A FICO score is a compressed representation of a person's credit product history. It answers a specific question: given this person's track record of opening, using, and repaying credit accounts, how likely are they to default on a new obligation? That's a useful question. It also requires a prerequisite that a large portion of credit applicants cannot satisfy: a credit product history to compress.
Cash-flow underwriting asks a different question. Instead of "what does their bureau file say?", it asks: "what does their bank account show us about how they manage money?" These are not the same question, and the distinction matters enormously when you're making lending decisions on the segment of the population that has never or rarely used credit products.
The mechanics of a bureau score, quickly
FICO 8, the most widely used score version in consumer lending, computes across five factor categories: payment history (35%), amounts owed (30%), length of credit history (15%), new credit (10%), and credit mix (10%). Each of these requires tradeline data — credit cards, installment loans, mortgages, auto loans — that has been reported to a bureau.
Without tradelines, there is no score. The model doesn't have an opinion on someone who has never opened a credit product. It doesn't score them as a 300 (terrible). It simply has no output at all. That applicant is not a bad credit risk in the model's framework — they're an absence in the model's input space.
What cash-flow underwriting reads instead
Bank transaction data — accessible via open banking APIs or direct account connection — contains a dense record of how someone actually manages their financial life. A 24-month transaction history for a typical account contains thousands of individual data points. From those raw transactions, a properly engineered feature set can extract:
- Income volume and consistency — how much flows in per month, from how many sources, how regularly
- Recurring obligation payments — rent, utilities, subscriptions, loan payments — and whether they're made on time, early, or delayed
- Net cash flow position over time — is this person's account balance growing, stable, or declining over the lookback period?
- Volatility patterns — does their account regularly dip near zero before income arrives, or do they maintain a buffer?
- Overdraft behavior — frequency, amount, recovery speed
- Discretionary spending ratios — what fraction of income goes to non-obligatory categories?
None of this requires the person to have ever used a credit card. A person who has rented the same apartment for three years, paid utilities every month, and maintained a consistent positive balance is demonstrating financial discipline that is directly relevant to repayment prediction — it's just not being captured by the bureau because none of those accounts are credit products.
Feature engineering is where this actually works or fails
Raw transaction data is not a credit signal. It's a sequence of debits and credits. The signal lives in the engineered features you extract from that sequence. This is where naive implementations of cash-flow underwriting fall short: they compute simple aggregates (average monthly income, average balance) and expect those to be predictive. They aren't, reliably.
The features that carry real predictive weight are structural: they describe the shape of a person's financial behavior over time, not just the level.
Take income variability. Two applicants might both show $3,200 average monthly income over 24 months. Applicant A receives a paycheck of exactly $3,200 on the first and fifteenth of every month without exception. Applicant B receives varying deposits of $1,800 to $5,100, some from one employer, some from what looks like gig work, some from transfers. Their average income is the same. Their income stability is not. A model that only captures the average misses the structure entirely.
Or consider obligation consistency. Someone who pays $1,100 rent on the 28th of every month for 24 consecutive months — sometimes showing that payment initiating two days before the due date — is demonstrating a behavioral pattern that is strongly associated with creditworthiness. The timing regularity and the planning behavior (initiating before the due date) are features. A simple flag for "has rent payment in transactions" is not.
The Gini and AUC-ROC case for cash-flow features
The legitimate question to ask of any alternative underwriting approach is whether it actually discriminates between good and bad credit risk. Practitioner skepticism here is warranted — there are plenty of "alternative data" vendors with compelling narratives and weak model validation.
For cash-flow features on thin-file populations, the validation picture is favorable. On populations where bureau data is absent or sparse, models trained on engineered cash-flow features routinely produce Gini coefficients in the 0.35–0.50 range on holdout sets. That's meaningful discriminatory power — not as strong as a mature FICO model on a well-characterized population (Gini values of 0.50–0.65 in ideal conditions), but substantially better than the alternative, which is either declining everyone or approving everyone regardless of risk.
At Lendiro, we validate on a rolling holdout basis, tracking AUC-ROC on each cohort of thin-file decisions. The important measurement isn't the absolute Gini — it's the Gini on the population that would otherwise be unscorable. Measuring the model's performance against a bureau benchmark on bureau-scoreable applicants misses the point entirely.
Where cash-flow underwriting is not the answer
We should be clear about the limits of this approach. Cash-flow underwriting is not a replacement for bureau-based scoring on populations with adequate bureau files. A 750 FICO score from someone with 15 years of on-time payment history represents an enormous amount of validated credit behavior. That signal is real, it's tested, and supplementing it with cash-flow features typically produces marginal improvement at best and noise at worst.
Cash-flow underwriting is also not reliable on very short lookback windows. Thirty days of transaction data tells you almost nothing about structural financial behavior. Ninety days is better. Twelve months is workable for many feature types. Twenty-four months is where the signal-to-noise ratio of the most important features — income stability, obligation consistency, cash flow trend — reaches meaningful predictive power. Short-window cash-flow underwriting is a different product than what we're describing here and should be evaluated separately.
There's also a data quality problem to design around. Bank transaction data categorization is noisy. Automatic categorization of transactions as "rent" vs. "transfer" vs. "other" varies significantly by bank and by the transaction description format used by each biller. Well-built cash-flow models require upstream data cleaning and validation logic that rejects or flags low-quality inputs rather than feeding garbage to the feature pipeline.
Integration considerations for lenders
The operational architecture for cash-flow underwriting is meaningfully different from bureau-based decisioning. Bureau score pulls are synchronous API calls that return in milliseconds. Cash-flow analysis requires either a bank account connection step (via open banking), a file upload from the applicant, or integration with a bank data aggregator. That step adds friction and latency.
The design question for a lender integrating cash-flow underwriting is: where in the application flow do you trigger the bank data request? Too early and you're asking for deep financial access from applicants who haven't yet committed to the product. Too late and you've already run bureau checks that came back unscorable, creating a disjointed experience.
The pattern that tends to work: trigger the bank data connection request at the point where the bureau check returns unscorable. The applicant experience reads as: "We weren't able to evaluate your application using traditional credit data. Would you like to connect your bank account so we can consider your financial history?" Conversion on that offer, in our experience building this flow, is meaningfully higher than you'd expect — applicants who have been declined elsewhere due to thin files often understand and welcome an alternative path.
The 64 million credit-invisible and thin-file adults in America are not a future market. They're in your application queue today. Cash-flow underwriting is the mechanism for hearing what they're trying to tell you.
