Chapter 03 · Section II · 15 min read
Credit scoring without credit bureaus
Most of Nepal is invisible to lenders. Why a working credit-decision model in this country looks nothing like a Western one — and what alternative data can and can't replace.
In Massachusetts a young person trying to get a credit card is, from the bank’s point of view, a thick stack of records — a student loan, a six-month rental history, three years of utility payments, a part-time job W-2. In Surkhet the same person is, from the cooperative’s point of view, a stranger. They have a citizenship card, a phone number, a face the loan officer may or may not know from the village. Almost everything you would call a “credit history” simply isn’t there.
This is the structural fact that makes credit in Nepal a fundamentally different problem from credit in a country with a working bureau — and what makes AI in this space both an enormous opportunity and a place where most foreign playbooks fail.
Why the Western credit model doesn’t transplant
The Western consumer credit stack assumes three institutions: a credit bureau that knows what loans every adult has and whether they repay; a salaried employment system that gives the bureau steady wage data; and a stable address history through utility bills and rental records. Take any of these away and the model wobbles.
Nepal is missing all three. CIB (Credit Information Bureau) exists, but it tracks formal bank credit only — invisible to it is anyone borrowing from a cooperative, a microfinance institution, an employer, a Marwari family business, or — most commonly — a relative. Roughly seventy percent of the adult population works informally, so wage data is sparse. Address histories rarely survive a move from village to town.
The result: a foreign credit-scoring model imported wholesale will look at the typical Nepali applicant, find almost no data, and refuse to lend. The applicant is not high-risk; they are invisible.
What alternative data can do
If you cannot see a credit history, you look for proxies. The serious work in this space in Nepal — done quietly inside the larger payment platforms and a handful of fintech startups — uses three families of signal.
Transaction history on the platform itself. Every Khalti or eSewa user has a history of payments: how often, how much, to whom, paid through which funding source. This is dense, structured, and — within the platform’s own boundaries — far better than any traditional credit bureau record. Someone who has been paying utility bills through eSewa monthly for two years is almost certainly stable in ways the formal banking system cannot see.
Phone and device signals. Length of phone ownership, recharge patterns, SIM swap history, device class. A user who has had the same number for five years and tops up Rs. 200 every Friday is a different risk than one with a brand-new SIM and an empty wallet. These signals are weak on their own but powerful in aggregate.
Network signals. Who has the applicant paid before, and how creditworthy are those counterparties? A young person whose first ten payments were to two grocery shops and a vehicle dealer that themselves have good history looks safer than one whose payments went to nothing recognisable. This is the data that platforms like Khalti hold uniquely and that no bank can see.
Where alternative data goes wrong
The promise is real. So are the risks, and they are worth taking seriously.
The data the platform sees is the data the platform sees. If the typical Khalti user is urban, salaried, under 35, with an Android phone, then the model learns to score urban-salaried-under-35-Android users well — and is unreliable on everyone else. A 55-year-old farmer in Achham who uses Khalti once a quarter is, to the model, a new and unfamiliar shape. The model may rate them poorly not because they are poor risk but because they look unlike the training data.
Proxy features can encode unfair patterns. Phone-recharge regularity correlates with salary stability — which correlates with formal employment — which correlates with caste, ethnicity, and geography in Nepal. A model that learns “recharges regularly = lower risk” may, by the back door, learn “high-caste urban = lower risk.” This is exactly the kind of harm a careful team checks for, and an incautious one ships into production.
The legal frame is unclear. Nepali law does not yet have a strong fairness or explainability standard for automated credit decisions. A bank can refuse to lend to you without explanation today. Whether the borrower has a right to know why — and whether the AI was the deciding factor — is currently an open question. The country will have to answer it.
What this means for cooperatives and microfinance
The most useful place for AI-assisted credit scoring in Nepal is not the big commercial banks. It is the cooperatives and the microfinance institutions — there are over thirty thousand cooperatives — that already lend to people the banks ignore. They suffer from high cost-per-loan and uneven repayment in ways an alternative-data model could meaningfully reduce.
A realistic pilot: a regional cooperative partners with a payment platform, anonymises members’ transaction histories, builds a small model to flag which applicants are likely to repay, and uses the model as a recommendation to the human loan officer — never as a sole decision. Done well, this is high-leverage AI-for-development work that no foreign team could ship as well as a domestic one.
Check your understanding
Quick check
—Why does a foreign credit-scoring model, imported into Nepal without modification, typically fail to lend to most of the adult population?
What comes next
The payments and credit story is one half of Nepali money. The other half — remittances — is the largest single source of foreign exchange in the country, and produces a stream of data that almost no one has used well for AI. The next section is about what remittance data could power, if we built for it.