Discover what a 2026 study reveals about alternative credit data, repayment risk, and the value of transaction history and digital footprints.
An applicant wants a credit card but has no conventional credit score. As a lender, how do you assess someone's repayment ability with limited information?
Researchers recently explored this question using alternative credit data from Mexico.
In their paper, FinTech Lending to Borrowers with No Credit History, Laura Chioda, Paul Gertler, Sean Higgins, and Paolina Medina examined lending decisions at RappiCard.
The study was revised in April 2026 and first appeared in November 2024. It offers fascinating evidence for lenders working with credit-invisible applicants.
The central question the paper raises is: Can someone's digital economic activity help lenders make responsible credit decisions?
The researchers worked with RappiCard, a Mexican fintech credit card provider.
Their dataset included applicants without conventional bureau scores. This made traditional credit scoring unavailable for the population being studied.
However, these applicants had other information lenders could examine.
Some had previously used Rappi, a popular delivery platform. Their activity created a history of purchases, payments, and everyday economic behavior.
The researchers combined several sources of information:
They used these inputs to build models predicting serious delinquency during the first year.
The researchers then tested how well these predictions matched observed repayment outcomes. They also compared models using different combinations of data.
This allowed them to examine which information sources contributed most to predicting credit risk. The results offer several interesting lessons for lenders.
A lack of credit history makes borrowers harder to assess. But does it mean lenders have no reliable information to work with?
The researchers found that alternative data could help predict creditworthiness even when traditional credit records were unavailable.
Their benchmark machine learning model achieved an AUC of 0.791, demonstrating a meaningful ability to distinguish between higher- and lower-risk borrowers using alternative data alone.
This does not diminish the value of traditional bureau information. Previous borrowing and repayment records remain important indicators of credit risk.
However, not everyone has an established credit history. Young adults, first-time borrowers, and other underserved applicants may have little or no information in their credit files.
The study shows that these applicants are not necessarily impossible to assess. Alternative data can provide predictive insights even when conventional credit scores are unavailable.
For lenders, this creates an opportunity to evaluate more applicants without relying exclusively on their borrowing history.
Alternative data is a broad category. Its predictive value depends on which signals lenders collect and how they analyze them.
The NBER researchers demonstrated this by testing different data categories separately.
Their results showed that Rappi transaction history performed particularly well. Meanwhile, a narrower set of digital-footprint variables delivered more modest results.
To understand these differences, we need to look at what each dataset actually contained.
The researchers analyzed applicants' previous activity on Rappi. This included purchase amounts, ordering frequency, shopping locations, and payment methods.
These records offered a detailed view of economic activity.
For example, repeated purchases can reveal spending patterns over time. Payment methods and ordering frequency provide additional context.
Source: Chioda, Gertler, Higgins, and Medina, revised April 2026.
The transaction-only model achieved an AUC of 0.769. It also generated 91% of the modeled profits achieved using all available information.
This suggests that transaction history carried substantial predictive information in this particular setting.
The study's digital-footprint dataset was considerably narrower than the term might suggest.
It included device types, operating systems, acquisition channels, and basic email characteristics. Researchers examined details such as email providers, capitalization, and numbers in email addresses.
These variables can provide useful information. However, they capture only a small part of an applicant's digital identity.
Modern digital footprint analysis can examine a much wider range of signals.
For example, RiskSeal analyzes email and phone history, online shopping activity, paid subscriptions, social media presence, and IP information.
This allows lenders to examine relationships between signals.
An email address alone may reveal little about creditworthiness. Its age, associated registrations, and consistency with other identity information can provide additional context.
The distinction matters because digital-footprint models are not interchangeable. Their predictive performance depends on the information available and how it is analyzed.
We explored this approach in RiskSeal's research involving over 6.1 million loan applications from seven Mexican lenders.
The results showed a consistent relationship between digital credit scores and repayment outcomes.
Default rates fell from 52% in the lowest score band to just 5% in the highest.
The research also examined how digital signals performed alongside traditional credit bureau information.
A digital-only model achieved an AUC of 0.67, compared with 0.69 for bureau data alone. Combining both increased AUC to 0.73.
These findings demonstrate how digital information can improve risk assessment when used alongside existing credit data.
Of course, the NBER and RiskSeal studies examined different populations and scoring approaches. Their results should not be treated as a direct performance comparison.
Still, both studies point toward an important conclusion.
Transaction records reveal patterns of economic activity. Broader digital-footprint analysis adds information about identity stability, online engagement, and other risk-related characteristics.
For lenders, the opportunity lies in identifying which data sources add meaningful predictive value to their existing models.
Another finding concerns the amount of historical information available.
The researchers grouped applicants by their previous Rappi activity. They found that prediction improved as transaction histories became richer.
For applicants with two or fewer transactions, the model achieved an AUC of approximately 0.742. For those with at least 29 transactions, AUC increased to approximately 0.828.
This difference illustrates the importance of behavioral history.
A single purchase reveals relatively little about someone's habits. Repeated purchases can show spending frequency, consistency, and changes over time.
These patterns give models more information to evaluate.
However, the finding also highlights a familiar challenge. Applicants with limited transaction histories may be harder to assess, much like those with thin credit files.
This is where digital footprint analysis offers another approach.
Rather than relying on extensive financial transaction records, lenders can access additional insights through email, phone, online registrations, and other digital signals.
This approach can also be more cost-efficient, particularly during onboarding. Lenders can analyze information already provided during the application process without requiring applicants to share detailed financial histories.
Of course, richer transaction histories remain valuable. Their predictive usefulness depends on the quality, consistency, and relevance of the information.
The goal is to balance data availability, predictive accuracy, and assessment costs. A limited transaction history does not necessarily mean limited information about an applicant. Other digital signals may help fill the gaps.
More information can also change how many applicants qualify for credit.
The researchers examined this through a simulated approval strategy. They selected approval thresholds designed to maximize modeled lending profits.
With all available information, the model would approve approximately 69.4% of applicants. Without Rappi transaction data, the approval rate fell to 49.5%.
The result is especially interesting because the applicants lacked conventional credit scores.
Better predictions helped the model identify borrowers it could serve profitably. Some applicants who appeared difficult to assess became more distinguishable.
Better risk separation can support portfolio growth alongside credit risk management. The same information that identifies higher-risk applicants can help recognize potentially reliable borrowers.
This means a lot for neobanks, BNPL providers, and other lenders targeting underserved segments. Their teams must balance approval rates, credit losses, and regulatory requirements. Additional predictive information can help navigate those competing priorities.
However, these approval rates came from a modeled strategy. They should not be interpreted as actual approval increases achieved in production.
The researchers also explored a less obvious question: Should lenders predict defaults or model customer profitability directly?
They developed two approaches. One predicted whether borrowers would default. The other estimated how much profit each borrower might generate.
The profitability model considered account-level revenues and costs over time.
You might expect the second approach to perform better financially. However, both models generated broadly similar profits.
This finding is interesting because borrower profitability involves several factors.
A low-risk customer might repay balances quickly. Their account could generate relatively little interest revenue.
Another customer might carry balances for longer. This could generate more revenue if repayments remain manageable.
As default risk increases, expected credit losses can eventually outweigh that revenue. Interest income, funding costs, repayment behavior, and losses all affect profitability.
For risk teams, the distinction matters when choosing model objectives. A default model estimates credit risk. A profit model examines the financial outcome of serving a customer.
These objectives can lead to similar decisions, as this study found. Nevertheless, neither approach removes the need for responsible lending policies and regulatory controls.
The study offers promising evidence, but it has limitations.
One important challenge is to reject inference. Researchers can observe repayment outcomes for borrowers who received credit. They cannot directly observe how rejected applicants would have performed with an approved card.
That creates uncertainty when applying the results to a broader population. A model may perform well among approved borrowers. Its performance across all applicants could differ.
The study also covers one lending ecosystem in Mexico. Its transaction information came from a specific delivery platform.
Other lenders may work with different customer populations, products, and data sources. These differences can affect predictive performance.
The findings therefore require careful interpretation before wider application.
A model trained on Mexican delivery-app transactions may perform differently elsewhere. Data quality, local purchasing habits, and product characteristics all matter.
This is particularly relevant when introducing alternative data into existing decision systems. In my view, historical backtesting remains essential before any large-scale implementation.
Risk teams need evidence that new variables improve predictions within their own portfolios. They also need to assess stability, explainability, and regulatory suitability.
The paper provides useful evidence for further testing. It cannot establish that every credit-invisible applicant can be scored safely.
The research offers several practical lessons for lenders considering alternative credit data.
1. Credit-invisible applicants may still have useful risk information.
Traditional credit records capture borrowing and repayment history. Other economic activities may provide additional predictive variables.
2. Alternative data sources contribute different levels of value.
The study's transaction data was especially informative. Its basic digital-footprint features also contained predictive information.
The strongest combination depends on the available sources and lending context.
3. Behavioral depth matters.
Repeated activity provides more opportunities to identify stable patterns. This makes the amount and continuity of historical data important considerations.
4. Better risk separation can support more approvals.
The researchers' modeled strategy approved more applicants with transaction data available. That finding supports further exploration of alternative data for financial inclusion.
For implementation, I would start by measuring incremental predictive value. A new dataset should contribute information beyond existing bureau scores, internal records, and decision rules.
Risk teams can test this through historical validation and controlled pilots. They should also examine performance across borrower segments and changing market conditions.
As more data becomes accessible, source selection will become increasingly important. The practical question is which new information meaningfully improves existing models.
Traditional credit reports provide valuable information about financial behavior. But they cannot describe every aspect of an applicant's economic activity.
The NBER study shows how transaction history and digital signals can help assess borrowers without conventional scores. It also demonstrates that data sources differ in predictive strength.
For lenders in 2026, this creates a practical opportunity. Carefully selected alternative credit data can improve risk assessment and potentially expand access to credit.
Discover the top risks BNPL providers must tackle, from fraud to thin files, to stay competitive and compliant in 2026.
Discover the power of alternative data in credit scoring to enhance financial inclusion.
.webp)
Explore the data gap behind credit invisibility and how to close it with smarter scoring.
