Your peer group is running out of peers

There were 18,443 FDIC-insured institutions at the end of 1984. On 30 June 2026 there were 4,260.

That figure is reconstructed from the FDIC's own directory rather than taken from a headline, by counting every institution whose establishment date precedes the measurement date and whose closing date does not. The FDIC's Quarterly Banking Profile reports 4,238 for the same day. The 0.5% gap comes from institutions with no recorded closing date, and it is worth naming rather than smoothing over.

Consolidation is slowing, not accelerating

The first thing the series says contradicts the way this subject is usually written.

Net attrition ran 214 institutions a year between 2016 and 2021. Between 2021 and 2026 it ran 118. The rate has roughly halved.

That is not a recovery. Nothing about it suggests new charters are replacing what is leaving. It means the pool has shrunk to the point where the same forces remove fewer institutions in absolute terms, because there are fewer left to remove. At 118 a year the population passes 4,000 in early 2028.

The consequence nobody measures

Scale arguments are well covered. The measurement problem is not, and it lands on the exact analysis every small bank is expected to produce for its board and its examiner.

Peer comparison assumes a peer group exists. Take the population at 31 March 2026 and sort it the way a bank actually thinks about itself, by state and by asset band:

Five asset bands across fifty-odd states produce 238 cohorts. Seventy-nine of them contain four institutions or fewer.

Add the first two rows. 426 banks, 9.9% of the industry, sit in a state-and-size cohort with fewer than ten members. For those institutions a median is being computed from single digits, a quartile boundary moves when one bank has an unusual quarter, and a peer average is an average of a handful of specific competitors who can be named.

Why this is a risk problem rather than a statistics problem

A thin cohort does not merely produce a noisy benchmark. It produces a benchmark that moves for reasons unrelated to the bank being measured.

Consider a nine-member cohort. One member sells a branch and books a gain. The cohort's return on assets jumps, and every other bank in it appears to have underperformed in a quarter where nothing about its own operations changed. Consider the same cohort after an acquisition removes a member. The comparison now runs against eight banks, the history runs against nine, and nobody restates the earlier periods.

This is the same problem in a different costume as the one that arrives when a credit union buys a bank. The institution leaves the Call Report population and its history stops. Anyone carrying a multi-year peer series through that event is comparing two different populations and calling it a trend.

What to do about it

Three things, none of them exotic.

Name the cohort before reading the comparison. If the peer group has eight members, the analysis should say eight, and a reader should be told. A percentile against a cohort of eight is a rank in disguise, and it should be presented as a rank.

Widen deliberately rather than by accident. Dropping the state constraint and comparing nationally within an asset band gives cohorts of 554 to 1,446 members at every size below $10 billion, which is statistically comfortable and geographically meaningless. Both framings are legitimate. Using one and describing it as the other is not.

Build the group by hand where the automatic one is thin. Ten institutions chosen for genuine similarity, in business model and funding mix rather than in postcode, will out-explain a mechanical cohort of four every time. That is more work, and it is the work.

The limit of this reading

The cohort arithmetic here uses state and asset size because that is how supervisory comparison is usually framed and how most peer tables are built. It is not the only reasonable definition, and a bank whose business is unlike its neighbours' will find any geographic cohort unsatisfying regardless of how many members it has.

The point is narrower than it may appear. It is not that peer benchmarking has stopped working. It is that the population has changed underneath a method that assumes plenty, and for one bank in ten the assumption no longer holds. That is a fact about your comparison rather than about your performance, and it should be stated on the page where the comparison appears.

---

Sources

• Institution directory, establishment and closing dates · Federal Deposit Insurance Corporation
• Quarterly Banking Profile, second quarter 2026 · Federal Deposit Insurance Corporation