Scaling a cryptocurrency exchange using Data Science & Machine learning
summary
Coinbase uses a continuous ML risk score to throttle buy limits and hold withdrawals, training on chargeback labels with models deployed via Vorpal Rabbit and SageMaker; production safety includes shadow-mode challengers, anomaly detection, clustering and a rule engine. Specifics: chargeback timelines show 95th percentile near eight days so the system holds funds up to 8 days for high-risk accounts, hot wallets keep 2% hot wallet online, a SIM-change ML check stops phone‑porting and prevents rings like the Verizon debit-card case — deployed and effective (no successful phone‑porting loss for a year).
AI-written summary · about
chapters (10)
- 0:00Start
- 2:30Proof of Work
- 8:30Coin Based Commerce
- 12:00User Flow Signup
- 22:49A Machine Learning Pipeline
- 23:22Why We Chose To Use Worple Rabbit
- 27:39What Are the Cases Where Machine Learning Does Not Work To Detect Fraud
- 28:12Anomaly Detection
- 30:13The Verizon Debit Card Ring
- 36:29User Security Challenges
related talks
33mSecuring Merchants: Combatting Scams and Fraud for Trust
27mDetecting fraud faster with Amazon SageMaker, featuring Coinbase
05mInterview
35mCoinbase's ML Training Evolution: From Sagemaker to Ray
46mSecuring Coinbase’s Edge Payments Infrastructure
30mScaling crypto payments for gaming volumes
47mTrain, Evaluate, Repeat: Building a Credit Card Fraud Detection System
30mGuarding Against Protocol Subversion at Coinbase