People whose careers center on bureau-based credit scoring often react skeptically when we explain prepaid mobile behavior. "Can phone-plan management really say anything about credit risk?" At first, the connection seems weak. Time spent examining the data makes the logic tangible. A prepaid account demands the same discipline as any constrained financial resource: deciding what to spend, when to spend it, and how to preserve reliable access. Those are precisely the behaviors credit scoring seeks to assess.
About 80% of Brazil's mobile subscriptions are prepaid. This is the mainstream, not a narrow segment, with heavy representation among lower-income households and younger adults, who are also most likely to be thin-file or no-file borrowers. Their mobile activity consequently provides one of the broadest-coverage alternative data sources in our work.
How the Signal Appears
Prepaid accounts produce many behavioral events, including recharges, calls, SMS, data use, and, when permissions permit, app activity. Their value as credit indicators differs. The features we have found most predictive generally fall into a handful of dimensions.
Recharge frequency and consistency ask two questions: how often is the account topped up, and how predictable is the timing? Someone recharging every 7 to 10 days for several months shows a steady financial rhythm. In our data, that pattern correlates with comparable regularity across other obligations. Alone it is not a powerful indicator, but alongside other inputs it provides meaningful lift.
Recharge amount stability concerns whether top-ups stay near similar levels or fluctuate sharply. Large swings may point to volatile income, while a steadier pattern suggests household spending is being kept within predictable limits. Here too, the signal's direction carries more weight than any individual transaction amount.
Consistency under pressure may be the most revealing feature. When the balance approaches zero, does the borrower permit service to lapse, or add funds before interruption? How an account is handled under constraint maps unexpectedly well to management of other financial obligations under similar pressure.
A Specific Scenario
Take a borrower in a mid-size Brazilian city who works in informal retail and has no formal credit history. She has kept a prepaid account with the same carrier for four years. During that period, she averaged 48 recharges per year, with fewer than 3 service interruptions per year. Top-ups average R$30-R$50 and show little variance. She has never gone more than 14 days without recharging.
For a credit model, this shows sustained management of a recurring financial obligation. It indicates she can fund phone access consistently, even with modest amounts, and that her behavior is disciplined and low-variance. Alongside regular utility payments and stable geolocation, our model rates this profile as a strong thin-file candidate for a small personal loan. A bureau would produce no result; alternative data provides a meaningful view.
This does not mean she will repay every loan. It does mean the data distinguishes her meaningfully from a borrower with erratic mobile use and repeated lapses. That distinction is worth retaining.
Feature Engineering Factors
Before raw mobile event logs can serve as model inputs, they need substantial processing. Several considerations guide our approach:
Carrier normalization matters because carriers differ in recharge minimums, the time before service degrades at low balances, and promotions that can temporarily raise recharge frequency. Someone recharging 20 times in a month is not necessarily more disciplined than someone recharging 4 times if the first carrier permits R$5 minimum top-ups and the second requires R$30. Cross-borrower comparisons therefore require carrier-specific normalization.
Seasonality also matters. Mobile use changes around major holidays, payroll cycles, and agricultural income periods in rural or semi-rural areas. Without seasonal controls, a model may give too much weight to brief unusual periods. We apply rolling 90-day windows and month-of-year controls to reduce those effects.
Mobile behavior may change quickly when income changes, so a six-month-old recharge pattern may not represent present finances. Our standard scoring window gives the last 60 days twice the influence of days 61 to 180, then discounts data older than 180 days further. In practice, recency weighting is among the most important controls in alternative data feature engineering.
Consent and Privacy
Using mobile behavior requires consent that preserves the borrower's control of the information. We use two routes. Under direct borrower consent, the applicant expressly allows the lender to obtain account behavior, usually through a carrier API or an aggregator with carrier partnerships. Under the second route, carrier-to-lender partnerships give the lender a formal data-sharing arrangement with one or more carriers.
Each route involves tradeoffs. Borrower consent is straightforward from a regulatory perspective, but creates the participation selection effect noted above. Carrier partnerships provide coverage across the full population, while demanding relationship investment and continuing legal work. In Brazil, LGPD principles require purposeful, disclosed, proportionate data use. Consent must clearly state which data will be accessed and the specific purpose.
Supersymm does not retain raw mobile event data. We derive features, calculate the relevant metrics, and provide a score together with its signal contributions. Once extraction is complete, the raw transactional data is not retained. This design materially simplifies governance and lowers the data liability lenders assume when working with us.
Important Limitations
Mobile behavior has genuine predictive value as a credit signal, but its boundaries are genuine too. The central issue is multi-SIM use. A significant minority of prepaid users in Brazil operate two SIM cards, frequently across different carriers, to reduce costs or address regional coverage differences. A model viewing only one SIM may see just part of the picture. If the borrower conducts primary financial activity through another SIM, the observed pattern may not represent them.
Coverage is limited for another reason. Mobile behavior is strongest when the borrower has used the same account for at least 12 months. Newer accounts lack enough history for stable features. Signal depth is therefore constrained for young adults who recently obtained their first device and borrowers who recently changed carriers.
These constraints do not negate the signal's usefulness; they establish where it works. A scoring system that combines mobile behavior with other inputs, uses sensible recency weighting, and marks low-confidence results can obtain real value from the data. Treating it as a silver bullet instead will expose its limits.