top of page

When Operational Issues Cause Credit Risk

Writer: Brandon Homuth
Brandon Homuth
1 day ago
5 min read

During a credit risk training early in my career, I remember learning from an experienced credit manager: credit risk is neither created nor destroyed, it's simply modeled better. The idea being that risk is a property of the borrower and the loan — your job is just to measure it accurately. Miss it, and you've got a modeling problem, not a risk problem.


It's a tidy axiom. It's also wrong. Servicing and collections don't just measure credit risk — they change it, in both directions. Two experiences, a decade apart, taught me that the hard way.


The product code


Early in my career, I was the credit officer for a point-of-sale lending business that was replacing manual, judgmental underwriting with an automated model. I'd set expectations for what delinquency and loss should look like on the new book, and for a while the data matched the story: first-payment failures were in line, and roll rates from current into early-stage delinquency (0 to 1) tracked my assumptions closely.


Then the roll rates from DQ1 through DQ4 came in meaningfully worse than expected. That's the part that should have worried me most — and it did, for a few days. But the next number was stranger: the roll from DQ4 to charge-off was better than expected, by enough to mostly offset the earlier deterioration. On paper, the loss curve looked almost fine. Underneath, two different stories were fighting each other, and neither matched what a model change should produce.


I dug in. The culprit wasn't the model at all. When the new underwriting engine launched, an internal product code had changed along with it — a tagging field with no credit meaning whatsoever. But the collections team's automated dialer campaigns were built to source accounts by product code, and nobody had updated the mapping. Loans originated under the new code simply weren't being worked in early-stage collections. They rolled through DQ1 to DQ4 on their own, with no calls, no reminders, no intervention — which is exactly what you'd expect to produce worse early-stage roll rates with no change in underlying credit quality.


But by the time those accounts hit the final bucket before charge-off, they were on a manual review list that collectors worked by hand, outside the broken automated campaigns — and that team picked them up and started calling. The credit quality underneath had never moved. What moved was who was calling whom, and when. The model got blamed for a routing bug.


A newer version of the same problem


I saw a variant of this recently with a lending client whose first-payment failure rate on one loan cohort spiked to nearly 40%, well above a normal range in the mid-20s. The obvious suspects got tested first, in order: had underwriting policy just changed? Had credit quality on this vintage quietly slipped? Had the mix of borrowers on autopay shifted?


All three came back negative. The new underwriting policy was, if anything, producing better credit quality on paper — stronger scores, lower expected loss — and it wasn't associated with either better or worse early failure rates once you controlled for it. Enrollment in autopay was right in its normal historical band. None of the usual levers explained a spike that size.


The team then found a real, confirmed operational cause: a batch of loans where a payment mandate had simply never been established with the servicer due to a file-transition error, so the draft was never attempted at all. That was a genuine, one-time servicing failure, and it explained a meaningful slice of the spike.


But backing those loans out of both the numerator and denominator still left a failure rate roughly six times the historical baseline for autopay-enrolled loans. The one clean explanation they had wasn't the whole explanation. That's the moment this stops being a data-quality exercise and starts being a real investigation — you have a confirmed operational issue, and it's still not big enough to close the gap.


The digging that followed looked a lot like my product-code story, just dressed differently: not "is the model wrong," but "is something upstream of the model quietly changing outcomes." The team sliced the failures by servicing channel, by which sub-populations were affected, and by how failed payments behaved after the initial failure — not just whether they failed, but whether they cured. What they found was a slower-moving version of the same mechanism: the rate at which failed automated payments got cured before they compounded into real delinquency had been declining for the better part of a year, not spiking overnight. A one-time servicing failure was layered on top of a second, quieter erosion in how returned payments were being worked. Both were operational. Neither was credit.


What this means practically


A few things I'd take from both stories, if you're building or advising a lending business:


Treat a sudden delinquency move as an investigation, not a verdict. The instinct when FPD or early-stage delinquency jumps is to start second-guessing the model or the credit box. Sometimes that's right. But before you touch underwriting, rule out the boring explanations: did anything change in how loans are enrolled in payment methods, routed to collections, or worked by servicing, around the same time the numbers moved?


Look at cure and recovery behavior, not just failure rates. A pure credit-quality problem tends to show up as loans failing and staying failed. An operational problem often shows up as loans failing and not being worked — the recovery or cure rate on failures is the tell, and it's a number most teams don't monitor as closely as the failure rate itself.


Segment before you conclude. In both cases, the aggregate number hid a concentrated story — a product code affecting a specific book, a servicing gap affecting a specific vintage. Whole-portfolio numbers can look like a credit signal even when the actual driver sits entirely inside one narrow, operational slice.


Build servicing and ops monitoring as seriously as you build credit monitoring. Most lenders have real infrastructure for tracking model performance and portfolio quality. Far fewer have equivalent rigor around monitoring the operational plumbing — payment processing, collections routing, servicer handoffs — even though a break there can produce a delinquency signal indistinguishable from a real credit problem until someone goes looking.


Credit risk isn't just measured by your model and your servicing stack — it's produced by them, together. The axiom should probably read: credit risk is neither created nor destroyed, it's created and destroyed constantly, by everything downstream of underwriting. Modeling it well is necessary. It was never sufficient.

bottom of page