Layer fraud rules under an ML model sensibly

Steps

1. 1Start with hard rules for non-negotiables (sanctions, velocity caps)
2. 2Score remaining traffic with a model trained on confirmed labels only
3. 3Define review queue capacity; set threshold to match it, not aspiration
4. 4Measure false-positive friction cost alongside fraud loss prevented
5. 5Champion-challenger rule/model changes in shadow mode first
6. 6Feed analyst outcomes back into labels weekly, closing the loop

Common Pitfalls

• ▲Blocking based on proxy features discriminating protected groups
• ▲Thresholds tuned once and forgotten as patterns shift

Commands

$ npx skills add aniruddhaadak80/skills --skill risk-modeling-fraud-rules-vs-model

$ npx skills add aniruddhaadak80/skills --skill risk-modeling-fraud-rules-vs-model -g

Tags