The Unemployment Rate Fell. That's the Worst Part.

The American economy lost jobs last month and the unemployment rate went down. If you read only one of those numbers, you read the wrong one.

On Friday the Bureau of Labor Statistics reported that US nonfarm payrolls fell by 23,000 in July — the first outright decline in months, against a consensus of roughly 83,000 jobs added. The unemployment rate ticked down, to 4.1% from 4.2%. Underneath it, labour force participation slid to 61.4%, its lowest in more than five years. The rate improved because people stopped looking, not because they found work.

Then came the revisions. May's gain was cut from 129,000 to 63,000 and June's from 57,000 to 20,000 — 103,000 jobs erased from months that had already been reported, analysed and traded on. Across three months the US economy added about 60,000 jobs in total. Wage growth cooled to 3.2% year on year, the slowest since May 2021. Futures markets promptly cut the odds of a September rate hike from around 55% to roughly 40%.

The composition tells you more than the count. Local government education shed about 50,000 positions, leisure and hospitality 40,000, retail trade 19,000, and financial activities 14,000 — a sector now down more than 120,000 from its May 2025 peak. Health care kept hiring. This isn't a broad cooling. It is a narrow one, concentrated precisely in the sectors that employ hourly workers and in the small firms that sell to them. Those firms are the ones I have spent my career trying to price.

But the revision is what I would actually flag to a risk committee. A 103,000-job correction is not a rounding error; it means decisions taken in June and July were taken on a picture of the economy that no longer exists. Every credit model I have built or bought in two decades — at eCredit, at Cortera, and now scoring thin-file MSMEs in India — treats macro inputs as fact. Employment, payroll growth, sector health: they go into the model as settled numbers. They are not settled. They are estimates that get quietly rewritten after the loan is booked, the limit is set, and the provision is made.

There is a serious argument on the other side, and it deserves saying. One month is noise. Seasonal adjustment around summer education payrolls is genuinely messy. An unemployment rate of 4.1% is historically strong, and the Fed's inflation problem — driven by an energy shock no rate can un-shock — has not gone away. The three hawks who dissented on 29 July have not been proven wrong by a single soft print.

My read is that they have been proven early, which in policy is a different kind of wrong. An economy averaging twenty thousand jobs a month with falling participation is not one you tighten into. And a labour market that softens while inflation stays sticky is the scenario every 2010s-calibrated risk framework handles worst, because it never had to.

The practical move is unglamorous: stop treating macro inputs as constants. Version them. Re-run the portfolio when the data changes, not only when the borrower does. In India, where MSME lending already runs on thin files and informal cash flows, the discipline of asking “how confident am I in the inputs?” is worth more than another decimal place of model precision.

If the numbers your decisions rest on are revised three months later — how would you even know which of those decisions were wrong?