Alternative Credit Data: The New Underwriting Logic for Frontier Startups

Where formal credit bureaus are thin or absent, the question of how to underwrite a startup is not academic. It is the difference between a business that scales and one that stalls at the bank’s front door. The answer emerging from a decade of fintech research and field deployment is clear: alternative data, drawn from mobile-money flows, digital payment histories, supplier and utility records, and verified traction proofs, can substitute meaningfully for bureau scores in frontier markets. But the substitution is imperfect, carries real bias risks, and works best when multiple data streams are layered rather than used in isolation.

Why is the bureau gap a structural problem, not a data problem?

The gap is structural because lenders lack credible information about borrowers, not because capital is scarce. Small and medium enterprises drive economies and jobs, yet face persistent challenges in obtaining the financing needed to start, sustain, and grow across emerging market and developing economies. The scale of that gap is now precisely measured: MSMEs make up over 90 percent of all firms and account, on average, for 60–70 percent of total employment and 50 percent of GDP worldwide, yet according to the SME Finance Forum there is currently a roughly $5.7 trillion financing gap for MSMEs.

The proximate cause is not a shortage of capital. It is a shortage of credible information. FounderWise treats the $5.7 trillion MSME financing gap as an evidence problem: the capital exists, but most founders cannot present proof a lender can verify. Traditional credit systems often exclude individuals and small businesses without formal financial histories, leaving many women entrepreneurs and low-income borrowers invisible to lenders. In markets where bureau penetration is low, the standard underwriting toolkit (credit scores, audited financials, collateral registries) simply does not exist for the majority of potential borrowers. This gap disproportionately affects individuals in developing economies, micro-entrepreneurs, and marginalized communities who lack access to formal financial systems.

The response from the fintech sector has been to treat the absence of bureau data not as a dead end but as a design constraint. By incorporating alternative data, from mobile money transactions and digital payments to business and platform records, new scoring models can better capture economic activity that previously went unrecognized. The question is no longer whether alternative data can work. The evidence base is now large enough to ask the harder questions: what does it reliably predict, where does it fail, and how should verified traction proofs sit alongside it?

What alternative data actually predicts, and what it does not

The signal strength of mobile-money flows

The most mature body of evidence comes from mobile-money ecosystems, and the most studied of those is M-Pesa in Kenya. The annual value of transactions on M-Pesa reached 7.2 trillion Kenyan shillings by 2023, equivalent to 55 percent of Kenya’s GDP. That volume of transactional data creates a longitudinal behavioral record that no bureau could replicate for the same population. Mobile money generates a digital record of transactions, which facilitates creating histories and credit scores, easing the inclusion in formal finance.

The practical proof of concept is M-Shwari, the savings-and-credit product built on M-Pesa’s rails by Safaricom and the Commercial Bank of Africa. M-Shwari is a fully digital bank account operating over the rails of mobile money; launched in 2012 through a partnership between the Commercial Bank of Africa and Safaricom, within two years of launch there were more than 4.5 million active users (nearly 20 percent of the adult population) and approximately 10 million accounts had been opened. More recently, Kenya’s M-Shwari has transformed microlending by utilizing mobile money transaction data, providing over 15 million Kenyans with their first formal credit products. The macro outcome is striking: in 2024, more than 84.8 percent of Kenyan adults had access to formal financial services, up from just 26.7 percent in 2006, and the Central Bank of Kenya is explicit that the broad use of mobile money services is primarily responsible for this growth.

What mobile-money data predicts well is short-term liquidity behavior: frequency and regularity of inflows, the ratio of outflows to inflows, the breadth of the counterparty network, and the consistency of merchant payment activity. A suite of digital credit products use data from mobile phones to construct alternative credit scores, now used to originate loans to hundreds of millions of historically unbanked individuals; the key insight is that some behavioral patterns, such as frequency of international calls or the breadth of one’s social network, correlate with repayment, and machine-learning algorithms can detect these patterns and construct credit scores for mobile phone owners who would otherwise be excluded from formal financial services.

The limits are equally important to understand. Mobile-money scores are calibrated on short-duration, small-ticket loans. M-Shwari loans are small, short-term (30-day) products, and the average loan size in research samples is around KSh 480 (approximately $4.80). Extrapolating a behavioral score derived from micro-credit repayment to a working-capital facility of $50,000 for a growth-stage startup involves a category error that many lenders have not yet resolved. The data predicts whether a borrower will repay a small, short-term obligation under normal conditions. It says very little about how a business will perform through a demand shock, a supply-chain failure, or a currency move.

Payment history, supplier records, and utility data

Beyond mobile money, a second tier of alternative data has gained traction: digital payment histories from merchant platforms, supplier invoice records, and utility payment data. Research has found that the use of non-traditional data, such as rent payment and utility bill payment transactions, can help individuals with thin credit files or no credit history establish a credit profile and qualify for credit scores within three to four months. For SMEs, the equivalent signals are B2B payment consistency (whether a business pays its suppliers on time) and utility payment regularity, which proxies for operational continuity.

Current market conditions may present a unique window to build on substantial interest among industry, advocates, and policymakers in using non-conventional data sources such as digital wallet information and supply chain records for credit scoring and underwriting. FinRegLab’s 2024 market context report on Kenya, drawing on extensive interviews with industry, government, and academic stakeholders, found that tapping multiple non-conventional data sources is particularly important to expanding credit access and growth among women-owned MSEs.

The predictive logic of supplier and payment data is different from mobile-money flows. Where mobile money captures personal liquidity behavior, supplier payment records capture business-to-business reliability: a signal that is closer to what a trade-credit insurer or a supply-chain financier would use. When both signals are present and consistent, the underwriting case is substantially stronger than either alone. Because there is no single, readily available source that would provide insight into all micro and small enterprises, assessing multiple individual sources and building a platform to support access to the most promising and cost-efficient data is more helpful than executing a series of incremental initiatives that may tend to concentrate on larger, more formal businesses.

Machine learning amplifies signal, and it amplifies error

The shift from parametric to non-parametric models reflects the evolving nature of credit risk; conventional models impose rigid assumptions that were often incompatible with the unpredictable behavior of borrowers in emerging markets, while machine learning thrives on high-dimensional data, learning patterns that evade traditional techniques. Empirical research has demonstrated that adopting machine-learning techniques and incorporating cash-flow data into credit underwriting can significantly increase predictiveness and expand credit access for consumers without increasing lenders’ default risk.

The amplification, however, runs in both directions. The increasing adoption of artificial intelligence algorithms is redefining decision-making across industries, and in the financial sector, where automated credit granting has undergone profound changes, this transformation raises concerns about biases perpetuated or introduced by AI systems. The bias risks in alternative-data models are not hypothetical. Research documents that gender-based disparities in lending decisions have been found both under traditional and algorithmic scoring regimes, and that switching to machine learning does not automatically resolve them. In the context of developing-economy MSE lending, this matters acutely: women-owned businesses are disproportionately informal, disproportionately excluded from the transactional data that models are trained on, and therefore disproportionately at risk of being mis-scored downward.

A related problem is proxy discrimination. Ethnographic research on credit scoring intermediaries documents how organizations developed multiple renditions of credit scoring models that incorporated behavioral and social data in ways that created new forms of stratification not anticipated by regulators: a practical illustration of how proxy relationships emerge from data architectures rather than from explicit discriminatory intent. A model trained on mobile-money data in a market where women have lower average balances will encode that disparity unless the training pipeline explicitly corrects for it. The use of non-traditional data, particularly from social media, raises concerns about data privacy and borrower consent, and future research must address the ethical implications of harvesting digital footprints for credit assessments.

Where do verified traction proofs fit in the underwriting stack?

They sit above the transaction-data layer, supplying the one signal raw data cannot: verifiable proof of business momentum. Alternative data, at its best, answers the question: has this entity demonstrated consistent financial behavior? It does not answer the question: is this business building something that will generate durable cash flows? That is the gap that verified traction proofs are designed to fill.

A verified traction proof is an auditable, third-party-confirmable evidence of business momentum: a signed customer contract with payment terms, a confirmed cohort of repeat purchasers, a supplier relationship with documented payment history, a revenue milestone confirmed against bank or mobile-money records. Unlike a pitch-deck projection, a traction proof is backward-looking and verifiable. Unlike a bureau score, it captures business-specific signal rather than personal financial behavior. The free FounderWise Traction Audit turns that definition into a working test: 12 questions across 4 categories, about 3 minutes, a score out of 100, and the 3 biggest gaps in a founder’s evidence file named plainly.

The combination is more powerful than either component alone. A founder with strong mobile-money transaction history but no verified customer traction presents a different risk profile than one with both. A business with documented supplier payment consistency and a confirmed revenue run-rate is making a claim that can be cross-checked against multiple independent data streams. Lenders and investors who require verified traction proofs alongside alternative credit data are, in effect, building a triangulated underwriting model: behavioral data from mobile money, operational data from supplier and utility records, and forward-looking signal from confirmed business traction. The triangulation reduces the risk of any single data source being gamed or misread.

This matters particularly for growth-stage startups seeking working capital or early-stage debt. The alternative-data stack that works for a $500 micro-loan does not scale directly to a $200,000 revenue-based financing facility. At that ticket size, lenders need to see not just that the founder manages personal liquidity well, but that the business has a demonstrable revenue engine. Verified traction proofs are the instrument that bridges the two.

Platforms building in this direction, whether in Southeast Asia, Latin America, or East Africa, are converging on a similar architecture: automated ingestion of mobile-money or open-banking transaction data, enriched with supplier and utility payment records, with a human-reviewed traction proof layer for facilities above a defined threshold. The mechanics of how a deal closes in these markets increasingly depend on that layered stack being present and legible to the capital provider.

Why is the bias and fairness problem not optional?

Because the populations most in need of alternative-data credit access (women-owned businesses, informal-sector operators, rural enterprises) are precisely the populations most likely to be mis-scored by models trained on incomplete or skewed historical data. Operators and lenders building alternative-data underwriting systems in developing economies cannot treat fairness as a compliance afterthought.

In a 2024 case study, MIT researchers created a new technique that identifies and eliminates the specific attributes of training data that are the strongest contributors to a model’s biases about minority subgroups, and this approach also preserves overall accuracy because it preserves more of the data compared to other options. The technical solutions exist. The organizational will to implement them is the variable.

Three practical disciplines are emerging as standard in responsible alternative-data underwriting. First, disaggregated model auditing: running model outputs separately across gender, geography, and business formality to identify systematic score gaps before deployment. Second, feature transparency: being explicit about which data inputs drive which score components, so that a rejected applicant can understand and potentially contest the decision. Although machine-learning models outperform traditional methods, their black-box nature complicates result interpretation, and regulatory bodies may resist adoption without clearer explanations of model outputs. Third, human review at the margin: research has found that combining algorithmic predictions with human judgment improves credit decision accuracy and reduces discriminatory biases.

The history of credibility systems in financial markets shows a consistent pattern: the tools that survive regulatory scrutiny are those that can explain their decisions, not just optimize their default rates. Alternative-data underwriting is no different. The lenders who build explainability into their models from the start will have a structural advantage as regulation catches up to practice.

Building a defensible alternative-data underwriting stack

Principles for operators and lenders

The evidence points toward several non-negotiable design principles for any alternative-data underwriting system operating in a thin-bureau market.

Layer at least three independent signal types. Mobile-money flows, supplier payment records, and utility data each capture different dimensions of financial behavior. A model that relies on any single source is vulnerable to both gaming and to the structural gaps in that source’s coverage. By incorporating alternative data from mobile money transactions and digital payments to business and platform records, new scoring models can better capture economic activity that previously went unrecognized.

Calibrate ticket size to data depth. The predictive validity of mobile-money scores has been demonstrated for short-duration, small-ticket products. For larger facilities, the underwriting stack must include business-level signals (supplier payment history, confirmed revenue traction, open-banking cash-flow data) that go beyond personal transaction behavior. Transitioning to machine-learning models that incorporate both cash-flow and credit bureau data can produce the largest overall improvements relative to traditional approaches.

Build for women-owned and informal-sector businesses explicitly. Innovations like digital credit scoring, which use alternative data such as mobile phone usage and utility payments, allow women to obtain loans despite lacking formal credit histories. But this only holds if the model is designed to capture the data signals that women-owned businesses actually generate, rather than defaulting to signals that correlate with male-dominated formal-sector activity.

Treat verified traction proofs as a required input, not an optional enhancement. For any facility above a defined threshold, require at least one independently verifiable traction signal: a confirmed customer contract, a supplier payment record, a revenue milestone cross-checked against transaction data. This is not bureaucracy; it is the mechanism that makes the underwriting defensible to a credit committee, a regulator, or a downstream capital provider. Understanding how trust develops between founders and capital providers is inseparable from understanding what evidence those providers require.

Audit model outputs disaggregated by protected characteristics before deployment. Run the model against a holdout sample stratified by gender, geography, and business formality. If systematic score gaps appear, investigate the feature set before going live. This is the minimum standard for responsible deployment.

The broader question of how alternative credit data connects to capital platforms in developing economies is still being resolved. Open-banking frameworks, data-sharing agreements between mobile-money operators and lenders, and standardized APIs for supplier payment data are all works in progress. But the direction is clear: the infrastructure is being built, and the operators who understand the underwriting logic now will be positioned to use it when the infrastructure matures.

What this means