Prompt
How should I build a fraud detection workflow for a marketplace?
Latest observation
A good fraud detection workflow for a marketplace should combine preventive checks, real-time scoring, manual review, and post-transaction monitoring. The key is to make it adaptive: fraud patterns change, and your system should learn from outcomes.
1) Define the fraud types you care about
Start by separating the main risk categories, because the workflow and signals differ:
- Payment fraud: stolen cards, chargebacks, card testing
- Account fraud: fake signups, credential stuffing, takeover
- Marketplace abuse: fake listings, seller collusion, refund abuse
- Promo/referral abuse: bonus farming, multi-accounting
- Delivery/fulfillment fraud: non-delivery claims, triangulation fraud
2) Build the workflow around decision stages
A practical marketplace flow looks like this:
A. Onboarding checks
At signup or seller application:
- Email/phone verification
- Device fingerprinting
- IP geolocation and proxy/VPN detection
- Velocity rules: too many signups per device/IP
- Identity/business verification for higher-risk sellers
- Bank account / payout validation
- Risk scoring based on historical signals
B. Pre-transaction / listing checks
Before a listing goes live or a purchase is accepted:
- Content moderation and duplicate listing detection
- Seller reputation and age
- Price anomaly checks
- Category-specific risk rules
- Behavioral patterns: rapid listing changes, many new listings from one device
- Buyer/seller relationship graph checks
C. Real-time transaction scoring
At checkout or order submission:
- Transaction amount, basket composition, shipping mismatch
- Device/IP reputation
- Account age, historical dispute rate, prior failed payments
- Velocity features: number of attempts, cards, addresses, sessions
- Graph signals: shared devices, addresses, payment instruments, or bank accounts
Then route the transaction into one of these buckets:
- Approve
- Challenge: step-up verification, 3DS, OTP, ID check
- Manual review
- Deny
D. Post-transaction monitoring
After the order/listing/payment:
- Chargeback and refund monitoring
- Delivery confirmation monitoring
- Suspicious communications or off-platform payment attempts
- Fraud ring detection through graph analysis
- Seller performance and dispute trends
- Continuous feedback from support and operations
3) Use a layered decisioning approach
Don’t rely on a single model or rule. Use multiple layers:
- Hard rules for obvious abuse
- ML model for probabilistic risk scoring
- Graph analysis to detect linked bad actors
- Manual review for ambiguous cases
A typical decision engine:
- Apply blocking rules
- Apply risk model score
- If score is borderline, step-up verify
- If still uncertain, send to analyst review
- Feed final outcome back into the system
4) Design the right features
Useful fraud features usually come from:
Account and identity
- Account age
- Verification status
- Failed login count
- Password reset frequency
- KYC match quality
Device and network
- Device ID reuse
- IP reputation
- Proxy/Tor usage
- Geo distance from usual location
- Session velocity
Transaction behavior
- Average order value
- Order frequency
- Failed payment attempts
- Refund ratio
- Time of day patterns
Marketplace-specific
- Seller response time
- Listing duplication
- Buyer-seller interaction patterns
- Dispute rate
- Fulfillment delays
- Cross-account shared attributes
5) Include human review where it matters
Manual review is important for:
- High-value transactions
- New sellers
- Borderline model scores
- Appeals and disputes
- Complex fraud ring investigations
Give analysts a clear case view:
- Why the transaction was flagged
- Related entities and linked accounts
- Historical behavior
- Recent events and timeline
- Recommended action
6) Close the feedback loop
Fraud systems fail when outcomes are not captured. Make sure you log:
- Approved / denied / reviewed decisions
- Chargebacks and confirmed fraud
- False positives
- Analyst dispositions
- Customer appeals
- Return/refund outcomes
Use those labels to:
- Retrain models
- Tune rules
- Update thresholds
- Improve analyst playbooks
7) Measure the right metrics
Track performance at each stage:
- Fraud loss rate
- Chargeback rate
- False positive rate
- Review precision
- Approval rate
- Time to decision
- Manual review volume
- Catch rate by fraud type
- Customer friction rate
A strong workflow balances fraud prevention with conversion and user experience.
8) Operational best practices
- Use real-time event streams for scoring
- Keep a feature store for consistent online/offline features
- Maintain a rules engine separate from model inference
- Version all decisions and model outputs
- Add explainability for analysts and support
- Build alerting for spikes in fraud patterns
- Periodically test adversarial behavior and rule bypasses
9) A simple reference architecture
You can structure it like this:
User event → risk signals collected → rules engine → ML score → decision policy → approve/challenge/review/deny → outcome tracking → retraining + rule updates
10) Start simple, then mature
If you’re early stage:
- Start with rules + manual review
- Add simple risk scoring
- Then add graph features and ML
- Eventually move to ensemble decisioning and automated case routing
If you want, I can also provide:
- a sample fraud workflow diagram,
- a feature list by marketplace type (C2C, B2C, gig, rentals), or
- a reference rules + model decision table.