Why the aggregate is the wrong unit

In an indirect book, loss attribution naturally starts with the channel: which sources sent the paper that went bad. That is its own method (Note 03). A direct-to-borrower book has no channel axis at all. The underwriting decision belongs to the lender, so attribution has to come from the credit box itself, which FICO band, vehicle type, vehicle age, mileage bucket, advance rate, term, payment burden, and geography are running hot relative to the rest of the book.

A book-level cumulative net loss of, say, 9% is not actionable on its own. It could be a uniformly mediocre book, or a mostly sound book subsidizing three segments running at twice the average. The credit-box response to those two situations is completely different, and the aggregate cannot tell them apart. Segment attribution exists to tell them apart.

Four numbers per segment

The method computes the same four quantities for every segment, on originated balance rather than loan count wherever dollars matter.

Cumulative net loss rate is gross charge-offs minus recoveries, divided by the segment's originated balance. Share of total loss is the segment's net-loss dollars over the whole book's net-loss dollars, the materiality axis. Default frequency, the PD proxy, is charged-off plus repossessed units over all units in the segment. Loss severity, the LGD proxy, is net loss on the defaulted loans divided by those same loans' originated balance. A fifth number, 60+ days past due among still-active accounts, serves as the forward indicator.

The frequency-severity split is the point of the exercise. A segment can carry outsized loss because too many of its loans default, deep credit tiers, stretched payment burdens, or because each default recovers badly, aged luxury collateral, very high mileage. The first is an approval and structure problem; the second is a collateral and advance-rate problem. Same loss dollars, different fix.

The segment grid

Bins are fixed, not data-driven. A segment definition that drifts with the book cannot be tracked across snapshots, so the cut points below stay constant and the book moves through them. These are the exact edges the method uses.

DimensionBin edges, as used
Credit (FICO)<520 · 520-539 · 540-559 · 560-579 · 580-599 · 600-619 · 620+
Vehicle typeas labeled on the tape (economy sedan … luxury SUV)
Vehicle age0-3 yr · 4-6 yr · 7-9 yr · 10-12 yr · 13 yr+
Mileage<60k · 60-90k · 90-120k · 120-150k · 150k+
Term≤48 mo · 49-60 · 61-66 · 67-72 · 73 mo+
LTV<100% · 100-114% · 115-129% · 130-144% · 145%+
Affordability (PTI)<10% · 10-15% · 15-20% · 20%+
Down payment0% · 0-5% · 5-10% · 10%+
Income typeas labeled (W-2 · 1099 · cash · fixed)
GeographyZIP subprime tier where present, else state
Vehicle age falls back to model year against the as-of date when age is not stated directly. These edges are method choices, tuned to subprime auto; a near-prime book would cut the credit and LTV axes differently.

The flag line: elevated and material

A segment is flagged only when two conditions hold at once: its loss rate runs at 1.2× the book average or worse, and it carries at least 3.5% of the book's total net loss, with a minimum of 20 loans in the bin. Both conditions matter. Without the materiality bar, a tiny 40-loan bucket at 3× book dominates the list while moving nothing; without the multiple, big segments get flagged just for being big.

Flagged segments are then ranked by an impact score: share of loss × (multiple − 1). That is, roughly, the loss the book would not have taken had the segment performed at book average, expressed in share points. A segment at 1.6× book carrying 21% of loss outranks one at 2.0× carrying 12%, which is the correct ordering for anyone deciding what to tighten first. Above 1.4× book, the flag is treated as severe. At book level, the context bands are: cumulative net loss above 8% reads elevated and above 13% severe; severity above 70% severe; 60+ DPD above 7% severe. In the per-dimension view, a bin above 1.35× book reads hot and below 0.7× reads cool.

RuleThreshold, as used
Segment flag, elevatedloss rate ≥ 1.2× book
Segment flag, material≥ 3.5% of total net loss
Minimum sample20 loans (25 for cross-cuts)
Ranking scoreshare of loss × (multiple − 1)
Severe segment≥ 1.4× book
Cross-cut flag≥ 1.5× book and ≥ 3% of loss
Bin heat, dimension viewhot > 1.35× · cool < 0.7×
Book bands: cum net loss / LGD / 60+ DPD8% and 13% / 70% / 7%
The full flag logic. When nothing clears both segment bars, the honest output is that risk is diffuse, the loss is everywhere and nowhere, and the credit box is not the lever.
InferenceA flag is a statement about where the lever is, not a verdict on the loans. A flagged credit band says the approval line is the instrument; a flagged collateral bin says advance rates and valuation are; a flagged geography usually says the box is fine but its application is not uniform. The method locates the decision. It does not make it.
The read The aggregate is calm because the majority is calm. In the synthetic example below, a book at 9.4% cumulative net loss is really a handful of segments running 12-19% stacked on a majority running near 6%. The decision-relevant number is the multiple against book, not the mean, the mean is what the multiple hides in.

The cross-cuts the aggregate hides

Single-axis attribution has a blind spot: risk factors arrive on the same loan. The method therefore crosses six fixed pairs, credit × term, vehicle age × mileage, vehicle type × vehicle age, geography × credit, LTV × term, and affordability × credit, and flags any intersection at 1.5× book or worse that carries at least 3% of total loss on 25+ loans. This is where the classic concentrations surface: deep-subprime credit written at 73 months and beyond, and old luxury collateral, where frequency and severity go bad together.

Here is a worked read on a fully synthetic direct book, roughly 2,400 loans, cumulative net loss 9.4%, default frequency 14%, severity 58%. The top flagged segments, ranked by the impact score above:

#SegmentCum net loss× bookShare of lossDefault freqLGD
1Term · 73 mo+15.0%1.6×21%24%64%
2Credit · <52018.8%2.0×12%31%66%
3LTV · 145%+14.1%1.5×14%22%63%
4Down payment · 0%12.2%1.3×16%19%61%
5Mileage · 150k+13.2%1.4×9%20%67%
Synthetic and illustrative, no actual portfolio. Segments overlap: the same loan sits in several rows, so shares do not sum. Note the ranking: #1 beats #2 on impact (21 × 0.6 > 12 × 1.0) despite the lower multiple.
One bin, two shares, the 73 mo+ term bucket
In the synthetic book, loans written past 72 months are about 13% of originated balance but carry roughly one net-loss dollar in five.
21% 79%, rest of book 73 mo+ share of net loss from 13% of originated balance
Synthetic, illustrative book. Not data from any actual lender or portfolio.

The cross-cuts sharpen the same picture. In the worked book, credit <520 × term 73 mo+ runs at 2.6× book, well past either parent bin alone, and vehicle age 10-12 yr × mileage 150k+ at 2.3×, with LGD in the low 70s. Neither intersection is visible in the single-axis table: the loans are spread across four separate rows there, diluted in each.

InferenceStacked risk factors on one loan behave closer to multiplicatively than additively in loss dollars. A borrower who is deep-subprime and at 84 months and above 145 LTV is not three moderate risks. It is one loan carrying all three, defaulting early with maximum balance at risk and minimum equity. The single-axis view disassembles that loan into bins; the cross-cut is what reassembles it.

Where the bin edges come from

The worked numbers on this page are synthetic, but the structure, where the lines are drawn and which direction the flags point, is not guesswork. It is set to match published, public data, none of it loan-level, borrower-level, or vehicle-history data:

In other words: a reader who wants to know whether the 73-month flag or the sub-620 flag is a real phenomenon can check it against the public record, not take this note's word for it. The full list is under Data & sources.

Limits

What this method cannot see is as important as what it can. It does not separate segment from seasoning. Cumulative loss favors older vintages; a segment concentrated in early originations will read hot partly because it is old. A vintage-controlled cut is the check. It does not establish causation. Segments correlate, deep credit tiers also carry longer terms and higher LTVs, so a flagged bin may be carrying a neighbor's risk. The materiality bars hide small, emerging problems by design; a new 30-loan segment at 3× book will not surface until it grows. Recovery lag flatters recent defaults, understating current-period severity. Input quality binds everything: LTV depends on which book value was used at origination, and PTI on stated income. And the method says nothing about what the segment earns, a 1.5× loss segment written at a sufficiently higher yield may still be sound economics. That is a different analysis.

Sources & notes The attribution method is published in full: every bin edge, formula, and threshold above is stated exactly as used. The cut points are calibrated to public, non-proprietary data, the Federal Reserve Bank of New York Household Debt & Credit report (Consumer Credit Panel/Equifax; 620 subprime line, auto delinquency by tier), Experian State of the Automotive Finance Market (term-length and subprime-delinquency shares), the Federal Reserve FEDS Note on subprime/BHPH lending (May 2026), the Fitch subprime auto ABS net-loss index, and Edmunds negative-equity data. None of these are loan-level or borrower-level. The exact edges remain this publication's own method choices, reasonable practitioners draw them elsewhere, but the direction and materiality of every flag can be checked against those public series. All portfolio figures in the worked table, the cross-cut examples, and the chart are synthetic and illustrative; they describe no actual lender, book, or borrower. Full source list under Data & sources. LendRisk Analytics is an independent research publication with no position in, and no affiliation with, any company mentioned. Not investment, legal, or accounting advice.
Have a question about the market, or a different view? Send it through →.