Bureau models are built on credit history — installment loans, revolving accounts, payment timing on tradelines. What they miss is the much larger category of financial obligations that most Americans meet reliably every month without those obligations ever being reported to a bureau: rent, utilities, phone bills, insurance premiums, streaming subscriptions, gym memberships.
For a thin-file borrower, this is the entirety of their financial track record. It's not small. An adult who has been paying $1,200 rent plus electricity plus phone plus car insurance for 36 months has executed somewhere in the range of 144–180 separate recurring payment obligations. If the hit rate is 97%+ — which is typical for this demographic — that's a demonstrated repayment behavior pattern that a bureau score of 540 catastrophically misrepresents.
This article is about how to extract that signal technically — the feature engineering choices that determine whether recurring payment consistency becomes predictive in a credit model, or gets lost in the noise.
The Signal Problem: Recurring Payments Don't Self-Label
Transaction data from a bank account doesn't come pre-tagged with "this is a recurring obligation." You see a debit on the 1st for $1,175 to an entity whose bank descriptor might say "PROPERTY MGMT LLC" or "RENT 1020" or "ZELLE DAVID K LANDLORD." Identifying that as a rent payment — and then identifying the same rent payment consistently over 18 months — is a classification and deduplication problem before it's a modeling problem.
The three main technical approaches:
Merchant Category Code (MCC) Filtering
The crudest approach. Many recurring obligations hit specific MCCs: utilities (MCC 4900), telecommunications (4813), insurance (6300s). The problem is that rent payments often don't have clean MCCs — they go through Zelle, ACH personal payments, or property management companies with generic MCCs. Relying on MCC alone misses the largest recurring obligation most renters have.
Temporal Pattern Detection
A better approach: identify transactions that recur at consistent intervals (within ±5 days of a target recurrence period) and within a consistent amount band (±15%). A transaction that appears on the 1st of the month ± 3 days, in the range $1,100–$1,300, 18 times in 18 months, is almost certainly rent regardless of how the payee is labeled.
This approach handles the labeling problem by not requiring it. The temporal recurrence pattern is the signal, not the merchant category. It works for rent, car insurance paid monthly, utility auto-pay, loan payments that aren't on-bureau, and subscription services. It misses obligations paid in irregular amounts (variable utilities in a seasonal climate, for example, require a different approach).
Obligation Clustering
The approach we use at Lendiro: cluster transactions by payee fingerprint (combining normalized merchant name, amount band, and recurrence interval) to create an obligation register for the account. Each obligation has an estimated due date, an expected amount range, and a 24-month payment history. Consistency scoring runs against that register — not against raw transaction data — which allows us to handle real-world messiness like payee name changes when a property changes management companies, or slight amount drift when a subscription increases its price.
The clustering step is where most of the engineering complexity lives. It requires handling ACH reversal patterns (a payment that hits and then reverses within 3 days should not be counted as a success), distinguishing between service upgrades (amount increases midstream) and late payment backdating, and normalizing for bank processing delays that shift apparent debit dates by 1–2 days.
The Features That Actually Matter
Once you have a clean obligation register, the feature engineering question is: which derived features are actually predictive of loan repayment, and which are noise?
From our validation work, the features with the most consistent lift in Gini coefficient on thin-file populations:
Obligation hit rate, 12-month. What fraction of identified recurring obligations were met within a 5-day window of expected due date? This is the most direct measure of payment discipline. Even a simple binary cut (hit rate ≥ 90% vs. below) moves the needle meaningfully. The nuanced version accounts for whether misses were followed by catch-up payments — a pattern that signals cash-flow stress but not delinquency intent.
Obligation persistence. How long has the longest continuous recurring payment streak lasted without interruption? A borrower who has been paying the same set of obligations for 30+ consecutive months has a financial stability signal that a 6-month streak doesn't. This feature is particularly useful for distinguishing between thin-file-by-choice (someone who avoids credit on principle but is financially stable) and thin-file-by-circumstance (someone who is financially fragile and hasn't needed credit yet).
Obligation-to-income ratio stability. What fraction of monthly inflows are committed to recurring obligations, and how stable is that ratio over time? A borrower whose obligations consume 42% of inflows consistently has a different risk profile than one where the ratio oscillates between 30% and 65% month to month. High variance here signals either income instability or obligation management difficulty — both relevant for loan repayment risk.
Obligation count trend. Is the borrower adding new recurring obligations over time, or shedding them? Adding obligations (a new gym membership, an additional streaming service) correlates with financial confidence. Shedding them — especially shedding higher-cost obligations like insurance — can be an early indicator of cash-flow pressure that precedes a credit deterioration event by several months.
Calibration Pitfalls
Recurring payment features are not universally predictive across all product types and populations. Some calibration notes from our experience:
Subscription inflation. The proliferation of low-cost subscription services (streaming, software, newsletters) means that modern borrowers have more recurring transactions than borrowers from 5 years ago. Simple obligation count as a standalone feature overfits to this pattern — a borrower with 14 subscriptions is not necessarily more creditworthy than one with 4. Weight by obligation size, not just count.
Seasonal utility variation. In climate-variable markets — Texas in August, Minnesota in January — utility bills fluctuate significantly by season. An amount-consistency feature that penalizes a $180 July electricity bill versus a $90 March bill is creating a false negative signal. Features derived from utility obligations should use seasonal normalization or amount bands that accommodate known seasonal range.
Housing payment type differences. Property tax payments (annual or semi-annual), HOA fees, and rent have different temporal patterns. Conflating them in a single "housing obligation" feature reduces precision. Distinguish between monthly rent (most common for thin-file populations), payment app disbursements (increasingly common for landlords accepting digital payments), and genuine property ownership payments, which indicate a different risk profile entirely.
We're not saying recurring payment features are a complete credit model on their own — they're not. A model built only on recurring payment consistency misses cash-flow volatility, income stability, and balance dynamics that matter independently. What we are saying is that for thin-file populations, recurring payment consistency is one of the three or four most predictive feature families available, and ignoring it because the data requires extra engineering work leaves meaningful predictive signal on the table.
The Credit Bureau Gap and What Comes Next
There is movement toward getting rent payment data into bureau files. Experian RentBureau and similar products have been around for years, and the CFPB has published guidance encouraging the use of alternative data including rent payment history. The obstacle is that reporting rent to a bureau requires a landlord to participate in a reporting program — which most small landlords don't.
Until bureau-reported rent data achieves meaningful coverage (which may be years away), lenders who want to use rent payment history for thin-file underwriting need to extract it themselves from transaction data. The feature engineering described above is that extraction process.
For lenders who do invest in building these features, the payoff is access to a population that has demonstrably been paying their most important financial obligation — housing — month after month, often for years. The question is whether you're willing to build the pipeline to see that track record, or whether you'd rather decline them on a FICO score that doesn't know they exist.
