overstates the case.

And the revisions cut the other way too. If the initial estimates are unreliable enough to move by 103,000, then the current month's figure of negative 23,000 is equally unreliable and may itself be revised upward. Treating a provisional number as bad news while treating its provisional nature as evidence of weakness is having it both ways.

The Bottom Line

Payrolls fell by 23,000 in July against expectations of 83,000 to 95,000 additions, and the revisions removed another 103,000 jobs from May and June. The three month average is now 20,000 a month against 34,000 over the prior year, which is a labour market that has stopped growing. The unemployment rate improved to 4.1 percent because participation fell to 61.4 percent, which is the rate falling for the wrong reason. Read the revisions before the headline and the participation rate before the unemployment rate, because in this release both of the secondary numbers were larger news than the ones that led the coverage.