The lot behind the branch.

Bottom line

Take every federally insured credit union with at least $10 million of vehicle loans in both March 2025 and March 2026. That is 2,284 institutions. Rank the ones reporting repossessed vehicle inventory by that inventory as a share of their vehicle book, cut them into quartiles, and put the non-reporters in their own group. Then wait a year and measure vehicle net charge-offs.

Twice the loss rate, top of the scale against the bottom, from one field most institutions do not fill in. The same construction run on every available annual pair back to September 2022 produces the same ordering all eleven times.

I · What the field measures, and what it does not

Account AS0024 on schedule FS220P: consumer vehicle foreclosed and repossessed assets. When a credit union takes a car back, the loan comes off the books and the car goes on, held at its estimated value until sold. The difference between loan balance and sale proceeds becomes the charge-off. So the field is a photograph of losses in transit: committed, sized approximately, not yet recognized.

The obvious objection: if the car is already on the lot, the loss is already made, so predicting next year's charge-offs with it sounds like predicting rain with a wet umbrella. The answer is that the objection concedes the point. A measure that mechanically precedes recognition is exactly what lead time is. Delinquency has the same relationship to charge-offs and nobody calls it a tautology; it just sits earlier in the same sequence. Payment stops, car comes back, car sells, loss books. The field reads the third position in that sequence, the last one before the number everyone waits for.

What the field does not do is beat the standard measures head to head, and this brief will not pretend otherwise. In the March 2025 cohort, ranked by Spearman correlation with next-year vehicle charge-offs: prior charge-offs +0.691, prior vehicle delinquency +0.480, repo inventory +0.387, indirect-lending concentration +0.372. Repo inventory is the weakest of the four as a standalone signal. Controlling for prior delinquency it retains +0.337; controlling for delinquency, prior charge-offs and concentration together it retains +0.156. It carries some information of its own. Not a lot.

II · The quadrant that matters

Split the cohort twice. Once at the median of prior vehicle delinquency, 0.576%, the number a board sees. Once at the median repo ratio among reporters, 0.122%, the number most boards do not. Four cells:

The top-right cell is the finding. One hundred ninety-six institutions whose delinquency sits below the cohort median, the profile a board reads as healthy, and whose lots are full. They charged off 1.11% over the following year against 0.67% for their clean-lot peers. Two thirds more loss, from a group the standard measure cannot separate. That gap is positive in all eleven cohorts.

III · Eleven cohorts, one ordering

One honest wrinkle, stated rather than smoothed: the bottom quartile of reporters is indistinguishable from the non-reporters. Q1 lands below the non-reporter group in seven of the eleven pairs, median difference minus 0.03 points. A little repo inventory means nothing. The gradient begins at quartile 2, which in the current cohort means a repo ratio around 0.09% of the vehicle book and up.

The ordering also survives the obvious confound. Big institutions report more and charge off more, so the gradient could have been a size effect wearing a costume. Within each of four asset bands separately, under $200 million, $200 to $500 million, $500 million to $1 billion, and over $1 billion, the high-repo half of reporters exceeds the non-reporters by 0.5 to 0.7 points. The spread lives inside every size class.

IV · A correction, and what concentration turned out to do

An earlier version of this analysis tested indirect-lending concentration against contemporaneous delinquency, found almost nothing, and said so. That test was wrong, and the error was mine: it compared the two measures on different clocks and different outcomes. On identical terms, concentration measured in March 2025 against vehicle charge-offs in the year to March 2026, concentration is monotonic in all eleven cohorts, with a median spread of 0.45 points from the lowest band to the highest and a standalone correlation of +0.372 against repo inventory's +0.387. Essentially tied. Under full controls, concentration retains slightly more independent signal, +0.182 against +0.156.

V · Seven in ten lots are dark

Now the part that turns a statistical note into a supervisory question. Among credit unions with at least $5 million of vehicle loans in March 2026:

The institutions least likely to fill the field in are the small ones, and the small band carries the highest vehicle delinquency of the four. Whatever sits on those lots is invisible to anyone reading the filings, including, at one remove, the examiners who set the exam calendar from them. A zero in this field means either an empty lot or an unanswered question, and the filing does not say which.

VI · What this does not show

Five things, so nobody has to find them for me.

One outside data point, for corroboration rather than support: Depository360, which runs charter-departure work on the same filings, found that delinquency separates ceased from continuing credit unions in all thirteen of its cohorts, but by margins around 0.12 points. Our own March cohorts reproduce that scale: +0.12, +0.09, +0.09. The headline credit measure separates reliably and thinly. That is the gap this field, and concentration beside it, exist to fill.

Proof: every figure, traced

All figures were recomputed from the raw NCUA call report archives for this publication. Recomputed means the number was independently rebuilt from the raw quarterly zip files and matched. Derived means calculated from recomputed figures, arithmetic shown.