We're Launching a New Product. Do We Need a New Model?

It's a question that comes up regularly when a lender is expanding — into a new geography, a new credit segment, a new vertical. You have an existing model. You're launching something new. Does the model come with you, or do you need to build something fresh?

The honest answer, as with most modeling questions, is: it depends. But the dependencies are specific enough to be useful, and there's a logical path through them.

What "New Product" Actually Means

Before getting into the modeling question, it's worth being precise about what kind of change we're talking about. In a lending context, a new product can mean several different things:

• A new market or geography — a new country, or a different region within the same one

• A new credit segment — moving upmarket or downmarket from your existing borrower base

• A new vertical or use case — say, adding auto lending to an installment loan business

• A new channel — moving from direct to indirect origination, for instance

• A meaningfully different product structure — different loan amounts, different terms, secured versus unsecured

The reason the definition matters is that not all of these changes are equally disruptive to an existing model. Some mean a genuinely different borrower population with different risk characteristics. Others mean a similar population with a different product wrapper. The modeling implications are very different depending on which situation you're in.

The underlying principle is that the risk of borrowers should be correlated across products. But the details are what matter. You would not expect someone's default behavior on a short-term installment loan to mirror what they'd do on an auto loan or a mortgage. Understanding how closely the new product aligns with what your model was built on is the starting point for the whole analysis.

The Key Dependencies

What model do you have today? A custom model built on your own data from a narrow, specific product may transfer poorly to a new context. A broad bureau-based score — Vantage, FICO — generalizes more widely by design, even if it's less powerful for any given product. If you're currently using a custom model that was optimized for a particular population, its applicability to a new one is genuinely uncertain until tested.

How different is the new population from the old one? Geographic moves are a good example of where this gets complicated. Lenders often assume that cultural similarities between neighboring countries or regions imply data similarities. That's not always true. In some geographies, credit bureaus collect rich positive and negative data, similar to the US. In others, only negative information is captured. The macro profile of the borrower population might look similar; the data infrastructure underpinning a model can look very different. Even within a single country, moving significantly up or down the credit spectrum means a different borrower profile, potentially with different predictive signals.

How much data is available? Is there historical performance data you can actually build on — either your own or accessible through a bureau retrospective study — or are you truly launching into uncharted territory?

What are your practical constraints? This is where theory meets reality. How quickly do you need to be in market? How much capital is available for the launch? What's the risk appetite for learning as you go versus getting things right upfront? A lender with a deliberate launch timeline, sufficient capital, and high expectations for day-one volume has both the ability and the incentive to invest in a proper new model build before launch. A lender moving quickly with limited resources may need to test into the new population with the tools already available. Both are legitimate strategies — they just require different approaches.

A Baseline Is Always Better Than Nothing

Whatever your situation, some baseline prediction is almost always available and almost always useful. That might be your existing model, a broad bureau score, or some combination. The rank ordering may be imperfect in a new context, but imperfect rank ordering still separates risk — and that's better than no separation at all.

What matters is understanding the limitations of whatever baseline you're using, and building a strategy around it that accounts for those limitations. The credit terms, the launch volume, the monitoring cadence, the early intervention thresholds — all of these can be calibrated to reflect the uncertainty in a model that hasn't been validated on the new population. With the right strategy around it, an existing score can be a perfectly reasonable starting point.

One practical option that falls between "use the old model" and "build a new one": commission a lightweight data study. Even without the budget for a full new model build, you can often engage a bureau to pull a targeted retrospective sample on the new population and test your existing model against it. This is a relatively low-cost way to get real signal on how well the model translates before you've committed to a full launch. It won't give you a new model, but it tells you how much confidence to place in the one you have — and that's valuable information.

The Case for a New Model — and How to Transition

If your new product is meaningfully different from what your model was built on, a purpose-built model is the right long-term answer. A model trained on data that actually reflects the new population — its risk characteristics, its repayment behavior, its relevant predictive signals — will outperform a repurposed model from a different context.

The question is often one of timing and sequencing. You don't necessarily have to choose between launching with the old model or waiting for a new one. A combined approach is often the most practical path:

Use the existing model or baseline score to support the initial launch. Build the new model in parallel as performance data accumulates on the new product. Then blend the two — weighting the new model more heavily as confidence in it grows. The transition doesn't have to be a hard cutover. An ensemble approach that gradually shifts weight from the old model to the new one lets you capture the benefits of the new build while maintaining continuity.

We've seen this play out in a couple of different ways. In cases where we've helped lenders build models for genuinely new-to-market products — with no prior performance data of their own — they've used large bureau retrospective studies to build the initial model. That gives you broad coverage and a model that should work across the credit spectrum. The challenge is that you still don't know exactly how it will perform on the specific through-the-door population you'll actually be funding until you're in market. In those situations, close early monitoring is essential, and recalibration based on early results is expected, not exceptional.

In another case, a lender whose existing model was built on a high-risk, downmarket product wanted to use it as a launching pad for moving upmarket. The old model wasn't well-suited to rank ordering lower-risk borrowers — that wasn't what it was built for. But it was quite good at identifying who was very high risk. That's still useful. It provided a guardrail that let the lender cut out a meaningful portion of the incoming population, even if it wasn't doing fine-grained segmentation at the better end of the credit spectrum. Sometimes an imperfect model is still doing real work.

Don't Forget to Recalibrate

One point worth emphasizing regardless of which path you take: even if your existing model rank orders well in the new product context, the calibration almost certainly needs to change.

As we've discussed in prior posts, we think of credit models fundamentally as rank-ordering tools. The score separates higher-risk borrowers from lower-risk ones. But the translation of that rank ordering into cutoffs, pricing tiers, and expected loss rates depends on calibration against the actual performance of the population you're decisioning. A new product — with different loan structures, different borrower behavior, potentially a different default definition — will produce different observed outcomes at any given score band.

If you carry over the old model without revisiting the calibration, you're making implicit assumptions about the new product that the model was never designed to support. Updating the calibration to reflect the new product's actual performance is a necessary step, separate from the question of whether to rebuild the model itself. In some cases, it's the only change you need to make — and that's a meaningful win for getting a new product to market efficiently.

The Practical Takeaway

A new product launch doesn't automatically require a new model. It does require a clear-eyed assessment of how much the new product differs from what your model was built on, what tools are available to fill any gaps, and what your practical constraints are around time, capital, and risk appetite.

In most cases, the right answer is somewhere in the middle: start with the best available baseline, monitor early performance closely, invest in a better model as data accumulates, and recalibrate as you go. The teams that do this well aren't necessarily the ones with the most resources — they're the ones who are honest about what their model does and doesn't tell them, and who build a strategy that accounts for that uncertainty from day one.