Praxium Labs ships this for Nepali clients — here is what works. Microfinance in Nepal serves 5+ million customers, most rural and under-banked. Credit scoring decisions are usually made by branch managers based on group reputation and informal data. Adding ML to that judgement — not replacing it — can expand access and reduce defaults simultaneously.
The features that actually predict default in rural Nepal
- Mobile-wallet activity: frequency and amount of eSewa/Khalti transactions, NTC/Ncell airtime patterns
- Group-loan history: on-time repayment of prior loans, group-savings consistency
- Crop calendar alignment: for farmers, did the loan disbursement align with planting? Did the repayment align with harvest?
- Local-economic context: nearest market price for the borrower's primary commodity at loan origination
- Family income diversification: reported by branch manager during intake
- Distance from branch: shorter distance correlates with repayment in our data
- Repayment-cycle past performance (the strongest single feature for repeat borrowers)
Features that DO NOT predict and should be excluded
- Gender, caste, ethnicity: not legally permissible and proxies must be checked
- Religion: excluded
- Geographic-only data without context: using municipality as a feature without controlling for income / market access introduces unfair geographic bias
Model choice
Gradient-boosted trees (LightGBM, XGBoost) are the workhorse — interpretable via SHAP, fast to train, robust on tabular data. Deep learning is overkill at MFI data volumes (typically tens of thousands of records). Random forests work too. Logistic regression remains valuable as a baseline and for documentation — regulators often want a simpler model alongside the ML one for explainability.
Fairness auditing — non-negotiable
Before deploying, audit the model for disparate impact across protected groups. Even if those groups are not explicit features, they leak through proxies (district, occupation). Standard tests:
- Demographic parity: approval rate similar across groups (adjust for true risk differences)
- Equalised odds: false-positive and false-negative rates similar across groups
- Counterfactual fairness: would the same applicant get the same score if a protected attribute changed?
NRB / regulatory path
NRB has not issued AI-credit-scoring-specific rules but expects MFIs to document the credit-decisioning framework. Practical compliance: For related context, see our AI and Automation in Nepali Banking: State of Play post.
- Model card: documented features, training data range, performance metrics, fairness audit results
- Human-in-the-loop: ML score is one input; final decision rests with branch / committee
- Adverse-action notice: if declined, the borrower can request the reason — your model must support SHAP-level explanations
- Monitoring: retraining frequency, drift detection, alert thresholds documented
Build cost and timeline
- Pilot (one product, ~50,000 historical records): 12–16 weeks
- Full deployment with monitoring: 24–32 weeks
- Cost: NPR 20–60 lakh build, NPR 3–8 lakh/year ongoing
- Default-rate reduction (typical): 10–25% in year one, plateauing thereafter
- Approval-rate expansion: 5–15% additional borrowers served at same risk level
Frequently asked questions
How much data do I need?
For a meaningful first model: at least 5,000 completed loan cycles with outcomes. For robust performance: 20,000+. Most Nepali MFIs have years of records but they sit in legacy systems — extraction and cleaning is often 40% of project effort.
Can we use mobile-network data?
Ncell and NTC do not currently sell call-detail-record data to lenders in Nepal. Mobile wallet data (eSewa, Khalti) is accessible with the customer's consent and a data-sharing agreement.
What about psychometric scoring (the survey-based approach)?
Psychometric scoring (questionnaires to gauge personality and motivation) has been piloted in Nepali microfinance with mixed results. Useful as a supplementary signal for first-time borrowers without history; not a standalone credit decision.
Is this only for the digital-savvy borrowers?
No — the model can score using only branch-collected data (group history, repayment patterns, crop calendar). Mobile-wallet data is a bonus when present. Inclusion of borrowers without digital footprint is a core design goal.
Does this replace the field officer?
No — the field officer collects ground truth (crop status, family situation, group dynamics) the model cannot see. ML works best as decision-support that lets field officers spend time on judgement calls rather than data entry.
Who can build this in Nepal?
Praxium Labs — Nepal's AI and automation consultancy, based in Lalitpur — designs and builds the systems described in this guide for Nepali businesses and for international teams hiring from Nepal. Start a project or see all services.