Differentiating Data Analytics from Data Analysis
In an interview with ZDNet, Josh Sullivan, leader of the data science and analytics practice at Booz Allen Hamilton, discusses the importance of human analysis complementing machine analytics. While the focus is on big data for any application, the discussion and points made provide insights and important considerations very relevant to modeling and analytics for fraud and risk management.
While card issuers and financial institutions have long employed modeling and analytics as a fraud prevention technique it is now becoming more common among eCommerce merchants. As a result we are seeing much more talk about machine learning, neural networks and advanced statistical models in the risk management marketplace. Although custom analytic modeling risk services require sophisticated platforms and technology, the reality is that a human element is required to ensure these services continue to run effectively.
Booz Allen Hamilton describes data science as the intersection of mathematics, computer science and “domain expertise” where the last term refers to an understanding of the details and nuances of the particular space where there is a problem that one is attempting to solve with data and analytics. Sullivan explains this further with the example of hiring several mathematicians and computer science PhDs that ultimately struggle to succeed without the expertise of people who have knowledge of the industry where the problem exists.
This concept rings true for modeling and analytics as it applies to fraud detection and risk management. Computer scientists are required to design and direct the infrastructure to manage, sort and handle all the data. Statisticians are needed to calculate the appropriate coefficients and properly weight the statistical models. Finally, both these parties will need to collaborate with experienced fraud and risk management professionals who can identify key patterns, risk signals and variables that are meaningful for detecting fraud.
While modeling and analytic services for fraud detection may utilize self-learning algorithms to retrain models, it is important to monitor the changes that are being made, and in many instances models will be updated based on human analysis independent of automated analytics. A statistical model left to learn on its own may draw connections between data points that have no meaningful relation, whereas human analysis works from the other direction by asking the right questions to find patterns that are confirmed with statistical analysis.
Historical data is a key component of modeling and analytics for risk management and other applications, and automated analytics are great at digesting and finding relationships in this data. But effective data science also requires cognitive ability and imagination which today is still best left to humans. As Sullivan put it, “machines do analytics,” but “humans do analysis.”
As it applies to risk management, organizations may have the resources to manage custom modeling and analytics entirely in-house, or they may use a third party service provider. In either case, the organization should have people with risk management experience contributing to the design and maintenance of models. When employing statistical models with machine learning capabilities the organization should ensure human analysis is utilized as well, which can be through internal expertise or managed by the service provider.
For more information: