The Two-Economy Problem in Credit Data

Take every delinquency series American Default tracks and rank each one against its own history. Not against a threshold. Not against last year. Against every quarter that series has ever recorded.

Bank-booked single-family mortgages sit at the 21st percentile of their own record. Credit cards at all commercial banks, the Late Fee, at the 32nd. Consumer loans excluding cards, the 34th.

Auto loans, the Repo Line, at the 99th.

Same country. Same quarter. Four kinds of consumer credit, and one of them is having a year the other three have never heard of.

The first three are Board of Governors of the Federal Reserve System bank-condition series retrieved via FRED. The auto series is the Federal Reserve Bank of New York’s Household Debt and Credit Report.

What the average is for

Those four series are the members of the American Distress Index delinquency domain. The domain score that summarizes them reads 46.7.

That number is a mean of the four percentiles above. It is not itself a rank of anything, and the ADI methodology says so explicitly. Read it as a rank and you have invented a statistic.

Read it correctly and it is still doing something strange. It is a summary of four series that describes none of them.

The whole composite behaves the same way. It reads 43.8, Typical (on average, its inputs sit higher than in 44% of their own quarterly histories). Its own gloss states the argument of this piece: the score is an average of where its inputs sit, and averages of dispersed things are not descriptions of any of them.

The personal saving rate makes the point from the other side. On the distress-oriented scale the ADI uses, which inverts a series where lower is worse, it sits at the 92nd percentile of its own history. An index that averaged that against the mortgage book would report calm.

This is not a complaint about the ADI. It is what an index is: a claim about the level of a population, bought by giving up the claim about where inside it.

Same product, two lenders

The cleanest version of the problem needs only one product.

The Board of Governors publishes credit card delinquency for all commercial banks and, separately, for banks outside the hundred largest by assets. Same quarter, same regulator, same collection, same loan type. The only thing that differs is who the lender is, and through that, who the borrower is.

All commercial banks: 2.92% in Q1 2026, the 32nd percentile of its record since 1991.

Banks outside the largest hundred, the Other Banks: 6.43% in Q1 2026, the 87th percentile.

A gap of 3.51 points. Neither series is wrong. Neither is more national than the other. Blend them and you get a number that has never described a cardholder.

The gap has pointed the other way

Here is what keeps this from being a story about small banks.

In Q3 2009, the gap was -2.38 points. Negative. The largest banks were the ones carrying the delinquency, and a reader who had built a thesis on community-lender fragility would have had it inverted underneath them at the worst possible moment.

The separation we see now opened later, and quietly. In Q1 2013 the two rates were 0.63 points apart. By Q4 2019 they were 4.07 points apart.

The aggregate barely moved across those seven years: 2.64% to 2.61%. The small-bank rate went 3.27% to 6.68%.

An analyst watching only the headline series through that window would have recorded nothing happening. Something was happening. It was happening entirely inside the part of the distribution the headline series averages away.

So the split is not a fact about small banks. It is a fact about a period, and the period turned over once already.

It is not an early warning

The tempting next move is to promote the disaggregated series to a leading indicator. The data does not support it.

Take the quarter-over-quarter change in each rate and correlate them at every offset out to 4 quarters in either direction. The strongest relationship is the one with no offset at all: r = 0.41 at zero lag, and no lead or lag tested beats it.

The formal test says the same thing with more teeth. American Default’s leading-indicator scan covers 73 series and tested 87,474 raw pairs. Of those, 30 reached a Granger stage, 15 passed it, and 12 survived out-of-sample validation.

None of the survivors is a small-bank-leads-large-bank pair. The small-bank card series appears in that file only as a follower.

Grade the claim honestly. There is an association between the two channels and a large, persistent divergence in their levels. There is no evidence of a temporal lead. Disaggregation is not a clock. It is a different measurement.

And it is easing

The finding that matters most for anyone reading this as a warning is that the concentration has been unwinding.

The gap peaked at 4.82 points in Q3 2022. It stands at 3.51, after 4 consecutive quarters of narrowing. The small-bank rate itself peaked at 7.86% in Q4 2023 and has been coming down since.

Which is the second half of the same argument, and the half that gets skipped. The aggregate did not show the split opening. It is not showing it closing either. A series that cannot see a divergence appear cannot see it resolve, and a reader who only checks in when the story is bad will keep getting the same non-answer.

What to do with this

Nothing here says the headline numbers are fake. They measure what they claim to measure.

They just answer a question about level when the interesting question is about location. The total delinquency rate can sit near the middle of its own history while one of its components sits near the top of its. Both readings are correct. Only one of them tells you where to look.

An average is an answer to “how much.” Ask it “where,” and it will still answer “how much,” in the same confident voice.