Case study · Banking

Risk and fraud models for a bank

Models that cut bad loans and quieted the noise in fraud alerts.

Delivery work from our track record. We've left the client unnamed; the numbers are real.

55%

fewer loan cancellations

15%

fewer false fraud alerts

+8%

loan-decision accuracy

The challenge

Underwriting leaned on manual checks, and the fraud system cried wolf so often the team had learned to ignore it. Both were costing real money on every application.

What we did

  • Built credit-risk models with careful feature work and ensembles - accuracy that held up out of sample.
  • Scored transactions for fraud with anomaly detection, and tuned out the false alarms.
  • Added churn and segmentation models to point retention and marketing at the right people.

Stack

Pythonscikit-learnAWSApache AirflowSQL

Services used

More work