The verdict in three sentences
Most merchants think you need a machine-learning model to fight fraud, when what you need first is deterministic velocity rules. Layering 10 simple rules before any model stops roughly 80 % of card-fraud attempts. Calibration target: false positives under 2 % and chargebacks under 0.5 %.
The 10 rules to layer, in order
These rules run in milliseconds, cost nothing to host, and need no training data.
| # | Rule | 2026 block threshold | Estimated fraud impact |
|---|---|---|---|
| 1 | Card velocity/hour | > 3 different cards/hour/device | High |
| 2 | Repeated failures | > 5 fails/10 min | High |
| 3 | Device fingerprint reuse | Same device, > 4 accounts | High |
| 4 | Billing country / BIN mismatch | Different countries | Medium |
| 5 | High-risk BIN range | Watched BIN list | Medium |
| 6 | First-order value cap | > 150,000 FCFA on new account | Medium |
| 7 | Blocklist shared across merchants | Email/card/device match | High |
| 8 | Disposable email address | Known temp domain | Low-medium |
| 9 | Odd hour + high amount | 2am-5am + > 200,000 FCFA | Low |
| 10 | IP / phone distance | > 500 km apart | Low-medium |
Calibrate without smothering real sales
The danger isn't blocking too little, it's blocking too much. Each rule must be measured on its blocked-fraud vs lost-legit-sales ratio.
| Metric | 2026 target | Alert zone |
|---|---|---|
| False-positive rate | < 2 % | > 4 % |
| Chargeback rate | < 0.5 % | > 1 % |
| Fraud blocked (rules only) | ~80 % | < 60 % |
| Orders in manual review | 3-6 % | > 10 % |
| Decision time per rule | < 50 ms | > 200 ms |
A blocklist shared across merchants (rule 7) is often the most profitable: a fraudster burned at one merchant is burned at all.
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Mini case study
Blessing runs an electronics store in Lagos with a 1.8 % chargeback rate. She turns on 6 of the 10 rules: card velocity, repeated failures, BIN mismatch, first-order cap, shared blocklist, device fingerprint. In six weeks her chargebacks drop to 0.4 %, she blocks ~82 % of attempts and loses only 1.6 % of legitimate sales. Estimated saving: 2,300,000 FCFA of avoided fraud per quarter.
FAQ
Do you really need machine learning? Not at first. 10 deterministic rules cover ~80 % of cases. ML only adds value beyond that, on the residual 20 %.
What's the right false-positive budget? Under 2 %. Above 4 %, you lose more in legitimate sales than in avoided fraud.
The most profitable rule? The shared blocklist: a device or card already flagged elsewhere is blocked instantly.
How to handle review orders? Aim for 3-6 % in manual review. Above 10 %, your thresholds are too aggressive.
Is mobile money affected? Less so: Wave/Orange Money push payments are hard to reverse, so card fraud is the rules' priority.
Let's talk about your project. We deploy an anti-fraud rules engine calibrated on your real numbers. WhatsApp +221 77 596 93 33.
Mohamed Bah
Fondateur, Kolonell
Passionate about digital and entrepreneurship in Africa, Mohamed has been helping Sénégalese businesses with their digital transformation since 2020. Founder of Kolonell, he believes every SME deserves a professional and accessible online présence.
