Head of Global Analytics, Danske Bank
Safety and Privacy
Danske Bank is a leading Nordic full-service bank, with local roots and bridges to rest of world. Founded in October 1871, they have helped people and businesses realize their ambitions for 145+ yrs. They serve personal, business and institutional customers through banking services, life insurance & pension, mortgage credit, wealth management, real estate and leasing services. Danske Bank adapts to the changes in the industry providing the best customer experience by making banking and financial decisions easy. Digitalization has increased efficiency, strengthened relationships and improved market position. One KPI for Danske Bank is to be data-driven, with understanding in the data, pattern definition, pattern recognition and efficient case handling.
Bank fraud is the use of potentially illegal means to obtain money, assets, or other property owned or held by a financial institution, or to obtain money from depositors by fraudulently posing as a bank or other institution. It is becoming more prevalent and more sophisticated with every passing month.
Danske Bank’s original fraud detection system was based on handcrafted rules that had been proactively applied by the business over time. With record numbers of false positives - at times reaching 99.5% of all transactions - the costs and time associated with investigation had become significant, with the bank’s fraud detection team feeling overworked.
Enterprise analytics are rapidly evolving and moving into new learning systems enabled by a renaissance in Artificial Intelligence. Danske Bank’s challenges included a lack of enterprise grade AI compute infrastructure, an incomplete data and analytic foundation for AI, and the inability to operationalize at scale.
Danske Bank started a multi-step program to enhance its existing fraud detection engine with AI technology. The problem was tackled using a heterogeneous agile work team with complementary capabilities: Platform and data engineers, data scientists, lines of business and investigative officers. The team was formed by Danske Bank employees and Teradata Think Big Analytics consultants, all working as one team.
The initial goals included:
- Develop a scalable and expandable platform which follows Danske Bank’s blueprint of digitalization.
- Data-driven approach to find data patterns and complement the existing fraud engine.
- Handle large data volumes for training models on transaction data.
- Implement a real-time solution that can score live transactions in under 300 ms.
- Reduce number of false-positives by at least 20-40%.
The solution finally included a new AI framework, Machine Learning techniques in production and Deep Learning challenger models.
The Danske Bank fraud detection unit recognises the need to use cutting-edge techniques to engage fraudsters not where they are today, but where they will be tomorrow. Using AI, Danske Bank already reduced false positives by 50% and as such can reallocate some of the fraud investigators to higher value responsibilities. Investigative officers were also brought in to better understand anomaly detection.
From the customer’s perspective, it improves their experience with Danske Bank in that the Danske Bank investigation team no longer needs to ‘hassle’ customers to verify the origins of each transaction. In addition, the AI models can identify an increased number of fraud cases, which the existing rule engine and the fraud investigators would have not caught.
The implementation of the new systems has also been thought to support the fraud detection unit, without damaging the current processes: AI in Danske Bank is meant to be a value addition, which supplements the human work carried out by the fraud analysts and investigators.
- Increased efficiency in the fraud investigation unit, with a reduction of false positives: Danske Bank has experienced an immediate 20-40% reduction in fraud false positives, which will be further reduced in the future. Danske Bank usually experienced thousands of false positives every day. That equates to people, time and money as investigative teams must examine the transactions and eventually call the customers to verify.
- Millions of money saved, benefitting the company’s bottom line: A 30-40% improvement in the detection of fraud (which will also be further improved over time), has a big monetary savings impact. Fraud attacks affect both the bank customers and finance. An improvement in the fraud detection is beneficial to the bank bottom line, but also increases the customer satisfaction.
- Increased response time with the Danske Bank AI real-time framework: Danske Bank now has a real-time engine that scores the transaction and provides immediately actionable insight, capable of handling many million transactions per year. The engine works in real-time, giving responses to transactions within 300 milliseconds. The framework has solved the first use case – fraud detection – and it is an asset for the future: It is designed to support further solutions for Danske Bank business units.
Teradata Think Big Analytics frameworks, such as: Analytics Ops; and LIME (Locally Interpretable, Model Explanation) used to explain outputs of algorithm models. This software helps the team using the model to explain factors that make them believe the model is solid. Machine Learning algorithms: Logistics regression, Bayesian models, Other technologies. Deep Learning algorithms: CNN, LSTM, Auto-encoders
Fighting fraud with AI is the new approach in the fast-changing market. It's no longer possible to create a fixed set of rules and hope they'll remain the same for the next 6mos. The company has one live model running in production - the ‘champion’; as well as a set of ‘challenger’ models which are continuously learning and adjusting. The ‘challenger’ model results are compared to the ‘champion,’ model and ultimately, they replace the ‘champion’ with the ‘challenger.’ This process enables the two models to be continuously trained and retrained, while the performance of the models are analyzed. The adopted technology pamphlet starts with simple machine learning models, regression, Bayesian models - all the way into deep learning - CNNs, LSTMs. and auto-encoders. Danske Bank has adopted an innovative use of deep learning techniques for financial transactions, by leveraging the experience from images processing to gain value in fraud detection. The success story from the adoption of AI to fight fraud had a synergy effect, taking the innovation wave to all involved systems, both within the Danske Bank AI framework and to all related systems. The experience from this initiative will be now reused to deliver new use cases across business units. The company is working on bringing the right AI tech to solve business problems and add business value.
The project main phase was developed and brought to production in 12 weeks. Lessons learned: Don’t buy a package; build it. Start with something concrete and small; can't place a man on the moon in 6mos. Leverage both AI and industry best practices; partnering with Think Big Analytics helped us to bring a product into production. Danske Bank Head of Global Analytics Nadeem Gulzar was nominated as a 2017 European Hortonworks data heroes finalist.