When Payrolls Matter Most, They May Mislead

The headline monthly payroll estimates are produced by the BLS through the Establishment Survey, part of the Current Employment Statistics (CES) program. The survey collects responses from roughly 119,000 businesses and government agencies, covering about 622,000 individual worksites.

Because these figures are derived from a sample rather than a full count of employment, they are subject to statistical estimation and revision. The closest approximation to a comprehensive count of jobs is the Quarterly Census of Employment and Wages (QCEW), which compiles administrative records from unemployment insurance filings and covers roughly 97% of US employment. For this reason, the QCEW serves as the benchmark to which nonfarm payrolls are periodically revised.

Each year, the BLS benchmarks the CES data to the QCEW. In this process, the payroll level for the previous March is compared with the QCEW estimate, and the difference is distributed evenly across the prior 12 months, rather than applying it all at once (a linear “wedging” adjustment). The result is that the level of nonfarm payrolls is brought into alignment with QCEW data for March of the given benchmark year.

In recent years, these benchmark revisions have been relatively large and persistently negative. Over the past three years alone, adjustments have reduced previously reported payroll employment by a combined 1.75 million jobs.

Many attribute these discrepancies to methodological issues with the survey itself, such as the Birth-Death Model, which estimates the net impact of new and closing businesses not yet captured in the monthly data. However, the pattern in revisions suggests a broader explanation.

If Establishment Survey estimates of nonfarm payrolls were independent of economic conditions, we would expect revisions to be randomly distributed over time. This is not what we observe in the data.

As shown in Exhibit 1, revisions exhibit clear regime dependence, tending to be positive during periods of labor market expansion and negative during periods of deterioration.¹

This analysis focuses primarily on the period after 2003, when CES methodology underwent significant changes. Notably, a similar pattern appears when extending the analysis back to 1955 across multiple methodological regimes (see Exhibit 2).

To examine this relationship more formally, labor market regimes are defined using payroll momentum.

Periods of labor market expansion are defined as observations where the six-month moving average of payroll growth is greater than or equal to +75,000 jobs per month. Periods of labor market contraction occur when that measure falls below +75,000.²

From 2003 through 2025, the average monthly revision during expansionary periods was +7,000, while the average revision during contractionary periods was –18,000.

A Welch two-sample test confirms that the difference is statistically significant, meaning it is unlikely to be due to chance (see Exhibit 3).

Additionally, testing each regime independently shows that revisions during expansions are not statistically different from zero, while revisions during contractions are statistically below zero at the 5% significance level (see Exhibit 4).

Taken together, these results suggest that the asymmetry in revisions is driven primarily by downward adjustments during weakening labor market conditions, rather than systematic upward revisions during expansions.

One important caveat is that monthly benchmark revisions are derived from an annual level adjustment and distributed using the linear wedging technique described above. The implication is that monthly revisions within a benchmark year are mechanically linked by construction and not fully independent. Accordingly, statistical inferences should be interpreted with appropriate caution.

Monthly revisions were chosen for this analysis as opposed to the annual headline figures to better illustrate regime dependence and cycle timing. However, the same general pattern appears in annual benchmark adjustments where no monthly wedging occurs. This suggests the result is not merely an artifact of the distribution procedure but reflects a more persistent cyclical tendency in the benchmark process itself (see Exhibit 5).³ 

If payroll estimates systematically overstate employment late in the cycle, relying on headline figures may lead investors to underestimate recession risk and misjudge the timing of policy shifts. Payroll data could therefore be an unreliable real-time indicator at precisely the point in the cycle when investors and policymakers need it most.

Payroll growth in 2025 was already extremely weak by historical standards, with the U.S. economy adding just 116,000 jobs for the year. Historically, years with fewer than 300,000 jobs added have either coincided with recessions or immediately followed them. This suggests the labor market may already be closer to a late-cycle phase than headline payroll figures imply.

Initial payroll data should therefore be interpreted with caution during late-cycle environments, particularly when other indicators point to slowing economic momentum. What appears as resilience may instead be a lagging reflection of conditions that have already begun to weaken. The jobs numbers can make the economy look healthy right when it’s actually starting to turn.

Footnotes:

1. Federal Reserve Bank of St. Louis, “All Employees: Total Nonfarm Payrolls (PAYEMS),” ALFRED (Archival Federal Reserve Economic Data), accessed March 1, 2026, https://alfred.stlouisfed.org/series?seid=payems.
2. Alternative thresholds ranging from +50,000 to +120,000 produce similar results, indicating that the regime dependence is robust to reasonable variations in classification.
3. For archived CES benchmark articles, see US Bureau of Labor Statistics, “CES Benchmark Revisions (Historical Archive),” Current Employment Statistics, https://www.bls.gov/web/empsit/cesbmkarch.htm