Three signals, read together
A dealer channel is scored on three numbers, each doing a different job. Net charge-off rate (NCO) is the realised loss number, settled fact, entirely backward-looking. 90+ days past due is the severe-lates share, the leading indicator: most accounts that reach ninety days roll forward into charge-off, so this is next quarter's loss number arriving early. Early payment default (EPD) is the share of loans that went 60+ days delinquent within their first six months on book. It is the cleanest dealer-level underwriting signal in the set, because a loan that fails almost immediately was usually never good to begin with, and origination is where the dealer sits.
The same arithmetic runs on any grouping key. An indirect lender cuts the book by dealer channel. A buy-here-pay-here operator can run the identical score by sales rep or by vehicle class. A multi-lot group can run it by location. The method does not change; only the grouping does.
The score: weights, caps, and the volume adjustment
The composite health score runs 0 to 100, higher meaning worse. Each signal is first normalised against a stress cap, divide the channel's rate by the cap, truncate at 1.0, so no single blown-out metric can dominate beyond its weight. The caps are 15% for NCO, 12% for 90+ DPD, and 10% for EPD, levels set against published subprime auto benchmark families (Fitch, S&P Global, TransUnion). The three normalised signals are then weighted 40 / 30 / 30 and multiplied by a volume weight.
The volume weight uses a log scale on the channel's trailing-90-day loan count: min(1.2, 0.7 + 0.3 · log₁₀(loans ÷ 60)). A channel with 60 loans or fewer is damped to 0.70; roughly 190 loans earns about 0.85; 600 loans reaches 1.00; the weight caps at 1.20 near 2,800 loans. The full specification:
| Component | Definition | Rule |
|---|---|---|
| NCO rate | Realised net charge-offs, trailing | 40% weight · stress cap 15% |
| 90+ DPD | Severe lates; leading loss indicator | 30% weight · stress cap 12% |
| EPD | 60+ DPD within first 6 months on book | 30% weight · stress cap 10% |
| Volume weight | Log scale on 90-day loan count | min(1.2, 0.7 + 0.3·log₁₀(loans÷60)) |
| Composite | 100 × (0.4·nNCO + 0.3·nDPD + 0.3·nEPD) × vol. weight | capped at 100 |
| Bands | High risk / watch / healthy | ≥70 · 50-69 · <50 |
Three bands, three different reads
High risk (score ≥ 70). The channel is harming the book more than it helps. The standard read: cap or restrict new originations from the channel, look at dealer recourse on the existing paper, and pull a set of representative loans for underwriting review.
Watch (50-69). The deterioration is visible but not yet conclusive. The standard read: pull a recent vintage sample and look for underwriting drift, and tighten stip verification or down-payment requirements before the channel migrates into the top band.
Healthy (below 50). A channel to grow into. The useful move is diagnostic in the other direction: study what its stip stack and deal structure look like, and check whether the weaker channels differ in ways that explain the gap.
A worked book: twelve synthetic channels
Here is the method applied to a synthetic twelve-channel portfolio built to mirror a typical mid-sized subprime book: 1,388 loans, $17.9M of unpaid balance. Weighted by balance, the book runs 7.5% NCO, 5.1% 90+ DPD, 4.3% EPD. Every name and number below is synthetic and illustrative.
| Channel | Loans (90d) | UPB | NCO | 90+ DPD | EPD | Score | Band |
|---|---|---|---|---|---|---|---|
| Auto Barn Motors · TX | 142 | $1.85M | 13.2% | 9.4% | 8.5% | 68 | Watch |
| Pilot Point Motors · OK | 96 | $1.24M | 11.7% | 8.1% | 7.2% | 56 | Watch |
| Coastal Drive · FL | 188 | $2.46M | 10.4% | 7.2% | 6.4% | 55 | Watch |
| Lonestar Cars · TX | 74 | $0.92M | 9.9% | 6.8% | 5.9% | 44 | Healthy |
| Sunrise Auto Group · CA | 121 | $1.58M | 8.1% | 5.4% | 4.6% | 39 | Healthy |
| Phoenix Auto Plaza · AZ | 88 | $1.12M | 7.4% | 5.0% | 4.2% | 34 | Healthy |
| Heritage Motors · NC | 156 | $1.98M | 5.8% | 3.9% | 3.1% | 28 | Healthy |
| Northstar Auto · MN | 64 | $0.85M | 5.2% | 3.5% | 2.8% | 22 | Healthy |
| Cascade Auto Sales · OR | 92 | $1.18M | 4.6% | 3.1% | 2.4% | 21 | Healthy |
| Sterling Motorcars · WA | 108 | $1.42M | 4.1% | 2.8% | 2.0% | 19 | Healthy |
| Riverside Auto · CA | 135 | $1.72M | 3.8% | 2.5% | 1.8% | 18 | Healthy |
| Atlantic Auto Group · NJ | 124 | $1.58M | 3.4% | 2.2% | 1.5% | 15 | Healthy |
Why a handful of dealers carry the deterioration
Deterioration in an indirect book rarely arrives evenly. Dealers differ in how disciplined their finance office is, in the customer base their lot attracts, and in how hard marginal deals get pushed through at month-end. Those differences compound: the channel with the loosest structure attracts the applications the tighter channels decline. The honest concentration measure is loss share against volume share. In the worked book, the worst channel produces 18.1% of the losses on 10.2% of the volume, a 1.8x ratio, and runs 1.8x the book's weighted-average NCO. Its six-month trend line is also the steepest in the book, up 3.4 points over the last three months of the series.
Limits
What this method cannot see. The stress caps are fixed choices. 15 / 12 / 10 are defensible against published subprime benchmarks, but a deep-subprime book and a near-prime book should not share caps; a real deployment recalibrates them to the book's own vintage history. EPD depends on honest data. Deferrals, re-ages, and due-date changes silently repair early payment default; a channel whose paper gets serviced generously will score better than its underwriting deserves. Small channels stay jumpy. The log damping softens small-sample noise but cannot fix it, a 20-loan channel's score is an anecdote, not a statistic. The score is descriptive, not causal. It cannot distinguish a dealer problem from a geography or collateral problem that happens to share a lot. And channel labels can hide common ownership, two clean-scoring lots run by the same finance office are one exposure wearing two names. The score ranks where to look. It does not explain what you will find.