How Do We Think About Score Cutoffs — and How Often Should We Revisit Them?

A model that ranks applicants by risk is only half the job. At some point, you have to draw a line — and how you draw it matters as much as the model itself.

Score cutoffs come up constantly in conversations with clients and prospects, and the questions around them are often underappreciated. Where should the line be? What's driving that decision? And once you've set it, when do you look at it again?

This post walks through how we think about cutoffs: what they're actually doing, how calibration makes them meaningful, the frameworks we use to set them, and why they require ongoing attention.

Two Distinct Uses for a Cutoff

Before getting into methodology, it's worth being precise about what we mean.

In an originations context, score cutoffs serve two distinct purposes. The first is the approve/decline decision: a borrower comes in, a score is generated, and you need a threshold that determines whether you even offer them a lending product. The second is product and pricing differentiation within the approved population. Most credit businesses offer a range of products — different loan amounts, terms, rates — and the model score is central to determining which tier a borrower qualifies for. A gold/silver/bronze structure, for instance, requires threshold decisions not just about who gets approved, but about where each approved borrower lands.

Calibration First

Before any cutoff can be set meaningfully, the score needs to be calibrated. We've written about this in more depth elsewhere, but the short version: we think of a raw model score as a rank-ordering tool, not a literal risk prediction. A gradient-boosted model that outputs a 0.12 score for an applicant may or may not indicate that person has a 12% chance of defaulting. But it for sure is saying that person is riskier than someone scoring 0.11 and less risky than someone scoring 0.13. The absolute number, on its own, doesn't mean much.

To turn that ranking into something you can actually build decisions around, you need to do two things.

The first is calibrating the raw score to the model's target variable. In most credit models, the target is something like "ever 90+ days past due at 12 months on book" — a meaningful early proxy for credit risk that doesn't require waiting years for data to mature. Calibration here means establishing, on a representative recent dataset, how a given score range actually translates to that outcome. What fraction of borrowers scoring between, say, 0.10 and 0.15 actually hit 90+ DPD at 12 months?

The second step is translating that model target to a longer-horizon business metric — typically something like a lifetime cumulative loss rate or an annualized loss rate — that feeds into your economic model and pricing assumptions. This isn't a modeling exercise; it's a financial forecasting one. But it's necessary if you want your cutoffs to connect to actual business outcomes rather than just to a development-era outcome definition.

One critical point on calibration: it needs to reflect your current population, not just your development data. If your business has grown into new channels, shifted its marketing, or changed its policy rules, the relationship between score and outcome on your old funded book may not hold for who's walking in the door today. Calibrating on recent vintages that are representative of your current applicant mix is what makes the scores actionable.

Two Basic Frameworks for Setting the Overall Cutoff

With a calibrated score in hand, the most straightforward way to frame cutoff decisions is through the lens of two core portfolio metrics: cumulative loss rate and approval rate.

A more powerful model, by definition, does a better job separating risk. That improvement can be deployed in one of two ways:

• Hold approval rate constant, reduce risk. Approve the same fraction of applicants you're approving today, but let the model's better rank-ordering select a less risky group. Your volume stays the same; your expected losses decline.

• Hold risk constant, increase approvals. Target the same cumulative loss rate you're currently running at, and let the better model expand the population of applicants you're willing to take on.

In practice, most businesses end up somewhere between these two poles, or shift between them depending on economic conditions, funding capacity, or growth targets. But as a framing device — both for thinking about how to deploy a model and for communicating its value to internal stakeholders — this two-axis view is useful.

The Better Approach: NPV-Positive Decisions at the Margin

The cumulative risk/approval rate framework is a useful starting point, but it has a fundamental limitation: it evaluates the portfolio in aggregate. What you really want to know is whether the next borrower you approve — at the margin of your cutoff — is going to be profitable.

That's the logic of setting NPV-positive cutoffs.

Rather than targeting some collective loss rate across all approved borrowers, you break the approved population into score bands — deciles, or finer buckets — and calculate the expected economics of each band independently. For each band, you can estimate expected revenue based on pricing, expected loss based on calibrated risk, and relevant cost assumptions. If a band is expected to generate positive returns, you approve it. You work down the score distribution until you find the band where the math no longer works, and that's where you draw the line.

The same logic applies to product and pricing tiers. Rather than asking "what's our cumulative loss rate across gold borrowers," you ask whether the marginal borrower at the gold/silver boundary generates a positive expected return in the gold product at gold pricing. If they do, they belong there. If not, you either move them to silver or rethink the pricing.

This approach requires more infrastructure — specifically, a robust NPV model or profitability framework that can translate risk estimates into return projections. But it produces fundamentally better decisions. You're not just optimizing for risk metrics; you're optimizing for the actual economics of the business.

One thing worth flagging about what this looks like in practice: the impact on headline metrics isn't always obvious. You may not see a clean increase in overall approval rate, and your cumulative loss rate may or may not improve in a simple comparison. What you tend to see instead is better product alignment — the right customers in the right product at the right terms — which drives better performance within tiers, improved customer selection, and increased conversion. We've seen this play out with auto lending clients in particular, where competition is intense and the distinction between approving someone and actually winning their business depends heavily on offering terms that match their risk profile. Getting more applicants into a lower-rate tier on the strength of a better model didn't change the overall approval rate — but it meaningfully moved conversion.

Cutoffs Aren't Set-and-Forget

However carefully you set a cutoff at model launch, it will eventually need to be revisited. The question is when and why.

The most common driver is economic environment. If macro conditions shift — rates move, unemployment changes, consumer stress increases — the actual performance of a given score band can change even if the model's rank ordering holds up fine. The model is still correctly identifying who is relatively riskier and less risky. But the absolute level of risk across those bands may have moved, which means cutoffs calibrated to a prior environment may no longer be generating the returns you expected.

A second driver is population shift. If your business expands — new geographies, new channels, relaxed policy rules that bring in applicants who weren't previously being scored — the model may encounter borrower profiles it hasn't seen before. Rank ordering may still hold for the core population, but you should have low confidence in how the model scores unfamiliar segments until you have performance data. The right response is closer monitoring and, often, more conservative cutoffs while you build that data.

What does monitoring actually look like? We recommend establishing early-stage risk indicators — short-window delinquency metrics, first payment defaults, early behavioral signals — and tracking them against expectations by score band from the time of launch. If early indicators are coming in hotter than projected, that's a signal to revisit the cutoffs, tighten the calibration, or both. You don't need to wait for 12-month outcomes to know something has shifted.

The key question is whether rank ordering is still holding. If the model is still correctly separating risk, the adjustment is relatively contained — you recalibrate the score-to-outcome mapping and adjust the bands accordingly. If rank ordering itself has degraded, you're in rebuild territory.

The Underlying Requirement

All of this — calibration, NPV-based cutoffs, monitoring frameworks — requires something beyond just a good model. It requires a strategic and analytical framework for the business that can translate model scores into profitability decisions.

A powerful, well-calibrated originations model is a critical input. But the cutoff decisions that sit on top of it are business decisions, not modeling ones. The teams that get the most value from custom models are the ones who've done the work to connect model outputs to the economics of their specific product, customer base, and competitive environment.

If you're thinking through how to structure that framework, or how to evaluate whether your current cutoffs are well-calibrated to your business, we're happy to dig into the specifics.