Why Your Consumer Credit Score Can't See Merchant Risk
In any lending model where a borrower transacts through an intermediary — a school, a medical provider, an auto dealer, a furniture retailer, a merchant on a BNPL rail — the same borrower profile can perform completely differently depending on where they walked in the door. A borrower with an identical credit score behaves differently financing a car at a franchise dealer with decades of service history than at a small independent lot that opened last year. The same is true financing an elective medical procedure at an established practice versus a newly opened clinic, or a furniture purchase at a retailer with a stable delivery and return track record versus one without.
The consumer score doesn't know the difference. It was never built to.
We've been working through this problem with several lending clients across different verticals, and it's the same underlying pattern every time: a meaningful share of loss variance sits at the merchant or provider level, not just the borrower level, and a pure consumer score leaves that risk on the table. Here's the framework we've landed on.
The Problem: Merchant Risk Hiding in Plain Sight
Most consumer credit models are, by design, agnostic to who the borrower is transacting with. That's usually the right call — a model that stays a clean read on the individual is more stable and more defensible. But when a meaningful share of loss variance sits at the merchant or provider level rather than the borrower level, a pure consumer score misses it.
The instinct is to reach for a category or segment cut — elective medical vs. home improvement, franchise dealers vs. independent lots, furniture vs. electronics — as a rough proxy. That instinct breaks down fast for three reasons:
Segment-level instability. A handful of new merchants entering or exiting a category can flip the aggregate signal even though nothing about underlying risk has actually changed.
Within-segment heterogeneity. Two merchants in the same category — two independent auto lots, or two elective medical practices — can have wildly different outcomes depending on their own operating characteristics, not the category they sit in.
Multi-line merchants. The same merchant can offer multiple products or service lines with very different risk profiles — a dealer selling both new inventory and a higher-risk buy-here-pay-here used book — and category-level cuts can't see below the merchant level to catch it.
The result is a risk signal that's directionally right but too noisy and too coarse to price against with any confidence.
The Fix: Decouple Consumer Risk From Merchant Risk
The architecture we landed on treats these as two genuinely separate assessments that combine, rather than one blended score:
The consumer score stays a pure read on the individual — no category, no merchant signal baked in.
Merchant risk becomes its own independent score, built from characteristics like business type, ownership and due-diligence data, industry-specific outcome metrics (repossession rates, completion rates, patient outcomes, return/complaint rates — whatever's relevant to the vertical), and trailing empirical performance.
The two don't get averaged into a single number. Instead, merchant risk sets the merchant discount fee and scorecut structure a given merchant operates under, and the consumer score slopes the borrower's rate and terms within that. A lower-risk merchant gets a lower discount fee and a more generous scorecut structure; a higher-risk merchant pays a higher discount fee and operates under tighter cuts, to compensate for the added risk they bring to the portfolio. Within either merchant's book, the consumer score still does the work of separating a strong borrower from a weak one. Same consumer score, different economics, because the merchant's risk tier sets the fee structure and scorecuts it's priced against.
This solves all three problems above in one move. New merchants entering a category don't distort anything because they're assessed on their own characteristics, not folded into an aggregate. Heterogeneous merchants within a category get separated by their actual features rather than a shared label. And multi-line merchants can carry different risk assessments for different products or service lines, because the assessment lives at the product/program level, not just the merchant level.
Build It Continuous, Ship It as Tiers
Once you've decided to build a separate merchant risk score, the next question is what form it should take: a continuous score, or discrete tiers.
Our recommendation is both, at different layers. Build the underlying model as a continuous, statistically-driven score — every merchant or program gets a real number, not a bucket. Then implement discount fees and credit policy off a small number of discrete tiers (three to four to start) that the continuous score maps into, each with its own fee schedule.
The case for continuous underneath is straightforward: it's more statistically honest, avoids arbitrary line-drawing at bucket boundaries, and gives you a foundation you can refit cleanly as new data comes in rather than re-litigating bucket definitions every quarter. It also lets the model use as many features as it needs — a dozen or more inputs — without that complexity leaking into how the business actually operates.
The case for tiers on top is operational. "This dealer moved from B to C" is something underwriting, sales, and leadership can act on immediately. "The modifier moved from 1.17 to 1.29" is not. Tiers are also a far easier pitch to the merchants themselves — a concrete fee schedule and a clear path from B to A is a real conversation. An ever-shifting continuous term is not.
A pure continuous modifier — applying the score directly as a multiplier on the loss assumption or discount fee, with no tiering layer at all — is worth naming as the more precise alternative. But it breaks down at the tails, since a linear multiplier doesn't behave cleanly near a fee floor or cap, and it doesn't produce anything you can hand to a merchant as an explainable fee schedule. We think of it as a viable end state to revisit later, not a starting point.
Build the Model on Raw Features, Not on Your Own Past Decisions
One design choice worth calling out separately: how you train the underlying model matters as much as what output format it produces.
Historical default and early-delinquency data at the merchant level is often already contaminated by a lender's own past policy reactions — merchants that were already tightened or cut off because performance looked bad. Train a model directly on that history and you risk re-learning your own past decisions rather than discovering true underlying risk.
The more durable approach is to cluster merchants on raw, independent characteristics — due-diligence data, product/service type, industry-specific outcome data — without reference to how the lender has already reacted to them. Then test whether those clusters explain risk better than the category cuts already in use. This is a useful analytical starting point regardless of whether the final output is continuous or tiered.
Category Doesn't Die — Its Job Changes
None of this means category or segment disappears from the business. It just stops trying to do a job it was never precise enough to do. The risk signal it used to carry gets better captured directly by granular merchant-level features. But category likely retains a real role as a commercial and communication construct — what the market will bear on pricing, forecasting, and internal reporting language — separate from the risk model itself.
New Merchants Without a Track Record
A merchant tier framework needs an answer for merchants or programs with no history yet. Rather than a blank assumption, new merchants should start at an informed prior — nearest-neighbor to similar existing merchants, or a category-level average if no close comparison exists — and then get promoted or demoted to a different tier, and a different fee schedule, as real performance data accumulates on a trailing monthly or quarterly basis.
The Takeaway
Merchant and provider risk is real, it's measurable, and it's sitting outside most consumer credit models today — whether you're financing a car, an elective medical procedure, furniture, or point-of-sale purchases generally. The lenders who get the most value from addressing it aren't the ones chasing the most precise possible score — they're the ones who build a statistically sound model underneath and a simple, explainable structure on top of it, so the sophistication does its job without becoming a liability in how the business actually operates.
If you're working through how to separate merchant risk from consumer risk in your own underwriting stack, we're happy to dig into the specifics.