DenmarkFinance and insurance2017–2018Danske Bank
Machine-learning fraud detection
Danske Bank replaced rule-based fraud alerts with machine learning and cut false positives
Problem
Danske Bank's fraud detection relied mainly on hand-written rules. The rules produced record numbers of false alarms, at times a 99.5% false-positive rate, and the large fraud team spent much of its time investigating legitimate payments.
Approach
Working with Think Big Analytics, the consulting arm of Teradata, the bank built machine-learning models that monitor banking and mobile payments for fraud in real time, developed in 12-week sprints. It put the machine-learning ensemble into production while testing deep-learning models in parallel. The project was presented publicly in October 2017.
Results
Takeaway
The bank reduced wasted investigations and caught more fraud, but figures varied by source and stage. The vendor reported pre-production cuts of 50% in false positives and a 60% rise in detection; an analyst report months later put the production models' false-positive reduction at 20% to 30%.
Categories
- Size
- Large enterprise
- Industry
- Finance and insurance
- Business function
- Finance, risk and legal
- Innovation type
- Process innovation
- AI technology
- Predictive analytics and ML
- Adoption
- Co-developed with a partner
- Evidence
- Company's own claim
Sources
Source summaries
Teradata said Danske Bank worked with its Think Big Analytics consultants to put machine-learning fraud detection into production while developing deep-learning models. Nadeem Gulzar, the bank's head of advanced analytics, said pre-production tests cut false positives by 50%, allowing half of the fraud unit to move to higher-value work. Teradata also reported a detection rate rise of about 60% in pre-production tests and said false positives under the old rules at times reached 99.5%.
Analyst Doug Henschen reported that rising fraud from 2016 led Danske Bank to build machine-learning models with Teradata's Think Big Analytics in 12-week development sprints. He said the machine-learning ensemble cut false positives by 20% to 30%, and that deep-learning models built on TensorFlow showed double-digit detection gains in testing but were not yet in production.
Full details are in the original sources.
