The peer group is the argument

Put a ratio in front of anyone senior and the first question is always the same: compared to what? A 3.1% net interest margin, a 187% CRE-to-capital figure, a 62% efficiency ratio: none of them mean anything standing alone. They acquire meaning only against a comparison set, which is why the quiet, procedural act of choosing that set is the most consequential analytical decision in the whole exercise.

It is also the one most often made by default.

Asset size is a weak peer group

The standard construction is a size band: banks between $500 million and $1 billion in assets, say. Size is the easiest variable to sort on and it is what most published peer data is organised around.

Size correlates with some things worth controlling for: overhead structure, regulatory burden, access to funding markets. It correlates badly with almost everything that drives the ratios people actually argue about. Two banks of identical size, one a commercial real estate lender in a metropolitan market and the other an agricultural lender in a rural one, will differ on margin, delinquency, concentration, funding mix and efficiency for reasons that have nothing to do with how well either is run. Comparing them produces variance that looks like performance and is actually business model.

The more defensible construction picks peers on what generates the ratio:

• Loan mix, as shares of the book. This is the single strongest predictor of margin, loss content and concentration ratios.
• Funding mix: core deposits versus wholesale, and the uninsured share.
• Geography, at the level the local economy actually operates.
• Charter and form, because what an institution reports depends on what it is required to file.

473 banks of the same size, and what they actually lend against

Commercial real estate loans as a share of total risk-based capital

One in seven of those institutions, 15% of them, sit at or above the 300% supervisory screen. The rest do not, and the spread beneath that line is the more interesting half: a tenth of the band carries less than 67% of its capital in commercial real estate, and the most concentrated carries more than eight times that. They are all the same size. Whatever the band is measuring, it is not the risk any of them is carrying, and a percentile drawn from it is a fact about the sorting variable rather than about the bank.

Even that chart required a decision. Two hundred and eighty of the 753 institutions in the band report no risk-based capital at all, having elected the community bank leverage ratio, and a measure defined against risk-based capital cannot include them. Left unstated, a choice like that is exactly how a percentile ends up arriving from nowhere.

A peer group of eleven institutions selected on those grounds is worth considerably more than a band of two hundred selected on size.

The comparison you cannot make yet

There is a timing problem underneath all of this that is easy to miss and produces confidently wrong conclusions.

Individual institutions file first. Aggregated peer data, industry and group averages, is assembled and published afterward, on its own schedule. For the most recent quarter, your own numbers exist and the peer set does not yet.

The failure mode is not that anyone lies about this. It is that a system asked to draw a comparison for the current quarter will silently reach back to the last quarter for which the aggregate exists, and present the result as if both sides were contemporaneous. In a stable period that is a small error. Across a turning point, a rate move or a credit inflection, it is a large one, and it points in whichever direction the intervening quarter moved.

The fix is not clever. It is to say so: this quarter has no peer comparison yet, here is your own trajectory, and here is the comparison as of the last quarter where both sides exist. An analysis that admits the gap is more useful than one that papers over it, and considerably easier to defend when someone eventually notices.

Choosing the peer group after seeing how the comparison turns out is the one move that invalidates the whole exercise. Define it first, in writing.

Level, rank, and direction

Three different readings come out of a peer comparison and they should not be conflated.

The level is your raw number. The rank is where it sits in the distribution. The direction is how that rank has moved across quarters.

Direction is usually the most informative and the least reported. Sitting at the 70th percentile on CRE concentration is a fact about the bank's strategy. Moving from the 40th to the 70th percentile over six quarters is a fact about its trajectory, and it is the one that will come up in the examination. A percentile with no time series behind it is a snapshot presented as a conclusion.

Honest rather than objective

There is no objectively correct peer group, and the search for one is a category error. Every construction encodes a judgement about which similarities matter, and that judgement is the analysis rather than a preliminary to it.

What is achievable is a comparison someone can argue with: a definition written down before the numbers were computed, held stable across quarters, and changed deliberately when it changes. A reader can then disagree with a specific decision, that this institution does not belong or that variable is the wrong sorting key, rather than with an unexplained percentile.

The distinction matters more than it sounds, because a peer comparison is one of the few analytical objects that is almost always presented as a fact and almost never as an argument. The percentile arrives with the authority of arithmetic, and the decision that produced it arrives with no authority at all, because it usually arrives invisibly.