Transfer Wise Data Talk 2
1 like203 views
The document discusses strategies for fighting fraud while maintaining a good customer experience. It describes transitioning from a rule-based fraud detection system to incorporating machine learning models, which can learn from data and update dynamically unlike static rules. By combining machine learning models with existing rule-based and manual review systems, companies can improve fraud detection abilities while balancing various priorities like price, speed, convenience and coverage.
Transfer Wise Data Talk 2
- 1. FIGHTING FRAUD WITH GOOD CUSTOMERS IN MIND
- 2. LAUR +
- 4. MAKING THE FINANCIAL SYSTEM FAIR AND TRANSPARENT
- 5. €700 MILLION IN TRANSFERS PER MONTH
- 6. €30 MILLION OF CUSTOMERS MONEY SAVED EVERY MONTH
- 7. TRUSTED ALL OVER THE WORLD CAN SEND MONEY FROM
- 8. TRUSTED ALL OVER THE WORLD CAN SEND MONEY TO
- 9. FRAUD 101
- 13. CHARGEBACK =
- 14. PRICE
- 15. WHAT TO DO ABOUT IT?
- 17. TEAM
- 23. RULE BASED SYSTEM - Become outdated - New rule for each pattern
- 24. PRICE
- 25. PRICE SPEED
- 28. TEAM
- 30. FEATURE 1 FEATURE 2 FEATURE 3DYNAMIC RULE MODEL
- 31. SOURCE CURRENCY AMOUNT TIER £ NUMBER IN EMAIL?DYNAMIC RULE MODEL FEATURE 1 FEATURE 2 FEATURE 3
- 32. SOURCE CURRENCY AMOUNT TIER £ NUMBER IN EMAIL? USD 100-200 Yes EUR 1000-1100 Yes GBP 500-600 No … DYNAMIC RULE MODEL DYNAMIC RULE 1 DYNAMIC RULE 2 DYNAMIC RULE N FEATURE 1 FEATURE 2 FEATURE 3
- 33. SOURCE CURRENCY AMOUNT TIER £ NUMBER IN EMAIL? USD 100-200 Yes EUR 1000-1100 Yes GBP 500-600 No … DYNAMIC RULE MODEL DYNAMIC RULE 1 DYNAMIC RULE 2 DYNAMIC RULE N FEATURE 1 FEATURE 2 FEATURE 3 30% 70% 76%
- 34. SOURCE CURRENCY AMOUNT TIER £ NUMBER IN EMAIL? USD 100-200 Yes EUR 1000-1100 Yes GBP 500-600 No … THRESHOLD: 65% DYNAMIC RULE 1 DYNAMIC RULE 2 DYNAMIC RULE N FEATURE 1 FEATURE 2 FEATURE 3 30% 70% 76% DYNAMIC RULE MODEL
- 35. SOURCE CURRENCY AMOUNT TIER £ NUMBER IN EMAIL? USD 100-200 Yes EUR 1000-1100 Yes GBP 500-600 No … THRESHOLD: 65% DYNAMIC RULE 1 DYNAMIC RULE 2 DYNAMIC RULE N FEATURE 1 FEATURE 2 FEATURE 3 30% 70% ADD +10 76% ADD +10 DYNAMIC RULE MODEL
- 38. FRAUD ENGINE STATIC RULES SCORING DYNAMIC RULES MANUALLY DEFINED: - SCORES - THRESHOLDS
- 39. FRAUD ENGINE STATIC RULES SCORING DYNAMIC RULES MANUALLY DEFINED: - SCORES - THRESHOLDS NEW RULE MODEL FOR NEW PATTERN
- 40. TEAM
- 41. MACHINE LEARNING
- 42. MACHINE LEARNING BLACK BOX Machine learning algorithm RULESOUTPUT FRAUD SCORE 0..100 PARAMETERS Θ . ..
- 48. 0% 25% 50% 75% 100% 0% 25% 50% 75% 100% * ILLUSTRATIVE
- 49. 0% 25% 50% 75% 100% 0% 25% 50% 75% 100% * ILLUSTRATIVE
- 52. JASON TAN “NOT A RACE BETWEEN ALGORITHMS”
- 53. BLACK BOX Machine learning algorithm TRANSACTIONDATA FRAUD SCORE 0..100 PARAMETERS
- 54. SCALE DATA FEATURE ENGINEERING • Number in email • Name in email • Device fingerprinting • Geo-location mismatch • …
- 55. OUR CHALLENGE
- 56. 80 - 20
- 58. THANKS! Q?