There's a specific kind of customer that every neobank has in large numbers and almost none of them are lending to: the deposit-active thin-file account holder.
This person has been banking with you for 14 months. Direct deposit hits every other Friday without fail. Rent goes out on the 1st. Subscriptions auto-renew monthly, and the account has never gone negative for longer than 36 hours. Their FICO score is 582 — or doesn't exist at all — because they've never had a credit card or installment loan.
Your current underwriting stack looks at that 582 and routes them to decline. The deposit data you already hold — the proof of repayment discipline sitting in your own ledger — never enters the decision.
That's the gap this article is about.
Why Deposit-Active Thin-File Customers Are Your Best Credit Risk
The intuition that a customer who banks responsibly will also repay a loan responsibly isn't novel. Community development financial institutions (CDFIs) have operated on this premise for decades. What's changed is the infrastructure to act on it at speed.
When we model the repayment performance of thin-file borrowers who have 12+ months of deposit account history showing regular inflows and consistent recurring payment behavior, the predictive signal is strong enough to separate the population into meaningfully distinct risk tiers. The point isn't that all thin-file depositors are low-risk — it's that the distribution is far more differentiated than FICO 580 implies.
A bureau model sees a single number that was computed from a credit history that largely doesn't exist. A cash-flow model sees 18 months of payroll regularity, expense stability, and net cash position trend. Those are different credit-relevant things, and conflating them by defaulting to the bureau score is what produces the 40% false-decline rate that digital lenders have normalized.
The Three Segments Worth Separating
Not all thin-file depositors present the same opportunity. Within that population, we find three distinct segments that warrant different credit strategies:
1. The Stable Earner
Regular direct deposit (bi-weekly or semi-monthly), payroll amount within ±8% over the lookback period, recurring obligations met consistently, average daily balance trending flat or upward. This segment has the highest approval lift from cash-flow underwriting — roughly 60–70% of this group that bureau models decline would perform within acceptable loss thresholds for a small-dollar personal loan or a modest credit line.
The risk-adjusted math on this segment often surprises product teams. A 6-month $1,500 personal loan at a competitive APR, funded from your own balance sheet via a credit line, has an expected loss rate in the 2–4% range for this population in our validation data — comparable to a bureau prime borrower at FICO 700+.
2. The Gig-Income Borrower
Irregular inflow timing but consistent inflow magnitude. Transaction patterns show multiple income streams rather than a single payroll deposit. Net monthly inflows are positive and relatively stable despite day-to-day variability. This segment is harder to underwrite with simple income verification rules because no single deposit looks like a paycheck — but the aggregate cash-flow picture over 12+ months is legible and risk-stratifiable.
Bureau models perform especially poorly here because gig income doesn't generate installment credit history the way traditional employment does. A driver who earns $3,800/month across platform payouts has zero bureau tradelines but demonstrable income stability. Declining them on bureau grounds alone is a false negative.
3. The Recently Banked
Fewer than 9 months of deposit history at your institution. Insufficient lookback for confident cash-flow modeling. This segment requires either a shorter lookback model with wider confidence bands, or a small starter product (a secured card, a small credit-builder installment) to accumulate the history you need before extending unsecured credit.
We're not saying this group should be treated identically to segment 1 — they shouldn't. The difference is that you have a pathway to eventual credit access, not a permanent wall. The strategy here is building toward the credit relationship, not abandoning the opportunity.
Where Neobanks Are Leaving Money on the Table
The product architecture problem is usually not the credit model itself — it's the integration point between deposit banking and credit underwriting. Most neobanks have these systems in separate stacks with no real-time data bridge. The deposit ledger doesn't feed the underwriting engine; instead, at loan application time, the platform either pulls a bureau report or asks for uploaded bank statements (which is friction that kills conversions).
Fixing this requires two things: a data pipeline that makes deposit transaction history available to an underwriting API at decision time, and a model that knows how to extract signal from that history. The second piece is harder than it sounds — raw transaction data contains a lot of noise, and feature engineering choices (what time windows to use, how to normalize for seasonal income variation, how to handle account-switching patterns) materially affect model performance.
The lenders who build this well get a compounding advantage: every new deposit customer is now a potential credit customer, and every month of account activity is data that improves decisioning precision for that individual. The customers you already have become more valuable over time rather than remaining permanently excluded from your credit product.
A Practical Expansion Sequence
For neobanks evaluating when and how to expand credit into thin-file depositors, we think the right sequence is:
Step 1: Audit your existing decline stack. Pull 6 months of declined credit applications from existing deposit customers. For each decline, check whether the applicant had 12+ months of deposit history with your institution. If 40% or more of your declines are existing deposit customers with substantial account history, you have a data infrastructure problem, not a credit quality problem.
Step 2: Build the data bridge before the model. The limiting factor is almost always pipeline, not ML capability. Getting permissioned transaction data flowing into a decisioning API in real time — without requiring the applicant to re-authenticate or upload statements — is a meaningful engineering lift. It's also where most of the value is created.
Step 3: Start with a narrow product. A $500–$2,000 personal loan or a $1,000–$3,000 credit line is an appropriate first product for this strategy. It limits downside exposure during model calibration while generating enough performance data to expand confidence intervals over 12–18 months. Don't launch with $10,000 unsecured to unproven thin-file borrowers and then conclude alternative data doesn't work when losses come in above expectation.
Step 4: Monitor disparate impact from day one. Cash-flow underwriting doesn't automatically eliminate demographic disparities — it can actually import income-based proxies that correlate with protected class membership if you're not careful about feature selection. Set up demographic analysis of approval and decline rates before you scale, not after regulators ask. ECOA applies regardless of what data you're using.
What This Looks Like in Practice
Consider a neobank with 180,000 active deposit accounts. Roughly 40% — 72,000 accounts — have been open for 12+ months with regular activity. Of those, historical decline data suggests about 28,000 had credit applications denied in the past year primarily on bureau grounds. If a cash-flow model can safely approve 55% of that group (a conservative estimate based on stable-earner and gig-income segments), that's roughly 15,000 new credit relationships in the first year of the program.
At a $1,500 average loan size, a 22% APR, and a 3% expected loss rate on a 12-month term, the unit economics on that cohort are substantially better than the same loan extended to a bureau-prime customer acquired through paid channels — because there's no customer acquisition cost. The relationship already exists.
The key constraint is not finding the customers. It's building the infrastructure to act on what you already know about them.
The Model Is Not the Hard Part
Risk teams sometimes frame alternative data credit expansion as a modeling problem. In our experience, the model is maybe 30% of the challenge. The larger challenges are data architecture (can you get 24 months of transaction history into a decisioning call in under 500ms?), product design (does the application flow create enough trust for thin-file customers to actually apply?), and regulatory posture (do you have adverse action codes ready for cash-flow-based declines?)
Each of those is solvable. But they require treating credit expansion into deposit-active thin-file customers as a product initiative, not just a model swap. The neobanks that approach it that way will find an acquisition channel with near-zero marginal cost. The ones that don't will continue declining customers they've already won, and watch those customers eventually get a credit card from someone else.
