412 Million Transactions Per Month,
39 Mil. Customers, 14 Card Types
The project involved a US based card issuer with 112 million transactions per month. Fraud rate was hovering around 0.71%. The existing system in place was rule based and employed a machine learning module that was trained once per month (this was the fastest train rate possible), with multiple manual performance checks and system downtime each time the model was deployed). The existing system was operated by a 6 member risk analyst team working night shifts and manually creating and maintaining rules. Rapid Initial Integration We introduced the customer to training a new model with our ModelFlow app, based on one year of historic marked data, with the data eventually centering around 187 key features. This was done using gradient boost + mSynapse, with the whole initial process taking 1 week from product introduction to scaled deployment. The model was one-click deployed to production by our customer operators and set to retrain automatically every day. The results were, as compared to the full existing system – i.e rule/model based + fully manned team vs letting the VeriFlow API decide without any rules or manual intervention : Fraud rates below 0.13%, 40% decrease in false positive rate. Currently the customer is reducing the analyst team to 2 members only, with only minor monitoring Immediate ROI Benefits Model Trained Using Built in Presets