Writing a Credit Policy for a New Loan Product With No Historical Data

You're launching a new loan product and you have no performance history to model on. No defaults, no roll rates, no vintage curves. The honest answer to "what's the right credit policy for a new loan product with no historical data" is this: you don't build a statistical model on data you don't have. You write a judgmental (expert) credit policy grounded in document-verified affordability and conservative cutoffs, you instrument every decision from day one, and you tighten the rules as real performance accrues. The model comes later. The policy ships now.

This is the cold-start problem, and credit and risk teams hit it constantly: a new segment, a new geography, a new product line, a first lending license. The mistake is waiting for data that only exists once you've lent. The discipline is lending carefully on purpose, capturing rich decision data, and converting that exhaust into a tightening loop. Below is the operator playbook.

A credit policy for a new loan product with no historical data starts judgmental, not statistical

With zero performance history, a scorecard has nothing to learn from. Logistic regression, gradient boosting, any supervised model needs labeled outcomes (who paid, who defaulted) and you have none. So you do what credit teams did before scorecards existed: you encode expert judgment as explicit rules.

A judgmental policy is a set of hard cutoffs and weighted criteria that a senior credit person would apply by hand, written down and applied consistently. The point is not that judgment is better than a model. It's that judgment is the only thing you have at launch, and an explicit, consistent judgmental policy beats inconsistent case-by-case discretion every time. Consistency is what later lets you measure which rules worked.

What goes into the first policy

• Eligibility gates (hard knockouts): minimum age of business or borrower, geography, legal status, sector exclusions, sanctions and KYC pass. Binary, non-negotiable.
• Affordability cutoffs: the core of a cold-start policy. Debt service coverage, disposable income after obligations, exposure caps as a multiple of verified income or turnover.
• Conservative loan caps: cap first-cycle ticket size and tenor hard. You can always raise limits later for borrowers who perform. You cannot un-lend.
• External signals you do trust: bureau scores, existing-relationship behavior, registry checks. Score-agnostic: bring whatever score you have, treat it as one input, not the whole decision.
• Fraud and consistency checks: document tampering, identity mismatch, claims that don't reconcile across documents.

Notice what's missing: no predictive risk grade, no probability of default. That's correct. At cold start you're not predicting default rates, you're enforcing affordability and screening out the obviously bad. For the distinction between a policy that decides and a model that scores, see credit decisioning vs credit scoring.

Why document-verified affordability is the backbone of a cold-start credit policy

When you have no performance history, you compensate with hard affordability evidence. A bureau score tells you how someone behaved on other people's credit. It does not tell you whether this borrower can service this loan. Affordability does, and affordability comes from documents: bank statements, payslips, financial statements, utility and tax records, invoices.

This is where most cold-start policies quietly fail. Teams write strong affordability rules ("require DSCR above 1.25", "average daily balance must cover three installments") and then have no reliable way to compute those numbers, because the inputs are PDFs, photographs of passbooks, scanned statements, and handwritten records. The rule is only as good as the data feeding it. Garbage affordability inputs make a conservative-looking policy effectively blind.

The fix is to treat affordability extraction as a first-class part of the policy, not an afterthought. You want hard, normalized signals computed directly from the source documents:

• Income normalization: recurring vs one-off credits, salary identification, seasonality smoothing.
• Cash-flow analysis: average daily balance, balance volatility, NSF and bounced-payment counts, end-of-month patterns.
• Debt service coverage (DSCR): verified inflows against existing obligations plus the proposed installment.
• Cross-document validation: does declared income match the bank statements, do the financials reconcile with the invoices.
• Tampering and fraud signals: edited PDFs, font inconsistencies, recycled documents.

This is exactly the gap Floowed's Document Intelligence closes. It reads and analyses any loan document at any quality, handwritten passbooks, photographed or skewed statements, scanned financials, and turns them into decision-ready affordability data: normalized income, ADB, DSCR, cross-document checks, fraud signals. This is not OCR or plain extraction. It's the analysis layer that other IDPs (Ocrolus, Rossum, Hyperscience, built for pristine US documents) struggle with on real-world, messy loan paperwork. For the affordability mechanics specifically, see cash-flow underwriting, and for statement-level extraction, bank statement scanning and extraction software.

Set conservative cutoffs on purpose, then earn the right to loosen

At cold start, asymmetry is everything. The cost of approving a bad loan is a full loss of principal. The cost of declining a marginal-but-good loan is a missed margin. Early on, before you can tell those two borrowers apart, you should bias hard toward declining the marginal case. You tighten first, loosen later, because loosening on evidence is safe and tightening after losses is expensive.

Concrete ways to build conservatism into the launch policy:

• Lower first-cycle limits. Lend small, observe, then graduate good borrowers up.
• Higher affordability margins. Require DSCR comfortably above 1.0 (say 1.25 to 1.4), not at the line.
• Manual review band. Auto-approve the clearly strong, auto-decline the clearly weak, route the middle to a human. The review band is your learning lab.
• Cohort caps. Limit total exposure to the new product until early vintages season.

Conservative does not mean slow. A judgmental policy can still auto-decide most applications in minutes. Conservative means the thresholds are tight and the limits are small, not that a human touches everything.

Instrument everything from the first application

This is the step that separates teams who escape the cold start from teams who stay stuck in it. The reason you have no data is that you never lent. The reason you'll still have no usable data in twelve months is that you lent without capturing why each decision was made. Capture is the whole game.

For every application, store: every input value (including the raw extracted affordability figures), which rule fired, the threshold it was tested against, the margin by which it passed or failed, the final decision, and the reason. Then tie that to outcomes as they arrive (paid, late, defaulted). Now you have a labeled dataset being built in real time, structured exactly around the rules you can act on.

This is where the tooling matters. If your policy lives in spreadsheets and loan-officer judgment, the decision data is unrecoverable: you'll never reconstruct why a 2026 application was approved. If your policy runs in a decisioning engine, every decision is logged with its full rule trace by default. Floowed's Decisioning Engine runs your credit policy on the document-derived data, every application, every time, and records the rules behind each call, audit-grade. That audit trail is also your training data.

Tighten as data accrues: the evolution loop

Once early repayment behavior arrives, the policy stops being static. The loop is simple and you run it on a cadence (monthly at first, then quarterly):

1. Pull decisions by rule. For each rule and threshold, look at the loans that passed near the cutoff and how they performed.
2. Find the rules that discriminate. If a threshold cleanly separates good from bad, trust it more. If it doesn't, it's dead weight, drop or replace it.
3. Recalibrate thresholds. A DSCR floor of 1.4 that shows zero losses at 1.25 is leaving good business on the table. Loosen it, on evidence.
4. Graduate limits. Borrowers who performed earn higher limits and longer tenors. This is your cheapest growth.
5. Introduce a score when you can. Once you have enough seasoned defaults, build a scorecard and run it inside the policy as one more input, with the judgmental guardrails still in place.

The faster you can change a rule, the faster this loop runs. A policy that takes a six-week IT change request per threshold tweak will lose to one a credit and risk team can edit themselves. No-code matters here precisely because the cold-start period demands frequent, small policy changes. See the no-code credit policy builder guide for how that authoring works in practice.

How Floowed runs the cold-start playbook

Two products, one platform, mapped directly onto the four moves above:

• Launch a judgmental policy fast. Credit and risk teams build the rules, cutoffs, and review bands in the Decisioning Engine without engineering, so you ship the policy in days, not a quarter.
• Get hard affordability signals despite no history. Document Intelligence reads and analyses real-world loan documents (handwritten, photographed, scanned) into normalized income, cash-flow metrics, DSCR, and cross-document validation, so your affordability cutoffs run on real numbers, not declared ones.
• Capture rich decision data from day one. Every decision is logged with its full rule trace, audit-grade, building the labeled dataset you'll model on later.
• Evolve rules as performance comes in. Edit thresholds, graduate limits, and later drop in a bureau score or your own model, score-agnostic, absorbed unchanged, without rebuilding the policy.

In production at Alon Capital, founder Rene de Jesus put it plainly: "Floowed reads the documents, runs our credit policy, and surfaces a decision in minutes." Pricing is consumption-based credits, sized on one short call, well under enterprise platforms.

Frequently asked questions

Can you launch a new loan product without any credit model?

Yes, and you usually must. With no historical defaults there's nothing to train a model on. You launch with a judgmental (expert) policy: explicit eligibility gates, document-verified affordability cutoffs, and conservative limits. The statistical model comes once you have seasoned outcomes to learn from.

How much historical data do you need before building a scorecard?

It depends on default rates and volume, but you need enough actual defaults (the rare event) to be statistically meaningful, not just enough applications. Many teams need several thousand seasoned loans and a few hundred defaults before a scorecard is trustworthy. Until then, the judgmental policy plus instrumentation is what generates that dataset.

How do you assess affordability with no credit history?

Through documents. Bank statements, payslips, financial statements, and tax records give you verified income, cash-flow stability, and debt service coverage independent of any prior credit behavior. The constraint is reading messy real-world documents reliably, which is what document intelligence solves. See cash-flow underwriting.

What's the difference between a judgmental policy and a rules engine?

A judgmental policy is the set of credit rules a senior expert would apply; a decisioning or rules engine is the software that runs those rules consistently on every application and logs the result. You need both: the policy is the thinking, the engine is the enforcement and the audit trail. See what is a decisioning engine.

How fast should you change the policy at cold start?

Often. The cold-start period is exactly when you learn the most per loan, so you want to recalibrate thresholds monthly at first. That's only feasible if credit and risk teams can edit rules directly, without an engineering change cycle for every tweak.

Ship the policy, then let the data tighten it

You don't need history to launch a new loan product. You need a disciplined judgmental policy, document-verified affordability doing the heavy lifting, conservative cutoffs, and complete instrumentation so the next twelve months produce the data your future model will run on. The teams that win the cold start are the ones who lend carefully on purpose and capture everything.

Floowed lets a credit and risk team do exactly that: launch a judgmental policy fast, run it on real document-derived affordability data, log every decision audit-grade, and tighten as performance comes in. Start free or book a demo.