When the target population lacks bureau data, building a credit model requires revisiting nearly every standard modeling assumption. The conventional process pulls bureau variables for training borrowers, defines default as 90 days past due, trains a gradient boosted classifier, and tests it on a holdout set. That process assumes infrastructure unavailable to thin-file populations. Starting fresh raises several linked questions: what forms the training set, how should the target be defined, and how can you verify that the model measures credit risk rather than another factor?
We addressed these issues while developing the Supersymm scoring engine. This post shares the reasoning behind our choices, not a definitive how-to guide, but an account of what we selected and why.
Training Set Challenge
Conventional credit models use borrowers who have previously taken loans. Their default outcomes are observed, bureau variables from origination are linked to those outcomes, and a model is trained to forecast later defaults. By definition, thin-file borrowers have not used formal channels, so their repayment behavior cannot be observed directly.
Two imperfect paths exist. One trains on a bureau-file population, assuming alternative data signals have comparable predictive relationships among thin-file borrowers. That is the transfer learning assumption, and it can fail. Relationships present among people with formal credit histories may not apply consistently to people whose financial lives are organized very differently. Still, when no alternative exists, it is a reasonable place to begin.
The other path is an origination experiment. Lend to thin-file borrowers through an initial heuristic or random assignment, observe repayment across a 90 to 180-day period, then use those outcomes to train a model designed for thin-file borrowers. Statistically, this is stronger, but loans must be originated before the model has full validation. Early lenders that run the experiment with careful records accumulate the richest training sets over time.
Supersymm has applied both methods in different pilots. Even small thin-file cohorts can produce valuable observed repayment data for lenders that have run them. For lenders reviewing the scoring layer before originating any thin-file loans, we apply the transfer method and document uncertainty bounds around prediction confidence.
Target Design
For bureau models, the usual target is 90 days past due on any obligation within 24 months. With thin-file borrowers taking a first loan product, 24 months is often impractical, while 90 days may not fit the offered terms. Repayment over 12 weeks for a microloan follows a very different path from repayment on a 24-month personal loan.
Our default target is "three or more missed weekly payments within a 90-day loan term, or any payment 30 or more days past due on a longer-term loan." This is stricter than the traditional definition. For microfinance-scale products, waiting for 90 days past due is nearly a lagging measure: when that clock ends, the loan is already a loss. A more sensitive outcome window is needed to detect early payment trouble.
The outcome window directly affects calibration. A model trained against 30-day-past-due produces score distributions unlike those from a 90-day-past-due target. When we give scores to lender partners, we record the training outcome definition and the calibration link between score and predicted default probability under that definition.
Selecting Features Without Bureau Data
Thin-file feature engineering is less about finding many variables than finding variables that predict well and are available throughout the population being scored. Possible alternative data is extensive, including hundreds of mobile metrics, utility payment dimensions, geolocation stability measures, and peer transfer patterns. Some lack predictive value, while others are unavailable for particular borrowers.
We assess the current feature stack using three tests. First is signal coverage: we include a feature only when it is available for at least 60% of the population. A signal present for just 20% of applicants leaves a model unable to score 80% of its target population with meaningful confidence. Second is predictive lift: in the training population, the feature must show meaningful information value, measured through IV scores and partial correlation with the outcome. Third is temporal stability: its distribution should not shift materially month over month. A feature that works in training but changes behavior six months later causes calibration drift.
Across our work, the features that have repeatedly met all three tests are: 90-day mobile recharge consistency rate, utility payment timing delta, meaning days before or after the due date averaged over 6 months, geolocation stability score, meaning variance in home and work location centroids over 60 days, and regularity of peer transfer cadence. They are not the model's only features, but they have the greatest individual IV and the widest coverage in the target population.
Validation Without Bureau Holdouts
Typical credit validation calculates population stability indices (PSI) and gini coefficients against bureau-file holdouts with known outcomes. Thin-file models lack that ground truth, so they need another validation framework.
We cross-validate using repayment outcomes observed in early pilot cohorts and combine that with calibration tests against benchmark populations that do have bureau data. Those benchmarks are not the model's main audience. They act as a sanity check: when bureau profiles are available, high scores for borrowers with strong profiles and low scores for those with weak profiles suggest the alternative signals capture something genuine.
We continuously monitor population stability as well. Each month, we compare current scoring feature distributions with those of the training population. PSI above 0.2 for any feature starts a review. Thin-file populations may change faster than bureau populations because mobile carriers and utility providers can alter behavior in response to economic conditions before those changes appear in bureau data.
Explainability and Adverse Action
Credit decisioning is regulated across most jurisdictions. In Brazil, LGPD principles state that automated decisions significantly affecting individuals must be explainable. For lending, this means adverse action notices: a borrower declined or given less favorable terms has the right to know the main reasons.
Common bureau adverse action codes, such as high credit utilization, short credit history, and too many recent inquiries, do not fit alternative data scoring. At score time, we produce reasons from the features that actually contributed. The output names the top three features that pushed the score below the approval threshold in language borrowers can understand. "Irregular utility payment history in the last 6 months" qualifies. So does "Low mobile recharge consistency rate over the past 90 days." Both describe real behaviors a borrower could address over time.
Adverse action wording is more than a cosmetic detail. It makes the scoring system usable by lenders subject to rules requiring explainability. Lenders asking about our compliance posture consistently request this output before completing integration. That scrutiny is proper, and we account for it.
Our Remaining Work
At writing, two issues remain unresolved. First is stability during economic stress. Our Brazil training data comes from a relatively stable period, so the model may not have encountered repayment shifts caused by a major income shock. We still lack enough stressed-period data to test behavior in those conditions, and we explicitly flag that uncertainty to lender partners.
Second is coverage among very young applicants. Borrowers under 22 may have under 12 months of mobile history, little utility footprint when living with family, and no meaningful peer transfer pattern. Coverage falls for this group and confidence bands widen. We are developing supplementary signal paths for young applicants, but results here should remain indicative rather than definitive.