Why job numbers keep getting revised down

February’s payroll report showed a gain of 275,000 jobs, alongside a 167,000 downward revision to the previous two months. Forecasters have spotted a pattern at the Bureau of Labor Statistics. Strong first prints are followed by sizeable cuts, which raises doubts about how far real-time jobs data can be trusted.

The biggest cut fell on January. Its first estimate of 353,000, which jolted bond markets a month ago and helped push back expectations of rate cuts, now stands at 229,000. The same report showed the unemployment rate rising to 3.9 per cent from 3.7 per cent, a less cheerful signal that came from a different survey altogether.

Pay told a softer story too. Average hourly earnings rose by just 0.1 per cent in February, after January's jump, which suggests that the earlier surge in wages owed something to quirks in the data. Taken together, the revision, the rise in unemployment and the slower pay growth describe a labour market cooling gently rather than one running hot.

Fewer firms fill in the form

The main culprit is the BLS establishment survey, whose response rate has dropped from over sixty per cent to barely forty per cent in a decade. First estimates lean heavily on imputation and on the birth-death model, which guesses how many businesses have opened. With fewer firms responding, early prints tend to overstate hiring at turning points and make the economy look stronger than it is.

How a first print is built

The establishment survey asks employers how many people were on their payrolls in the pay period that includes the 12th of the month. Many firms reply late, so the first estimate uses the responses in hand and imputes the rest, assuming that missing firms behaved like those that answered. Two further monthly revisions fold in the late replies. Once a year the whole series is benchmarked to the Quarterly Census of Employment and Wages, a near-complete count drawn from unemployment-insurance tax records, which can move the level by hundreds of thousands of jobs.

The birth-death model fills a gap the sample cannot see. New firms are not in the survey until they are found, and firms that close simply stop replying, so the BLS adds an estimate of net business creation based on past patterns. That works when the economy is steady. At a turning point the model keeps adding jobs from firms that were never born. If late responders are disproportionately struggling firms, the imputation flatters the first print as well.

Why markets trade the wrong number

The asymmetry in market reactions compounds the problem. Bond and currency traders respond within seconds to the first estimate, because it is the only new information on the day. Revisions arrive a month or two later, bundled into another report, and rarely move prices on their own. So January's figure of 353,000 pushed up two-year Treasury yields sharply when it was published, while the downgrade to 229,000 was absorbed without much fuss. If first prints are biased upward, markets systematically price a stronger economy than exists, and correct only slowly.

The Fed is less exposed, since it looks at many indicators and has warned against overreacting to any single month. Jerome Powell told Congress this week that cuts would probably be appropriate at some point this year, once officials are more confident about inflation. A labour market that is weaker than the headline numbers suggest would strengthen that case.

What happened last time

The clearest precedent is 2008. As the economy slid into recession, the first estimates understated the pace of job losses month after month, and the annual benchmark later removed hundreds of thousands of jobs from the record. The Fed and investors spent much of that year reading a labour market that looked sturdier than it was. The pattern now is milder, but the direction is the same.

A later episode shows the same bias outside recession. In August 2019 the preliminary benchmark cut 501,000 jobs from the level of employment, as the quarterly census revealed that hiring had been slower than the monthly survey suggested during a period when the Fed was already debating cuts. The Philadelphia Fed now publishes its own early benchmark using the census data, and its recent estimates have pointed to weaker hiring than the official series.

The two surveys have also drifted apart. The household survey, which produces the unemployment rate, has shown much weaker employment growth than the payroll survey over the past year. Each has flaws: the household survey is smaller and noisier, while the payroll survey counts jobs rather than people, so one worker with two jobs appears twice.

The case against

Revisions do not always go down. In the recovery from the pandemic they often raised earlier figures, and a single month's cut can reverse the next. The Congressional Budget Office has also raised its estimates of recent immigration sharply, which suggests that strong payroll gains partly reflect a bigger labour force rather than statistical error. That is a fair point about the level of hiring, but it does not explain why the revisions keep leaning one way.

The indicators to watch are the preliminary benchmark revision the BLS publishes in late summer, the quarterly census data as they arrive, weekly jobless claims and the unemployment rate. A continued rise in unemployment alongside strong payroll prints would be the clearest sign that the first estimates are overstating the labour market.

US jobs figures are getting less reliable on first release. Investors who trade the headline number should expect it to be revised down, particularly once the economy starts to turn, and the late-summer benchmark is likely to confirm that hiring over the past year was weaker than first reported.