Skip to main content

Fraud Detection

TL;DR
  • Signals = Data points indicating risk (device, velocity, behavior, identity)
  • Rules = Fast, explainable, good for known patterns
  • ML models = Find patterns you haven't seen, but they need training data
  • Stack by stage: Starter (rules + AVS) → Intermediate (+ device ID + ML) → Advanced (+ behavioral)
  • Detection is layered. Nothing catches everything on its own

Something's getting through and you want to catch it earlier. Build the stack your volume justifies, not the one a vendor's selling.


How Detection Works

ComponentPurposeExample
SignalsRaw data pointsDevice ID, IP, velocity, AVS result
RulesKnown-pattern matching"Block if >5 cards in 1 hour"
ML ModelsPattern discoveryAnomaly score from transaction features
ReviewHuman judgmentEdge cases, high-value orders

Core Topics

Evidence Framework

How to sort signals into Tier 1 and Tier 2:

  • Tier 1: High confidence, standalone indicators
  • Tier 2: Supporting evidence, combine for confidence

Rules vs. ML

Which one you've actually got a use for:

  • When rules work best
  • When ML excels
  • Hybrid approaches

Detection Methods

MethodCoverageUse Case
Velocity RulesTransaction patternsReal-time decisioning
Data EnrichmentIP, email, phone signalsEnriching transaction data
Building Fraud RulesRule sets, allow/block listsDay-one setup and lifecycle
Fraud Model FeedbackML feedback loopsModel accuracy and monitoring
Device FingerprintingDevice/browser attributesAccount-level linking
Behavioral AnalyticsUser behavior patternsATO, bot detection
Identity VerificationIdentity confirmationApplication, step-up
Manual ReviewComplex/edge casesHigh-value decisions

Building Your Detection Stack

Starter Stack

  1. Basic velocity rules
  2. AVS/CVV verification
  3. Simple device ID
  4. Manual review queue

Intermediate Stack

  1. Advanced velocity rules
  2. Device fingerprinting service
  3. Data enrichment (IP, email, phone intelligence)
  4. ML scoring (vendor or custom)
  5. Fraud rule lifecycle management (shadow mode, allow/block lists)
  6. Case management system

A full-stack fraud platform (Sift, Sardine, Kount) bundles items 2-4 into one integration. You can assemble the same thing from separate vendors. It's just more plumbing. Fraud vendors covers when each way makes sense.

Advanced Stack

  1. Real-time ML models
  2. Behavioral biometrics
  3. Network analysis
  4. Custom feature engineering
  5. Automated decision engine
  6. ML feedback loops and model monitoring
  7. Operational cadence (daily/weekly/monthly reviews)

When to Escalate

It's all in the evidence framework: Tier 1/Tier 2 indicators and when to escalate.