Fraud detection for mobile-money payments in a small Haitian marketplace

Question

Our marketplace (MonCash payments, cash-on-delivery too) is seeing fraud: fake accounts, stolen phones used to pay, and sellers who never ship. We don't have a data science team. Where do we start?

Accepted Answer

Start with **rules and good data**. Machine learning only helps once you have labeled history.

**1. Rules that catch most early fraud**
- **Velocity**: many orders or payment attempts from one account, phone or device in a short time.
- **New account plus high value**: first order far above your median, especially electronics or phone credit.
- **Mismatch signals**: many accounts sharing one MonCash payer number, or one account paying from many numbers.
- **Seller signals**: many "not received" disputes, and sellers who request payouts right after sales.

~~~ts
const score =
  (ordersLastHour(user) > 3 ? 30 : 0) +
  (accountAgeDays(user) < 2 && order.totalHtg > 10_000 ? 40 : 0) +
  (payerLinkedToAccounts(payment.payer) > 2 ? 30 : 0);

if (score >= 60) holdForReview(order);    // don't auto-reject: review
~~~

**2. Controls that remove whole categories**
- **Hold seller payouts** until delivery is confirmed (escrow-style), with a delay for new sellers.
- Verify phone numbers by OTP at signup, and limit accounts per number.
- For cash on delivery, confirm by phone call for first orders above a threshold.

**3. Label everything.** Every chargeback, dispute and confirmed fraud goes in a table with the order id. After a few months that becomes your training data.

**4. Then ML.** A gradient-boosting model (LightGBM or XGBoost) on features like account age, velocity, amount and history, evaluated with **precision/recall**, not accuracy, because fraud is rare. It should **rank orders for review**, not silently cancel them. Customers wrongly blocked are a real cost too.

1 Answer