Credit scoring has always been, at its core, an attempt to predict future repayment behavior from past signals. Bureau data does this by recording what a borrower did with formal credit obligations: did they pay on time, did they miss payments, how much did they owe relative to their limits. The underlying theory is that past formal credit behavior predicts future formal credit behavior.
The thin-file problem is precisely that this theory requires formal credit history to exist. For borrowers with no history, the theory has nothing to work with. But that's not the same as saying those borrowers have no payment behavior history. It means their payment history has been happening outside the formal credit system, and we haven't built the tools to read it.
Repayment velocity in mobile wallet transactions is one of the clearest alternative signals we've found for predicting formal loan repayment. The mechanism is intuitive: a borrower who consistently clears small informal obligations quickly, even when there's no formal enforcement mechanism requiring them to do so, is demonstrating something about their relationship to financial obligations that generalizes to formal credit. We ran this hypothesis against cohort data across four markets and found consistent support for it.
Defining Repayment Velocity
The concept needs a precise operational definition to be useful, and getting the definition right matters more than it might seem. Several plausible definitions produce features that behave differently and have different levels of predictive validity.
The most predictive definition we found treats velocity as the time elapsed between an incoming obligation-like transfer and the corresponding outgoing settlement transfer, measured across all transaction pairs that fit the pattern. Not all mobile wallet transactions are obligation-related: merchant payments for immediate goods and services are spot transactions, not deferred obligations. What we're looking for are the transaction patterns that resemble deferred obligations: peer transfers that occur in clusters with a regular repayment pattern, small recurring payments tied to subscription or installment structures, and phone top-up patterns where the borrower maintains balance to cover regular charges.
The operational challenge is identifying which transactions are obligation-like without access to the counterparty metadata that would make it obvious. We use a combination of: transaction timing regularity (recurring transactions at similar intervals suggest an obligation structure), amount stability (fixed or near-fixed amounts suggest installments rather than ad-hoc transfers), and directionality patterns (a consistent outflow to a specific counterparty following an inflow from a different counterparty suggests settlement behavior). None of these individually is definitive; together, they allow probabilistic classification of obligation-type transactions with reasonable reliability.
What the Cohort Data Showed
We measured repayment velocity across four cohorts of thin-file borrowers who took out personal loans in Vietnam and Indonesia between 2024 Q2 and 2025 Q1, with 12-month performance observation windows. The cohorts were constructed to control for income level and loan size, so that velocity differences wouldn't simply reflect capacity differences.
Within each cohort, we split borrowers into quartiles based on their pre-origination repayment velocity score, computed from the 12 months of wallet transaction history prior to loan origination. The velocity score captured both the speed of settlement (measured as median days between obligation-type inflow and settlement outflow) and the consistency of that speed (measured as the fraction of obligation-type transactions that settled within a defined window, typically 3 to 5 days).
The gradient across velocity quartiles was consistent across all four cohorts. Borrowers in the top velocity quartile showed lower 90-day delinquency rates than those in the bottom quartile, with the middle quartiles falling in between in roughly monotonic order. The effect was present in both Vietnam and Indonesia cohorts, though the magnitude differed: the spread between top and bottom quartile delinquency rates was somewhat larger in the Vietnam cohort, which we attribute to differences in how informal obligation networks function in each market rather than a fundamental difference in the signal's validity.
We want to be appropriately cautious about what this demonstrates. This is a finding from production data across four cohorts, not a controlled experiment. We can't rule out that velocity correlates with some third variable we haven't fully controlled for. But the consistency across markets and cohorts, combined with a theoretically coherent mechanism, makes us confident it's a genuine signal rather than an artifact.
The Velocity-Capacity Confound
The most important methodological concern when using repayment velocity is separating velocity from capacity. A borrower who settles obligations quickly might simply be settling them quickly because they have ample funds; a borrower who settles slowly might be managing a tight cash flow that doesn't allow for immediate settlement. If velocity is primarily capturing capacity (i.e., it's a proxy for income level), it adds limited incremental value over direct income measures.
We tested this by stratifying by income decile and measuring the velocity-performance relationship within each stratum. If velocity were primarily a capacity proxy, it should be concentrated in high-income strata and should lose predictive power when income is controlled for. Instead, we found the velocity signal persisting within income strata, particularly in the middle income bands where capacity constraints are real but not severe. Borrowers with similar cash-flow levels showed meaningfully different velocity behaviors, and those differences tracked their subsequent loan performance.
This supports the interpretation that velocity captures behavioral disposition toward obligations, not just the ability to settle them. A borrower with moderate income who consistently prioritizes settlement of small obligations is demonstrating something about their payment priorities that is independent of their absolute capacity.
We're not claiming the capacity dimension is irrelevant. Capacity-to-repay is a distinct and necessary component of creditworthiness. What we're saying is that repayment velocity adds incremental predictive information beyond what capacity measures alone provide, and that this incremental information is meaningful for portfolio outcomes.
Practical Feature Construction
Translating the velocity concept into production features requires resolving several engineering choices that affect the feature's behavior.
Observation window. A longer observation window (12 months vs. 3 months) captures more behavioral data but is less sensitive to recent behavioral change. For thin-file borrowers who may have recently shifted their financial habits, a shorter window may be more relevant. We use a hierarchical approach: primary features computed over a 6-month window, with a secondary feature capturing the ratio of the most recent 3-month velocity to the prior 3-month velocity. The ratio feature captures behavioral trajectory without abandoning the stability information in the longer window.
Transaction classification accuracy. The quality of obligation-type transaction classification directly affects feature quality. Over-classification (treating too many transactions as obligation-type) introduces noise; under-classification misses real signal. We treat classification confidence as a feature weight: high-confidence obligation-type classifications contribute fully to the velocity feature; low-confidence classifications contribute proportionally. This probabilistic weighting reduces the impact of classification errors on the final feature value.
Minimum sample size. Velocity is unreliable when computed over very few obligation-type transactions. A borrower with only 2 or 3 classifiable obligation-type transactions in the observation window may show fast velocity simply because they had 2 transactions that happened to settle quickly. We apply a minimum count threshold (currently 8 qualifying transactions over the observation window) below which the velocity feature is flagged as low-confidence and down-weighted in the model. Borrowers falling below this threshold often also fall below our minimum history requirements for the primary cash-flow score, so this threshold seldom creates a discontinuity in practice.
Where Velocity Adds the Most Value
Looking at where velocity features contribute most to model lift, two borrower profiles stand out.
The first is the near-zero bureau file borrower who has been in the digital financial system for 12 to 18 months. Bureau data is insufficient for scoring; basic cash-flow features are present but limited by the relatively short history. Velocity features, which can be meaningful over a 6-month window, add incremental discrimination in this population where the standard features are weakest.
The second is the borrower who looks borderline on income-based features but has strong behavioral consistency. This borrower's cash-flow stability score sits near the decision threshold, where the correct decision is uncertain. Velocity features act as a behavioral tiebreaker, moving borderline-approve and borderline-decline decisions based on evidence of payment behavior that income features don't capture.
The populations where velocity adds least are those with strong bureau histories and those with very high or very low income relative to the loan amount requested. Strong bureau borrowers already have a rich direct record of formal credit behavior; behavioral proxies add limited incremental value. Very high-income borrowers have ample capacity, making behavioral tiebreakers less important. Very low income borrowers relative to loan size are primarily a capacity problem, and no behavioral feature compensates for a fundamental capacity mismatch.
Building velocity features into a production model requires an upfront investment in transaction classification infrastructure that doesn't exist in most conventional credit origination stacks. For lenders willing to make that investment, the payoff in accuracy at the decision boundary for thin-file populations is material. For lenders not yet at that stage, the basic cash-flow features we've described in other posts provide the majority of the benefit with less engineering complexity.
