Tearing apart the BLS jobs number machine

This morning CNBC will put a number on screen and the market will react to it as if it were a measurement. It isn't. It's an estimate of an estimate, built on top of a model that has systematically overstated job creation for three years running and just got caught doing it again.

The mainstream read is that April payrolls will confirm whether the labor market is holding up under the Iran war energy shock. That's the question everyone is asking. It's the wrong question. The right question is whether the number that hits the screen this morning has any reliable relationship to the number the BLS will quietly revise into existence a year from now — and based on three consecutive years of evidence, the answer is: probably not.

This isn't a clerical error. It's a structural feature of how the BLS builds the number. The establishment survey samples roughly 119,000 businesses — covering about 26% of all nonfarm payroll jobs. First-closing response rates have dropped from around 70% before the pandemic to approximately 60% today, according to Michael Horrigan, a former 33-year BLS veteran. The headline number you'll see in forty-seven minutes is based on 60% of a sample that covers 26% of the workforce. The BLS extrapolates the rest using statistical models — including one called the birth-death adjustment, which is where the mechanism gets interesting.

Every month, the birth-death model adds jobs for businesses it assumes were born and created workers — businesses that don't yet show up in the survey frame because there's an unavoidable lag between a company opening and its appearance on the unemployment insurance rolls. The model is calibrated on historical patterns of business formation. In a normal economy, this is defensible. In an economy where business formation is slowing and energy costs have doubled, it runs the same calculation it always runs. And for three years, it ran that calculation in one direction and was wrong in the same direction every single time.

Numbers. No narrative.

Strip the one-time reversals out of March and the underlying economy was adding roughly 51,000 jobs per month. That's the trend the market isn't looking at. The consensus for April's print is 55,000 — but ADP's April private sector count came in at 109,000, and the ADP-to-BLS divergence has been so consistent that Raymond James issued a specific warning this week: do not read the ADP print as predictive of the BLS number. The two surveys have decoupled.

Here's the mechanical flow as it runs through the system today:

The Employment Cost Index for Q1 2026 printed at +0.7% for private sector wages — annualizing to roughly 2.8% — which is below the Fed's threshold for alarm, and also below inflation. Real wages are flat to slightly negative. That's the actual labor market condition entering today's print. Initial jobless claims for the week ending April 25 came in at 189,000 — the lowest since 1969 — which sounds like a tight labor market right up until you remember that the people who have already been laid off aren't filing new claims, and that a low-firing environment and a high-hiring environment are mechanically different things. This economy has one of them.

The weak link isn't the jobs number. It's the transmission from the jobs number to the Fed's posture — and the fact that the market is pricing that transmission off a data series with a three-year track record of being wrong by an average of 403,000 jobs per year.

The Fed is holding at 3.50 to 3.75%. The market prices a 25% probability of a hike and roughly 30% odds of even one cut by year-end. A strong April print today — anything north of 120,000 — will push those probabilities further toward no cut in 2026, and the 10-year Treasury yield, already sitting at 4.40%, will reprice up. That repricing transmits directly to the 30-year fixed mortgage rate, currently near 6.8%. Each 10 basis points on the 30-year fixed adds approximately $65 to the monthly payment on a $400,000 mortgage. That's not abstract. That's the cost of buying the house.

Here's the specific fault line. The birth-death model is calibrated on business formation patterns during periods of normal or expanding economic activity. When business formation slows — as it does when fuel costs double, credit tightens, and consumer confidence hits 74-year lows — the model keeps adding phantom births based on the historical average. It is, structurally, a lagging indicator of business optimism dressed up as a real-time count of workers. I've seen this play out before, most clearly in 2007 and 2008, when the birth-death model kept printing positive monthly additions to construction and finance employment while the actual loading docks and trading desks were quietly going dark.

The second weak link is the survey response rate, which I haven't seen discussed once in the last three months of payrolls coverage. When 40% of the sample doesn't respond in time for the initial print, and the BLS imputes those non-responses using prior patterns, you get an estimate that is mechanically biased toward resembling last month's trend. In a turning-point environment — which this one arguably is — trend imputation is the worst possible tool. It smooths the turn instead of showing it.

The sequence from today's print to your mortgage rate runs like this, and it runs fast:

The reverse chain also runs. If April prints below 55,000 — which is the consensus floor — the Fed cut probability surges. The 10-year yields pull back. The 30-year mortgage rate softens toward 6.5%. That's real relief in the housing market, immediately, for anyone with a rate-lock in process or a pending sale. The directional signal matters enormously. The specific number doesn't — because in eleven months, the specific number will be different.

The place to watch isn't the headline number. It's average hourly earnings — up 3.5% year-on-year through March, against an inflation rate running at 3.3%. Real wage growth is barely positive. If April's wage print accelerates even 20 basis points above the March rate, the Fed's inflation calculus changes faster than any payrolls beat. The wage number is harder to overstate through birth-death adjustment, because wages are surveyed directly. It's the one part of the release that's actually close to a measurement.

The pattern is consistent enough that it should change how you read every jobs Friday for the rest of the year. The headline is a first draft. The revision is the final copy. The market sets rates based on the draft and rarely gets to correct them when the final copy arrives. That's not a flaw in the system. It's how the machine was designed — and like most machines, it works exactly as designed right up until it doesn't.