Why Credit Scoring Breaks for 2 Billion People, And What’s Replacing It?

Mahmoud Khalil6 min read·1 hour ago

Lending is a bet on trust. A lender hands a stranger money today on the promise of repayment tomorrow. The hard problem in any lending operation is finding credible signals about who will and will not pay back. Credit scoring is the industrialisation of that problem. It compresses thousands of variables about a person into a single number that estimates their likelihood of repayment.

For decades, the dominant model has rested on three pillars.

First, credit bureaus: Centralised data utilities such as Experian, Equifax, and TransUnion in most Anglo markets, iScore in Egypt, and CRB in Kenya, that aggregate borrowers' history from regulated lenders.

Second, credit reports: The raw data: existing credit lines, payment history, defaults, public records.

Third, models: Historically, FICO in the United States and analogous proprietary scores elsewhere, almost all built on logistic regression of bureau data, scaled 300 to 850 (FICO) or 0 to 999 (UK).

This system works well in markets where most adults engage with formal credit and lenders systematically report to bureaus. The United States, the United Kingdom, Western Europe, and parts of East Asia. In those markets, traditional credit scores are accurate, regulated (in the US under the Fair Credit Reporting Act since 1970 and the Equal Credit Opportunity Act since 1974), and have withstood decades of fair-lending scrutiny.

The system breaks for everyone outside these markets!

The thin-file problem

The World Bank’s most recent Global Findex puts the global unbanked population at roughly 1.4 billion adults. The underbanked, those with some banking touchpoint but no formal credit access, add another billion-plus. Even in developed markets, about 25% of US households are unbanked or underbanked.

These people are not necessarily higher credit risk; they simply have no credit history that a traditional model can see. Industry insiders call them “thin files” or “no files.” A 24-year-old in Khartoum running a profitable kiosk, paying her landlord on time and topping up her mobile every week, looks identical to a defaulter in a FICO-style model. She is invisible.

The cost is enormous. Lenders pass on profitable customers. Customers pay loan-shark rates because formal credit refuses them. Whole sectors, informal SMEs, gig workers, refugees, displaced households, sit outside the financial system entirely.

The two faces of “alternative data”

Robinson and Yu’s foundational 2014 report, Knowing the Score, draws a distinction that still matters in 2026. Every operator in this space should be able to articulate it cleanly.

Mainstream alternative data. Same kind of signal as traditional bureau data, just from new sources: monthly utility, telco, rent, and BNPL payments. Same logic (“did they pay regular bills on time?”), same fair-lending scrutiny, just fed by new pipes. Reporting telco and utility data into US bureaus has been the most consequential financial-inclusion move of the last 15 years. Experian Boost claims tens of millions of US consumers have raised their scores using this approach.

Fringe alternative data. Fundamentally different signals: smartphone metadata (Credolab), social-graph signals (Lenddo’s psychometric era), e-commerce behaviour, GPS patterns, and even SMS tone. The case for fringe data is that it works where bureaus are absent. The case against is that its predictive value, fairness, and regulatory standing are all unproven and contested.

Operators building in emerging markets are usually forced into mainstream-alt or fringe-alt because bureau coverage is poor or non-existent. The art is sequencing. Start with mainstream-alt where possible (telco, mobile money, utilities). Use fringe data as a complement, not the foundation.

Why ACS dominates emerging markets

Three structural conditions push emerging-market lending toward alternative data.

Bureau gaps. Sub-Saharan Africa has bureau coverage well under 10% of adults in many countries. Even where bureaus exist, the reporting requirement on lenders is patchy. In many fragile markets, formal credit-bureau data on the adult population is functionally close to zero outside the regulated banking sector.

Mobile leapfrog. Smartphone penetration is 60 to 80% across much of Africa and MENA. Mobile-money penetration in Kenya, Tanzania, Uganda, and increasingly Sudan is well above formal bank-account penetration. Telco data is often better than bureau data because the underwriter sees live transaction velocity rather than a stale monthly snapshot.

Cost economics. A traditional retail lender’s underwriting cost is too high for a $50 loan. ACS makes micro-credit underwriting feasible by replacing humans with models running on data the lender already collects.

This is why almost every successful African and MENA fintech of the last decade, M-Pesa and M-Shwari in Kenya, Tala, Branch, JUMO, Carbon, valU, Tabby, runs on alt-data underwriting at its core, even when the front-end product looks like a payment or BNPL app.

The 2026 state of the industry

A snapshot of where the global ACS market sits today:

Market size: roughly $13.5 billion in 2026, projected to reach about $37 billion by 2035 at an 11.91% CAGR per Global Growth Insights. Other estimates peg the narrower “AI-in-credit-scoring” segment at 23% CAGR.
Geographic concentration: Asia-Pacific holds about 26% of the market. Roughly 70% of APAC fintech platforms now rely on alt data. Africa and Latin America are smaller by revenue but the highest-growth regions.
Adoption indicators: 62% growth in alt-scoring usage among emerging economies. AI integration in scoring is at 58% of providers. Cloud deployment is now the default.
Top concerns from operators: data privacy (62%), model transparency (57%), regulatory compliance (58%), integration complexity (55%), customer trust gaps (53%).

Trend lines for 2026:

Explainable AI (XAI) is becoming mandatory, with Singapore’s MAS, the EU AI Act, and Kenya’s CBK draft guidance all pointing in the same direction.
Psychometrics is having a quiet second wind, mostly as a complement rather than the primary signal.
Real-time scoring (sub-24-hour decisions) is now table stakes for any consumer-facing product.
Regulators across Africa are tightening. Kenya’s 2022 amendment to the CBK Act subjected digital lenders to direct CBK supervision. Similar moves are underway in Nigeria and South Africa.

What stakeholders in low-bureau markets should take from this

Three takeaways for anyone building or evaluating an ACS product in an emerging market.

Bureau absence is a structural reality, not a feature gap. In markets where bureau coverage is functionally zero, ACS is not optional. It is the entire underwriting strategy. Anyone trying to retrofit a bureau-augmented model into a market without bureaus is solving the wrong problem.

Sequence data sources. Do not chase the flashiest first. Mainstream alt-data (mobile-money transaction history, telco top-up patterns, utility and rent receipts) is the first ring. Fringe alt-data (device metadata, app usage, behavioural) is the second ring, used to break ties on customers that the mainstream layer cannot decisively rate. Building an underwriting model around fringe data alone, however compelling it sounds in a deck, is the temptation Robinson and Yu were already warning operators about in 2014. The warning has aged well.

Early loan volume is training data in disguise. No model performs well at first. Operators do not have a real underwriting model until they have a real default cohort. The first few hundred to a few thousand loans are training data, and any signal not instrumented during that period is permanently lost as a feature for the model. Log everything reasonable, even data that has no immediate use, because feature engineering on a backfilled dataset is a fundamentally weaker exercise than feature engineering on data captured live.

What’s next

In the upcoming articles, this series goes deep on alternative data sources: which signals actually predict repayment, how operators turn raw streams (a mobile-money ledger, a telco top-up history) into model features, and where the predictive ceiling sits for each data type.

If you build, invest in, or regulate emerging-market lending, this series is for you. New article every morning for the next 13 days.